attributes on type in new-expression are silently ignored

[llvmbot]

Bugzilla Link	47877
Version	trunk
OS	Linux
Reporter	LLVM Bugzilla Contributor
CC	@AaronBallman,@DougGregor,@zygoloid

Extended Description

Hi, the following code allocates memory for an integer variable a with 5, which is not an integral power of two. However, Clang accepts it.

Please take a look at paragraph [4]:
http://eel.is/c++draft/basic.align

$ cat s.cpp
void * a = new int __attribute__ ((aligned (5)))();

$ g++ -c s.cpp
s.cpp:1:48: error: requested alignment ‘5’ is not a positive power of 2
    1 | void * a = new int __attribute__ ((aligned (5)))();

[zygoloid]

I suspect attributes in this position are being ignored entirely.

[zygoloid]

Clang also silently accepts:

void * a = new int __attribute__ ((goldfish(fruitbasket)))();

[AaronBallman]

We are sort of ignoring the attributes. We parse the attributes, but explicitly diagnose the C++ attributes as we don't currently support any type attributes using them:

llvm-project/clang/lib/Parse/ParseDecl.cpp

Line 3015 in 702529d

ProhibitCXX11Attributes(attrs, diag::err_attribute_not_type_attr);

But we currently have a FIXME that we should be diagnosing this case too:

llvm-project/clang/lib/Parse/ParseDecl.cpp

Line 2404 in 702529d

/// TODO: diagnose attribute-specifiers and alignment-specifiers.

As best I can tell, we basically don't support attribute((aligned)) as a type attribute at all: https://godbolt.org/z/PzTxWh and I'm not certain GCC does either: https://godbolt.org/z/6xrxvd

[zygoloid]

The alignments you picked in your testing all happened to be <= the default new alignment. For a larger alignment (https://godbolt.org/z/EcMaKn) GCC calls the aligned form of operator new.

[AaronBallman]

It figures I managed to pick all the wrong alignments to test that out. Thank you, Richard.

So it sounds like the real issue here is that we don't support __attribute__((aligned)) as a type attribute. Once we start doing that implementation work, we'll have to fix up the diagnostics in some places.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: None (llvmbot)

Details

| | | | --- | --- | | Bugzilla Link | [47877](https://llvm.org/bz47877) | | Version | trunk | | OS | Linux | | Reporter | LLVM Bugzilla Contributor | | CC | @AaronBallman,@DougGregor,@zygoloid |

Extended Description

Hi, the following code allocates memory for an integer variable a with 5, which is not an integral power of two. However, Clang accepts it.

Please take a look at paragraph [4]:
http://eel.is/c++draft/basic.align

$ cat s.cpp
void * a = new int __attribute__ ((aligned (5)))();

$ g++ -c s.cpp
s.cpp:1:48: error: requested alignment ‘5’ is not a positive power of 2
    1 | void * a = new int __attribute__ ((aligned (5)))();

[shafik]

CC @erichkeane