Previously: When compiling for host, our constructed call graph went
*through* kernel calls. This meant that if we had
host calls kernel calls HD
we would incorrectly mark the HD function as known-emitted on the host
side, and thus perform host-side checks on it.
Fixing this exposed another issue, wherein when marking a function as
known-emitted, we also need to traverse the callgraph of its template,
because non-dependent calls are attached to a function's template, not
its instantiation.
llvm-svn: 284355
Summary:
Emitting deferred diagnostics during codegen was a hack. It did work,
but usability was poor, both for us as compiler devs and for users. We
don't codegen if there are any sema errors, so for users this meant that
they wouldn't see deferred errors if there were any non-deferred errors.
For devs, this meant that we had to carefully split up our tests so that
when we tested deferred errors, we didn't emit any non-deferred errors.
This change moves checking for deferred errors into Sema. See the big
comment in SemaCUDA.cpp for an overview of the idea.
This checking adds overhead to compilation, because we have to maintain
a partial call graph. As a result, this change makes deferred errors a
CUDA-only concept (whereas before they were a general concept). If
anyone else wants to use this framework for something other than CUDA,
we can generalize at that time.
This patch makes the minimal set of test changes -- after this lands,
I'll go back through and do a cleanup of the tests that we no longer
have to split up.
Reviewers: rnk
Subscribers: cfe-commits, rsmith, tra
Differential Revision: https://reviews.llvm.org/D25541
llvm-svn: 284158
Summary:
Together these let you easily create diagnostics that
- are never emitted for host code
- are always emitted for __device__ and __global__ functions, and
- are emitted for __host__ __device__ functions iff these functions are
codegen'ed.
At the moment there are only three diagnostics that need this treatment,
but I have more to add, and it's not sustainable to write code for emitting
every such diagnostic twice, and from a special wrapper in SemaCUDA.cpp.
While we're at it, don't emit the function name in
err_cuda_device_exceptions: It's not necessary to print it, and making
this work in the new framework in the face of a null value for
dyn_cast<FunctionDecl>(CurContext) isn't worth the effort.
Reviewers: rnk
Subscribers: cfe-commits, tra
Differential Revision: https://reviews.llvm.org/D25139
llvm-svn: 284143
Previously, this was an immediate, don't pass go, don't collect $200
error. But this precludes us from writing code like
__host__ __device__ void launch_kernel() {
kernel<<<...>>>();
}
Such code isn't wrong, following our notions of right and wrong in CUDA,
unless it's codegen'ed.
llvm-svn: 283963
match other CUDA preference orders, per discussion with jlebar. We now model
this in an attempt to match overload resolution as closely as possible:
- First, we throw out all non-callable (due to CUDA host/device mismatch)
operator delete functions.
- Then we apply sizedness / alignedness preferences based on whether the type
is overaligned and whether the deallocation function is a member.
- Finally, we use the CUDA callability preference as a tiebreaker.
llvm-svn: 283830
Summary: This matches the idiom we use for our other CUDA wrapper headers.
Reviewers: tra
Subscribers: beanz, mgorny, cfe-commits
Differential Revision: https://reviews.llvm.org/D24978
llvm-svn: 283679
Summary:
Move CheckCUDACall from ActOnCallExpr and BuildDeclRefExpr to
DiagnoseUseOfDecl. This lets us catch some edge cases we were missing,
specifically around class operators.
This necessitates a few other changes:
- Avoid emitting duplicate deferred diags in CheckCUDACall.
Previously we'd carefully placed our call to CheckCUDACall such that
it would only ever run once for a particular callsite. But now this
isn't the case.
- Emit deferred diagnostics from a template
specialization/instantiation's primary template, in addition to from
the specialization/instantiation itself. DiagnoseUseOfDecl ends up
putting the deferred diagnostics on the template, rather than the
specialization, so we need to check both.
Reviewers: rsmith
Subscribers: cfe-commits, tra
Differential Revision: https://reviews.llvm.org/D24573
llvm-svn: 283637
Summary:
We'd attempted to allow this, but turns out we were doing a very bad
job. :)
Making this work properly would be a giant change in clang. For
example, we'd need to make CXXRecordDecl::getDestructor()
context-sensitive, because the destructor you end up with depends on
where you're calling it from.
For now (and hopefully for ever), just disallow overloading of
destructors in CUDA.
Reviewers: rsmith
Subscribers: cfe-commits, tra
Differential Revision: https://reviews.llvm.org/D24571
llvm-svn: 283120
Also add a test that we disallow
__constant__ __shared__ int x;
because it's possible to break this without breaking
__shared__ __constant__ int x;
Reviewers: rnk
Subscribers: cfe-commits, tra
Differential Revision: https://reviews.llvm.org/D25125
llvm-svn: 282985
__attribute__((amdgpu_flat_work_group_size(<min>, <max>))) - request minimum and maximum flat work group size
__attribute__((amdgpu_waves_per_eu(<min>[, <max>]))) - request minimum and/or maximum waves per execution unit
Differential Revision: https://reviews.llvm.org/D24513
llvm-svn: 282371
Summary: This functionality is used by Thrust.
Reviewers: tra
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D24581
llvm-svn: 281543
CUDA target attributes are used for function overloading and must not be merged.
This fixes a bug where attributes were inherited during function template
specialization in CUDA and made it impossible for specialized function
to provide its own target attributes.
Differential Revision: https://reviews.llvm.org/D24522
llvm-svn: 281406
Summary:
Some function calls in CUDA are allowed to appear in
semantically-correct programs but are an error if they're ever
codegen'ed. Specifically, a host+device function may call a host
function, but it's an error if such a function is ever codegen'ed in
device mode (and vice versa).
Previously, clang made no attempt to catch these errors. For the most
part, they would be caught by ptxas, and reported as "call to unknown
function 'foo'".
Now we catch these errors and report them the same as we report other
illegal calls (e.g. a call from a host function to a device function).
This has a small change in error-message behavior for calls that were
previously disallowed (e.g. calls from a host to a device function).
Previously, we'd catch disallowed calls fairly early, before doing
additional semantic checking e.g. of the call's arguments. Now we catch
these illegal calls at the very end of our semantic checks, so we'll
only emit a "illegal CUDA call" error if the call is otherwise
well-formed.
Reviewers: tra, rnk
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D23242
llvm-svn: 278759
Summary:
When we emit err_ref_bad_target, we should emit a "'method' declared
here" note. We already do so in most places, just not in
BuildCallToMemberFunction.
Reviewers: tra
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D23240
llvm-svn: 278195
Summary:
Specifically, this patch adds testcases for all three calls to
EraseUnwantedCUDAMatches. The addr-of-overloaded-fn test I accidentally
neutered in r264207, which moved much of
CodeGenCUDA/function-overload.cu into SemaCUDA/function-overload.cu.
The coverage from overloaded-delete test is new.
Reviewers: tra
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D21913
llvm-svn: 275232
Summary:
You can overload a destructor in CUDA, and SemaOverload needs to be
tweaked not to crash when it sees an explicit call to an overloaded
destructor.
Reviewers: rsmith
Subscribers: cfe-commits, tra
Differential Revision: http://reviews.llvm.org/D21912
llvm-svn: 275231
Some people have weird CI systems that run each test subdirectory
independently without access to other parallel trees.
Unfortunately, this means we have to suffer some duplication until Art
can sort out how to share these types.
llvm-svn: 270164
According to Cuda Programming guide (v7.5, E2.3.1):
> __device__, __constant__ and __shared__ variables defined in namespace
> scope, that are of class type, cannot have a non-empty constructor or a
> non-empty destructor.
Clang already deals with device-side constructors (see D15305).
This patch enforces similar rules for destructors.
Differential Revision: http://reviews.llvm.org/D20140
llvm-svn: 270108
Codegen tests for device-side variable initialization are subset of test
cases used to verify Sema's part of the job.
Including CodeGenCUDA/device-var-init.cu from SemaCUDA makes it easier to
keep both sides in sync.
Differential Revision: http://reviews.llvm.org/D20139
llvm-svn: 270107
a base class via a using-declaration. If a class has a using-declaration
declaring either a constructor or an assignment operator, eagerly declare its
special members in case they need to displace a shadow declaration from a
using-declaration.
llvm-svn: 269398
With this patch, by a constexpr function is implicitly host+device
unless:
a) it's a variadic function (variadic functions are not allowed on the
device side), or
b) it's preceeded by a __device__ overload in a system header.
The restriction on overloading __host__ __device__ functions on the
basis of their CUDA attributes remains in place, but we use (b) to allow
us to define __device__ overloads for constexpr functions in cmath,
which would otherwise be __host__ __device__ and thus not overloadable.
You can disable this behavior with -fno-cuda-host-device-constexpr.
Reviewers: tra, rnk, rsmith
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D18380
llvm-svn: 264964
Summary:
* -fcuda-target-overloads
Previously unconditionally set to true by the driver. Necessary for
correct functioning of the compiler -- our CUDA headers wrapper won't
compile without this.
* -fcuda-disable-target-call-checks
Previously unconditionally set to true by the driver. Necessary to
compile almost any external CUDA code -- almost all libraries assume
that host+device code can call host or device functions.
* -fcuda-allow-host-calls-from-host-device
No effect when target overloading is enabled.
Reviewers: tra
Subscribers: rsmith, cfe-commits
Differential Revision: http://reviews.llvm.org/D18416
llvm-svn: 264739
Summary:
Previously we were using the codegen test to ensure that we choose the
right overload. But we can do this within sema, with a bit of
cleverness.
I left the constructor/destructor checks in CodeGen, because these
overloads (particularly on the destructors) are hard to check in Sema.
Reviewers: tra
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D18386
llvm-svn: 264207
Summary:
Principally, don't hardcode the line numbers of various notes. This
lets us make changes to the test without recomputing linenos everywhere.
Instead, just tell -verify that we may get 0 or more notes pointing to
the relevant function definitions. Checking that we get exactly the
right note isn't so important (and anyway is checked elsewhere).
Reviewers: tra
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D18385
llvm-svn: 264206
__global__ functions are present on both host and device side,
so providing __host__ or __device__ overloads is not going to
do anything useful.
llvm-svn: 261778
This is an artefact of split-mode CUDA compilation that we need to
mimic. HD functions are sometimes allowed to call H or D functions. Due
to split compilation mode device-side compilation will not see host-only
function and thus they will not be considered at all. For clang both H
and D variants will become function overloads visible to
compiler. Normally target attribute is considered only if C++ rules can
not determine which function is better. However in this case we need to
ignore functions that would not be present during current compilation
phase before we apply normal overload resolution rules.
Changes:
* introduced another level of call preference to better describe
possible call combinations.
* removed WrongSide functions from consideration if the set contains
SameSide function.
* disabled H->D, D->H and G->H calls. These combinations are
not allowed by CUDA and we were reluctantly allowing them to work
around device-side calls to math functions in std namespace.
We no longer need it after r258880.
Differential Revision: http://reviews.llvm.org/D16870
llvm-svn: 260697
Summary:
Turns out the variadic function checking added in r258643 was too strict
for some existing users; give them an escape valve. When
-fcuda-allow-variadic-functions is passed, the front-end makes no
attempt to disallow C-style variadic functions. Calls to va_arg are
still not allowed.
Reviewers: tra
Subscribers: cfe-commits, jhen, echristo, bkramer
Differential Revision: http://reviews.llvm.org/D16559
llvm-svn: 258822
Summary:
Warn for NVCC compatibility if you declare a static member function or
inline function as __global__.
Reviewers: tra
Subscribers: jhen, echristo, cfe-commits
Differential Revision: http://reviews.llvm.org/D16261
llvm-svn: 258263
Value, type, and instantiation dependence were not being handled
correctly for CUDAKernelCallExpr AST nodes. As a result, if an
undeclared identifier was used in the triple-angle-bracket kernel call
configuration, there would be no error during parsing, and there would
be a crash during code gen. This patch makes sure that an error will be
issued during parsing in this case, just as there would be for any other
use of an undeclared identifier in C++.
Patch by Jason Henline.
Reviewers: jlebar, rsmith
Differential Revision: http://reviews.llvm.org/D15858
llvm-svn: 257839
Summary: Thanks to jhen for helping me figure this out.
Reviewers: tra, echristo
Subscribers: jhen
Differential Revision: http://reviews.llvm.org/D16129
llvm-svn: 257554
Currently, when there is a global register variable in a program that
is bound to an invalid register, clang/llvm prints an error message that
is not very user-friendly.
This commit improves the diagnostic and moves the check that used to be
in the backend to Sema. In addition, it makes changes to error out if
the size of the register doesn't match the declared variable size.
e.g., volatile register int B asm ("rbp");
rdar://problem/23084219
Differential Revision: http://reviews.llvm.org/D13834
llvm-svn: 253405
Clang needs to know target triple for both sides of compilation so that
preprocessor macros and target builtins from both sides are available.
This change augments Compilation class to carry information about
toolchains used during different CUDA compilation passes and refactors
BuildActions to use it when it constructs CUDA jobs.
Removed DeviceTriple from CudaHostAction/CudaDeviceAction as it's no
longer needed.
Differential Revision: http://reviews.llvm.org/D13144
llvm-svn: 253385