[C++20] [Modules] Regression module declaration must occur at the start of the translation unit

[ChuanqiXu9]

Reproducer: https://godbolt.org/z/eP84dnKGv

# 1 __FILE__ 1 3
export module a;

I believe this should be accepted no matter what spec says.

[ChuanqiXu9]

CC @yronglin

[llvmbot]

@llvm/issue-subscribers-clang-modules

Author: Chuanqi Xu (ChuanqiXu9)

Reproducer: https://godbolt.org/z/eP84dnKGv

# 1 __FILE__ 1 3
export module a;

I believe this should be accepted no matter what spec says.

[yronglin]

Hmm, according to the wording in P1857R3:

A module directive may only appear as the first preprocessing tokens in a file (excluding the global module fragment.)

and the wording in [cpp.pre]

module-file:
    pp-global-module-fragment[opt] pp-module group[opt] pp-private-module-fragment[opt]

IIUC, # is the first pp-token in the file, so it was rejected by clang. Should line markers be exempted from this rule?

CC @Bigcheese

[Bigcheese]

Hmm, yeah, they really should be exempted from this rule. The goal is to not allow any preprocessor conditionals or most state changes, but these don't fit that.

This line format is an extension so we should just allow them. I'll additionally file and NB comment and write a paper to allow #line.

[yronglin]

Thanks, I'll be happy to follow up on the new rule.

[shafik]

CC @mpark

[AaronBallman]

Hmm, yeah, they really should be exempted from this rule. The goal is to not allow any preprocessor conditionals or most state changes, but these don't fit that.

What kind of state changes are problematic? GNU line markers can change states FWIW (particularly, they change "this is a system header" kind of state).

[boris-kolpackov]

BTW, the linked #151985 show that this issue precludes Clang from compiling its own -frewrite-includes output.

[AaronBallman]

Do we think we can get this one fixed for the Clang 21.x release? (Should this be a release blocker, @Bigcheese @ChuanqiXu9 ?)

[Bigcheese]

This is pretty important to fix as there are a number of tools that rely on -E/-frewrite-includes output.

State change would mean most semantically observable preprocessor state, particularly anything that is order dependent. Global flags like being a system header/module shouldn't matter. Some #pragmas would also be fine, but I'm not aware of any that couldn't also come after the module;.

I'm going to be out for the next few days, but I can review a change here. @yronglin would you be able to do a patch for that? I don't have formal wording for the standard yet, but just allowing things that Clang identifies as a line markers is what we should do for now.