P1857R3
Modules Dependency Discovery

Published Proposal,

This version:
http://wg21.link/p1857r3
Author:
(Apple)
Audience:
EWG, SG2, SG15
Project:
ISO/IEC JTC1/SC22/WG21 14882: Programming Language — C++

Abstract

The changes to import in [P1703r1] were both too much and not enough to enable fast dependency scanning via partial preprocessing. It made import at the start of a line anywhere in the program an import-keyword, breaking things like import->doImport();. Also, no similar changes were applied to module. This paper fine-tunes the context-sensitivity of import and module, applying the fixes discussed on the reflector. This resolves US026, US121, US125, GB126, US127, US128, US136, US137, US138, US139, US140.

1. Change Log

2. Effects of This Paper

At the start of phase 4 an import or module token is treated as starting a directive and are converted to their respective keywords iff:

Otherwise the token is treated as an identifier.

Additionally:

Failure to meet these additional requirements makes the program ill-formed (rather than not being interpreted as a directive).

Status Quo This Paper
export
module x
;                      // ✅
// -Dm="export module x;"
m                      // ✅
module;
#define m x
export module m;       // ✅
module;
#if FOO
export module foo;     // ✅
#else
export module bar;     // ✅
#endif
module;
#define EMPTY
EMPTY export module m; // ✅
#if MODULES
module;
export module m;       // ❌
#endif
#if MODULES
export module m;       // ✅
#endif
export
module x
;                      // ⛔
// -Dm="export module x;"
m                      // ❌
module;
#define m x
export module m;       // ✅
module;
#if FOO
export module foo;     // ❌
#else
export module bar;     // ❌
#endif
module;
#define EMPTY
EMPTY export module m; // ❌
#if MODULES
module;
export module m;       // ❌
#endif
#if MODULES
export module m;       // ❌
#endif
module y = {};         // ❌

::import x = {};       // ✅
::module y = {};       // ✅

import::inner xi = {}; // ⛔
module::inner yi = {}; // ✅

namespace N {
  module a;            // ✅
  import b;            // ⛔
}

#define MAYBE_IMPORT(x) x
MAYBE_IMPORT(
  import <a>;          // ☠️
)
#define EAT(x)
EAT(
  import <a>;          // ☠️
)

void f(Import *import) {
  import->doImport();  // ⛔
}
module y = {};         // ⛔

::import x = {};       // ✅
::module y = {};       // ✅

import::inner xi = {}; // ✅
module::inner yi = {}; // ✅

namespace N {
  module a;            // ⛔
  import b;            // ⛔
}

#define MAYBE_IMPORT(x) x
MAYBE_IMPORT(
  import <a>;          // ☠️
)
#define EAT(x)
EAT(
  import <a>;          // ☠️
)

void f(Import *import) {
  import->doImport();  // ✅
}

Legend:

3. Motivation

3.1. Fast Dependency Scanning

Fast dependency scanning relies on partial preprocessing. Several people are working on fast dependency scanning for modules, and they all share the common trait of skipping over non-directives. clang-scan-deps is one such tool and is a representative motivating usecase for these changes.

clang-scan-deps works by minimizing source code via partial preprocessing. Partial preprocessing works by running phases 1-3 of translation. Minimization proceeds by throwing away every line which is not a preprocessing directive, then removing all preprocessing directives that can’t impact the set of dependencies for that file. Up to C++17 this means it only needs to keep #define, #undef, #if*, #else, #elif, #endif, and #include. And only the bodies of #ifs that contain a preprocessing directive that could impact dependencies. This reduces most files down to their header guards and list of #includes, and is correct for standard C++ except for abuse of __LINE__ such as #if __LINE__ > 456. The only other things that break it are compiler extensions which do not follow the parsing rules of # directives such as _Pragma("push_macro(\"X\")"), or .incbin in inline assembly. Additionally, abuse of the compiler extension __COUNTER__ can break this.

This minimization is context free, and is done once for each file in the entire build.

Then the minimized source file is fully preprocessed by running phases 1-4, with #includes resolving to their minimized equivalent, macros expanded, and all preprocessing directives executed.

Minimization followed by using clang’s full preprocessor currently provides about a 9x speed up (~3 seconds vs ~28 seconds out of a 15m build on a 18 core iMac Pro) over no minimization at scanning all of llvm and clang (~7k files and ~3.8m LoC). We do not have access to a large C++20 modules only (no header units) codebase to do tests on, but that case still requires scanning the entire file due to the potential for #include based x-macros which will not be going away with modules, as they are intentionally non-modular includes. There will still be a difference in that we no longer have normal includes, and thus don’t need to preprocess a header for each TU it’s used in.

Of the 3s only 33ms are spent minimizing, while the majority of the time is spent in the full lexer, preprocessor, and header search. We believe we can reduce this overhead. This overhead is important as it’s on the critical path. Any speedup here turns into a direct reduction in build latency, which is important for extremely parallel builds.

C++20 currently breaks this approach. [P1703r1] resolved the issue for import, but dependency scanning also needs to find module declarations, and the current rules are not enough.

3.2. import is Too Relaxed

[P1703r1] went too far in fixing the dependency scanning issue with import. Now any line starting with import is treated as an import directive, even if it obviously couldn’t be. This breaks a decent amount of real code with function arguments or local variables named import, as any use of them without a prefix is treated as an import directive, and prepending :: doesn’t work, you must wrap some part of the expression in ().

4. Discussion

4.1. Implementation

There were some concerns expressed during the discussion of [P1703r1] that there may be performance or complexity issues. I have implemented this fix in Clang. It was rather simple to implement, with the most complex part being doing token lookahead while lexing (which isn’t that complicated). Measurment did not show any performance impact, which I expect to be due to the codepath only occuring with the import and module tokens.

4.2. One Line Restriction

To be fully resiliant, a modules directive must be entirely on a single logical line. You could have a rule that the ; must not come from a macro and say that the directive extends to the next ;, but that is incorrect:
#define f(x) "blah" #x "blah"
import f(
  ;
);

4.3. Extra Dependencies

#define eat(x)
eat(
  import <a>;
);

In the WD this has undefined behaivor due to [cpp]/2

A preprocessing directive consists of a sequence of preprocessing tokens that satisfies the following constraints: The first token in the sequence, referred to as a directive-introducing token, is a # preprocessing token, an import preprocessing token, or an export preprocessing token immediately followed by an import preprocessing token, that (at the start of translation phase 4) either begins with the first character in the source file (optionally after white space containing no new-line characters) or follows white space containing at least one new-line character.

and [cpp.replace]/11

If there are sequences of preprocessing tokens within the list of [macro] arguments that would otherwise act as preprocessing directives, the behavior is undefined.

A minimizer will transform this into just import <a>; which when preprocessed will form a valid module import. There are two outcomes of this:

This is not the best outcome, as an implementation is not guaranteed to emit a diagnosic, from my tests of clang, msvc, gcc, and icc, only clang rejectes #include in this case, and no compiler rejects an import directive (I plan to fix this in clang as part of implementing P1703 and this paper). This should be fixed by changing [cpp.replace]/11 to be ill-formed instead of UB, but that deserves a separate paper. With that change though the possible outcomes would be:

Both of these outcomes are fine as their result is the same. If a partial preprocessing would ever get the dependencies wrong, the compiler will reject the program as ill-formed.

4.4. Ship Vehicle

This needs to ship with modules as this is a backwards incompatible change. Thus it needs to be in C++20.

5. Wording

5.1. [lex.pptoken]

preprocessing-token:
    header-name
    import-keyword
    module-keyword
    export-keyword
    ...
4 The import-keyword is produced by processing an import directive ([cpp.import]) , the module-keyword is produced by preprocessing a module directive ([cpp.module]), and the export-keyword is produced by preprocessing either of the previous two directives. and has no [Note: None have any associated grammar productions observable spelling . end note]

5.2. [lex.key]

Add import-keyword, module-keyword, and export-keyword to Table 5: Keywords [tab:lex.key].

translation-unit:
    top-level-declaration-seqopt
    global-module-fragmentopt module-declaration top-level-declaration-seqopt private-module-fragmentopt
private-module-fragment:
    modulemodule-keyword : private ; top-level-declaration-seqopt

...

3 A token sequence beginning with exportopt module and not immediately followed by :: is never interpreted as the declaration of a top-level-declaration.

5.4. [module.unit]

module-declaration:
    exportexport-keywordopt modulemodule-keyword module-name module-partitionopt attribute-specifier-seqopt ;

5.5. [module.import]

module-import-declaration:
    exportexport-keywordopt import-keyword module-name attribute-specifier-seqopt;
    exportexport-keywordopt import-keyword module-partition attribute-specifier-seqopt;
    exportexport-keywordopt import-keyword header-name attribute-specifier-seqopt;

5.6. [module.global]

global-module-fragment:
    modulemodule-keyword ; top-level-declaration-seqopt

5.7. [cpp.pre]

1 A preprocessing directive consists of a sequence of preprocessing tokens that satisfies the following constraints: At the start of translation phase 4, the The first token in the sequence, referred to as a directive-introducing token, begins with the first character in the source file (optionally after white space containing no new-line characters) or follows white space containing at least one new-line character, and is that (at the start of translation phase 4) either begins with the first character in the source file (optionally after white space containing no new-line characters) or follows white space containing at least one new-line character. The last token in the sequence is the first token in the sequence that is immediately followed by whitespace containing a new-line character that follows the first token in the sequence . [Note: A new-line character ends the preprocessing directive even if it occurs within what would otherwise be an invocation of a function-like macro. end note] [Example:
#                      // preprocessing directive
module ;               // preprocessing directive
export module leftpad; // preprocessing directive
import <string>;       // preprocessing directive
export import "squee"; // preprocessing directive
import rightpad;       // preprocessing directive
import :part;          // preprocessing directive

module                 // not a preprocessing directive
;                      // not a preprocessing directive

export                 // not a preprocessing directive
import                 // not a preprocessing directive
foo;                   // not a preprocessing directive

export                 // not a preprocessing directive
import foo;            // preprocessing directive (ill-formed at phase 7)

import ::              // not a preprocessing directive
import ->              // not a preprocessing directive
end example]

preprocessing-file:
    groupopt
    module-file

...

module-file:
    pp-global-module-fragmentopt pp-module groupopt pp-private-module-fragmentopt

pp-global-module-fragment:
    module ; new-line groupopt

pp-private-module-fragment:
    module : private ; new-line groupopt

control-line:
    # include pp-tokens new-line
    exportopt import pp-tokens new-line
    pp-import
    
    ...

2 A text line shall not begin with a # preprocessing token sequence of preprocessing tokens is only a text-line if it does not begin with a directive-introducing token . A conditionally-supported-directive shall not begin with any of the directive names appearing after a # in the syntax. A conditionally-supported-directive is conditionally-supported with implementation-defined semantics.

3At the start of phase 4 of translation the group of a pp-global-module-fragment shall neither contain a control-line not starting with a # preprocessing token nor a text-line.

5.8. [cpp.module]

Add a new section before [cpp.import]: Module directive [cpp.module]
pp-module:
    exportopt module pp-tokensopt ; new-line

1 A pp-module shall neither appear in a context where module is an identifier defined as an object-like macro nor where export is an identifier defined as an object-like macro if the first token of the pp-module is export.

2 Any preprocessing tokens after the module preprocessing token in the module directive are processed just as in normal text. [Note: Each identifier currently defined as a macro name is replaced by its replacement list of preprocessing tokens. —end note]

3 The module and export (if it exists) preprocessing tokens are replaced by the module-keyword and export-keyword preprocessing tokens respectively. [Note: This makes the line no longer a directive so it is not removed at the end of phase 4. —end note]

5.9. [cpp.import]

Insert a new paragraph immediately after the grammar.

A pp-import shall neither appear in a context where import is an identifier defined as an object-like macro nor where export is an identifier defined as an object-like macro if the first token of the pp-import is export.

Insert a new paragraph after paragraph 1

If an import-directive is produced by source file inclusion (including by the rewrite produced when a #include directive names an importable header) while processing the group of a module-file, the program is ill-formed.

Update paragraph 2

In all three forms of pp-import, the import and export (if it exists) token preprocessing tokens is are replaced by the import-keyword and export-keyword token preprocessing tokens respectively . [Note: This makes the line no longer a directive so it is not removed at the end of phase 4. —end note] Additionally, in the second form of pp-import, a header-name token is formed as if the header-name-tokens were the pp-tokens of a #include directive. The header-name-tokens are replaced by the header-name token. [Note: This ensures that imports are treated consistently by the preprocessor and later phases of translation. —end note]

5.10. [cpp.glob.frag]

Remove this section.
pp-global-module-fragment:
    module ; pp-balanced-token-seq module
    pp-balanced-token-seq:
        pp-balanced-token
        pp-balanced-token-seq pp-balanced-token
    pp-balanced-token:
        pp-ldelim pp-balanced-token-seqopt pp-rdelim
        any preprocessing-token other than a pp-ldelim or pp-rdelim
    pp-ldelim: one of
        (    [    {    <:    <%
    pp-rdelim: one of
        )    ]    }    :>    %>

1 If the first two preprocessing tokens at the start of phase 4 of translation are module ;, the result of preprocessing shall begin with a pp-global-module-fragment for which all preprocessing-tokens in the pp-balanced-token-seq were produced directly or indirectly by source file inclusion ([cpp.include]), and for which the second module preprocessing-token was not produced by source file inclusion or macro replacement ([cpp.replace]). Otherwise, the first two preprocessing tokens at the end of phase 4 of translation shall not be module ;.

5.11. [diff.cpp17.basic]

The editor should consider a different section of annex C to move this to.
Affected subclauses: [basic.link], [module.unit], and [module.import]
Change: New identifiers with special meaning.
Rationale: Required for new features.
Effect on original feature: Top-level declarations Logical lines beginning with module or import may be either ill-formed or interpreted differently in this International Standard. [ Example:
class module {};
module *m1;         // ill-formed; previously well-formed
::module *m2;       // OK
module m1;          // was variable declaration; now module-declaration
module *m1;         // OK

class import {}; import j1; // was variable declaration; now import-declaration ::import j2; // variable declaration

end example ]

6. Acknowledgments

Thanks to Boris Kolpackov for writing P1703, Mathias Stearn and Walter Brown for feedback on this paper and to Richard Smith, Corentin, Gabriel Dos Reis, Ben Craig, Matthew Woehlke, and Nathan Sidwell for discussion regarding this issue.

References

Informative References

[P1703r1]
Boris Kolpackov. Recognizing Header Unit Imports Requires Full Preprocessing. 19 July 2019. URL: https://wg21.link/p1703r1