[patch] emit .weak for zero initialized weak variables

Project Infrastructure LLVM Dev List Archives
#1

The attached patches makes all backends print a .weak directive for
zero initialized variables. GCC does it on x86-64 and ARM. I assume
that it should be done on all architectures.

The patch also adds a test to the ARM backend? Should I copy it to all backends?

Comments?

Thanks,
Rafael

P.S.: I find the doFinalization code a bit confusing. Would you
welcome a cleanup patch?

llvm.patch (5.95 KB)

#2

I'm not sure this is right. C code like this:

int X;

should compile to a .comm directive, not a .weak directive, right?

-Chris

#3

I'm not sure this is right. C code like this:

int X;

should compile to a .comm directive, not a .weak directive, right?

right.

According to GCC, "int x" should compile to ".comm x", but "int x
__attribute__((weak))" should compile to

#4

Is there a semantic difference between the two? Just because GCC compiles them to different things, it doesn't mean they do different things...

-Chris

#5

Is there a semantic difference between the two? Just because GCC compiles
them to different things, it doesn't mean they do different things...

Sorry for the long delay...

On a x86 box I am able to use both declarations interchangeably. There
is a problem if we have both a declaration and an use on the same
file: We must print a ".weak" because of the use, but if we printed a
".comm" for the declaration, the assembler will reject the file:

Error: symbol `a' can not be both weak and common

Best Regards,
Rafael

#6

Error: symbol `a' can not be both weak and common

Actually this error happens only with an old version of my patch that
prints ".weak" for both weak and extern_weak symbols (similar to gcc).

Sorry for the confusion.

Rafael