[clang] Warning for incomplete array initializer lists

[Sinitax]

Summary

This patch implements a warning for incomplete array initializer lists.

Without warnings a 'gap' in an array can go unnoticed when the size increases. Consider what happens with the array const char *err_msg[ERR_TYPE_COUNT] = { ... } when a new ERR_.. enum value is added.

Related: https://github.com/llvm/llvm-project/issues/22692

[Sinitax]

Remove new warning from -Wuninitialized group for backwards-compatibility.

[Sinitax]

Add new warning to its own group.

[Sinitax]

Apply git-clang-format.

[shafik]

Adding reviewer for more visibility

[philnik]

clang/include/clang/Basic/DiagnosticGroups.td

732

Why do you make this a new warning group instead of just adding it to -Wuninitialized?

clang/lib/Sema/SemaInit.cpp

983

You'll want to add tests for this in clang/test/Sema

[erichkeane]

Needs release notes and tests. Patch needs full-context added.

clang/include/clang/Basic/DiagnosticGroups.td

732

I think the new group makes sense, but it should be a part of Uninitialized as well (see 733).

clang/lib/Sema/SemaInit.cpp

991

should probably just do range-for here, use the .inits here instead. Perhaps even better is a llvm::find of nullptr.

[Sinitax]

Rename warning to -Wincomplete-array-initializer-list for clarity, and add test case.

[Sinitax]

Rebased changes onto commit tagged 'llvmorg-17-init', and added release notes.

[Sinitax]

Why do you make this a new warning group instead of just adding it to -Wuninitialized?

I think a separate group is best, since the warning emits false positives: an intentional empty-initialization (such as char buf[BUF_SIZ] = { 0 }) is indistinguishable from unintentionally incomplete initialization (such as int pair[2] = { ENUM_A /*, ENUM_B */ }).

For this reason users would want to selectively enable it..

#pragma clang diagnostic push
#pragma clang diagnostic warning "-Wincomplete-initializer-list"

const char *errmsg[ERR_COUNT] = {
    [ERR_A] = "error a",
};

#pragma clang diagnostic pop

.. but I'm not really satisfied with this solution. Feedback on how to differentiate between the aforementioned cases, if possible, is appreciated.

[Sinitax]

Rebase patch against branch main.

[Sinitax]

Fix test.

[cor3ntin]

Can you test in c++ mode?

I'd like to see a test for something like

struct {
	int a, b, c, d  = 1; // d has an initializer
} bar2 = {
	.a = 1,
	.b = 1,
	.c = 1
};

Otherwise looks good