Incorrect clang-format pointer alignment when variables are declared in conditionals

[JohnC32]

Given:

void foo1(const T1* p1) {
    if (T2* v; check(&v, p1)) {
        goo(v);
    }
}

and _clang-format:

---
BasedOnStyle: Google
IndentWidth: 4
# Google style default is PointerAlignment: Left,
PointerAlignment: Left
...

clang-format 15 yields the following. Notice the pointer alignment is no longer left, despite
asking for it to be left.

void foo1(const T1* p1) {
    if (T2 * v; check(&v, p1)) {
        goo(v);
    }
}

In this case, we scoped the variable declaration in the if conditional. If we move it out of the
conditional, pointer alignment is as expected:

void foo2(const T1* p1) {
    T2* v;
    if (check(&v, p1)) {
        goo(v);
    }
}

[llvmbot]

@llvm/issue-subscribers-clang-format

[jeremyong]

I'm seeing this also with an even simpler repro:

for (int i = 0; Foo& foo : foos)

turns into

for (int i = 0; Foo & foo : foos)

even with PointerAlignment: Left. I am assuming that in the presence of scoped declaration statements, subsequent & operators are parsed as bitwise-and operators.

[swalkner]

I can confirm this behaviour and would like to gently bump this report.

if (MyClass * obj{GetMyClass()})

But - for some reason - auto and primitive-types seem to work as expected:

if (auto* obj{GetMyClass()})

if (int* obj{GetInt()})

I have reproduced that in clang-format:

• 17.0.3 (packaged with VS)
• 18.1.3 (and other versions via clang-format-configurator.site)

The only config option I changed:

PointerAlignment: Left

[CDAlexanderW]

Is there a chance for a fix for this?
@owenca

[owenca]

We have the TypeNames option for user-defined types when clang-format can't tell if an identifier before */&/&& is a type or variable:

$ clang-format -version
clang-format version 19.1.0
$ cat foo.cpp
if (T2 *v; check(&v, p1))
  ;
for (int i = 0; Foo &foo : foos)
  ;
if (MyClass *obj{GetMyClass()})
  ;
$ cat .clang-format
TypeNames: [T2, Foo, MyClass]
$ clang-format foo.cpp | diff foo.cpp -
$ 

[CDAlexanderW]

We have the TypeNames option for user-defined types when clang-format can't tell if an identifier before */&/&& is a type or variable:

$ clang-format -version
clang-format version 19.1.0
$ cat foo.cpp
if (T2 *v; check(&v, p1))
  ;
for (int i = 0; Foo &foo : foos)
  ;
if (MyClass *obj{GetMyClass()})
  ;
$ cat .clang-format
TypeNames: [T2, Foo, MyClass]
$ clang-format foo.cpp | diff foo.cpp -
$ 

So just to be sure. Technically we dont have any better option besides providing this additional TypeNames? For very large code bases with millions of types (although not all used in such cases but still a lot) it may not be possible to provide all the types. For this we mas have to live with it, correct?

Thx!

[owenca]

Technically we dont have any better option besides providing this additional TypeNames?

In general, TypeNames can be used if there are no better alternatives. (See https://reviews.llvm.org/D155273 for some background.) Nevertheless, I'll see if something similar to #109229 can be done here.

[owenca]

I'm seeing this also with an even simpler repro:

for (int i = 0; Foo& foo : foos)

turns into

for (int i = 0; Foo & foo : foos)

even with PointerAlignment: Left. I am assuming that in the presence of scoped declaration statements, subsequent & operators are parsed as bitwise-and operators.

This is related but different. I have moved it to #109358.

[owenca]

I can confirm this behaviour and would like to gently bump this report.

if (MyClass * obj{GetMyClass()})

But - for some reason - auto and primitive-types seem to work as expected:

if (auto* obj{GetMyClass()})

if (int* obj{GetInt()})

I have reproduced that in clang-format:

• 17.0.3 (packaged with VS)
• 18.1.3 (and other versions via clang-format-configurator.site)

The only config option I changed:

PointerAlignment: Left

This is different as there is no semicolon. Please split it to a new issue or use the TypeNames option.

[swalkner]

Thanks so much for your reply. I created #109371