llvm-project
llvm-project copied to clipboard
static __thread T* ptr(0); error: initializer for thread-local variable must be a constant expression (while T* ptr = 0 will work)
| Bugzilla Link | 21690 |
| Version | trunk |
| OS | Linux |
| Reporter | LLVM Bugzilla Contributor |
| CC | @hfinkel |
Extended Description
I bumped Clang to pre-3.6 (65d8b4c4998b3a0c20934ea72ede72ef4838a004, r222081, ~15 Nov). Have not checked the trunk version.
$ cat test.cc
template <typename T>
void dummy() {
static __thread T* ptr(0);
}
$ clang++ -std=c++11 test.cc
test.cc:3:25: error: initializer for thread-local variable must be a constant expression
static __thread T* ptr(0);
^~~
test.cc:3:22: note: use 'thread_local' to allow this
static __thread T* ptr(0);
^
1 error generated.
The following will work:
static __thread T* ptr;
static __thread T* ptr = 0;
static __thread T* ptr = nullptr;
static thread_local T* ptr(nullptr);
The following will not work:
static __thread T* ptr{0};
static __thread T* ptr(0);
It works with c-like initialization, but not with constructor initialization or uniform initialization.
We are assigning initial value of pointer (nullptr), isn't nullptr/0 are constant expression here?
Reconfirmed on trunk with following:
template <typename T>
void dummy() {
static __thread T* ptr(0);
}
I am looking at where this diagnostic is generated:
if (var->getType().isDestructedType()) {
// GNU C++98 edits for __thread, [basic.start.term]p3:
// The type of an object with thread storage duration shall not
// have a non-trivial destructor.
Diag(var->getLocation(), diag::err_thread_nontrivial_dtor);
if (getLangOpts().CPlusPlus11)
Diag(var->getLocation(), diag::note_use_thread_local);
} else if (getLangOpts().CPlusPlus && var->hasInit()) {
if (!checkConstInit()) {
// GNU C++98 edits for __thread, [basic.start.init]p4:
// An object of thread storage duration shall not require dynamic
// initialization.
// FIXME: Need strict checking here.
Diag(CacheCulprit->getExprLoc(), diag::err_thread_dynamic_init)
<< CacheCulprit->getSourceRange();
if (getLangOpts().CPlusPlus11)
Diag(var->getLocation(), diag::note_use_thread_local);
}
}
and I think this looks like a bug but I would have do more digging.
CC @zygoloid
We shouldn't be trying to check if a dependent variable/initializer is constant.
Whether this check makes sense at all is a separate issue.
The other calls to checkConstInit are guarded by:
if (getLangOpts().CPlusPlus && !type->isDependentType() && Init &&
!Init->isValueDependent() &&
... some of which seem to be missing here. The isDestructedType check for __thread also should be skipped for a variable with a dependent type, but we seem to get away with that because we seem to treat dependent types as trivially-destructed.