When clang emits an inheriting C++ constructor it may inline code
during the CodeGen phase. This patch ensures that any debug info in
this inlined code gets a proper inlined location. Otherwise we can end
up with invalid debug info metadata, since all inlined local variables
and function arguments would be reparented into the call site.
Analogous to ApplyInlineLocation this patch introduces a
ApplyInlineDebugLocation scoped helper to facilitate entering an
inlined scope and cleaning up afterwards.
This fixes one of the issues discovered in PR32042.
rdar://problem/30679307
llvm-svn: 296388
Essentially, as a base class constructor does not construct virtual bases, such
a constructor for an abstract class does not need the corresponding base class
construction to be valid, and likewise for destructors.
This creates an awkward situation: clang will sometimes generate references to
the complete object and deleting destructors for an abstract class (it puts
them in the construction vtable for a derived class). But we can't generate a
"correct" version of these because we can't generate references to base class
constructors any more (if they're template specializations, say, we might not
have instantiated them and can't assume any other TU will emit a copy).
Fortunately, we don't need to, since no correct program can ever invoke them,
so instead emit symbols that just trap.
We should stop emitting references to these symbols, but still need to emit
definitions for compatibility.
llvm-svn: 296275
2nd attempt: the first was in r296231, but it had a use after lifetime
bug.
Clang has logic to lower certain conditional expressions directly into llvm
select instructions. However, it does not emit the correct profile counter
increment as it does this: it emits an unconditional increment of the counter
for the 'then branch', even if the value selected is from the 'else branch'
(this is PR32019).
That means, given the following snippet, we would report that "0" is selected
twice, and that "1" is never selected:
int f1(int x) {
return x ? 0 : 1;
^2 ^0
}
f1(0);
f1(1);
Fix the problem by using the instrprof_increment_step intrinsic to do the
proper increment.
llvm-svn: 296245
Clang has logic to lower certain conditional expressions directly into
llvm select instructions. However, it does not emit the correct profile
counter increment as it does this: it emits an unconditional increment
of the counter for the 'then branch', even if the value selected is from
the 'else branch' (this is PR32019).
That means, given the following snippet, we would report that "0" is
selected twice, and that "1" is never selected:
int f1(int x) {
return x ? 0 : 1;
^2 ^0
}
f1(0);
f1(1);
Fix the problem by using the instrprof_increment_step intrinsic to do
the proper increment.
llvm-svn: 296231
Teach ubsan to diagnose remainder operations which have undefined
behavior due to signed overflow (e.g INT_MIN % -1).
Differential Revision: https://reviews.llvm.org/D29437
llvm-svn: 296214
C requires the operands of arithmetic expressions to be promoted if
their types are smaller than an int. Ubsan emits overflow checks when
this sort of type promotion occurs, even if there is no way to actually
get an overflow with the promoted type.
This patch teaches clang how to omit the superflous overflow checks
(addressing PR20193).
Testing: check-clang and check-ubsan.
Differential Revision: https://reviews.llvm.org/D29369
llvm-svn: 296213
The goal of this is to fix a bug in modules where we'd merge
FunctionDecls that differed in their pass_object_size attributes. Since
we can overload on the presence of pass_object_size attributes, this
behavior is incorrect.
We don't represent `N` in `pass_object_size(N)` as part of
ExtParameterInfo, since it's an error to overload solely on the value of
N. This means that we have a bug if we have two modules that declare
functions that differ only in their pass_object_size attrs, like so:
// In module A, from a.h
void foo(char *__attribute__((pass_object_size(0))));
// In module B, from b.h
void foo(char *__attribute__((pass_object_size(1))));
// In module C, in main.c
#include "a.h"
#include "b.h"
At the moment, we'll merge the foo decls, when we should instead emit a
diagnostic about an invalid overload. We seem to have similar (silent)
behavior if we overload only on the return type of `foo` instead; I'll
try to find a good place to put a FIXME (or I'll just file a bug) soon.
This patch also fixes a bug where we'd not output the proper extended
parameter info for declarations with pass_object_size attrs.
llvm-svn: 296076
Fix the fact that we don't assign profile counters to constructors in
classes with virtual bases, or constructors with variadic parameters.
Differential Revision: https://reviews.llvm.org/D30131
llvm-svn: 296062
This patch makes use of the prefix/suffix ABI argument distinction that
was introduced in r295870, so that we now emit ExtParameterInfo at the
correct offset for member calls that have added ABI arguments. I don't
see a good way to test the generated param info, since we don't actually
seem to use it in CGFunctionInfo outside of Swift. Any
suggestions/thoughts for how to better test this are welcome. :)
This patch also fixes a small bug with inheriting constructors: if we
decide not to pass args into an base class ctor, we would still
generate ExtParameterInfo as though we did. The added test-case is for
that behavior.
llvm-svn: 296024
This fixes an assertion failure in cases where we had expression
statements that declared variables nested inside of pass_object_size
args. Since we were emitting the same ExprStmt twice (once for the arg,
once for the @llvm.objectsize call), we were getting issues with
redefining locals.
This also means that we can be more lax about when we emit
@llvm.objectsize for pass_object_size args: since we're reusing the
arg's value itself, we don't have to care so much about side-effects.
llvm-svn: 295935
Summary: We implement structured exception handling (SEH) by generating filter functions for functions that use exceptions. Currently, we use associative comdats to ensure that the filter functions are preserved if and only if the functions we generated them for are preserved. This can lead to problems when generating COFF objects - LLVM may decide to inline a function that uses SEH and remove its body, at which point we will end up with a comdat that COFF cannot represent. To avoid running into that situation, this change makes us not use associative comdats for SEH filter functions. We can still get the benefits we used the associative comdats for: we will always preserve filter functions we use, and dead stripping can eliminate the ones we don't use.
Reviewers: rnk, pcc, ruiu
Reviewed By: rnk
Differential Revision: https://reviews.llvm.org/D30117
llvm-svn: 295872
Meta: The ultimate goal is to teach ExtParameterInfo about
pass_object_size attributes. This is necessary for that, since our
ExtParameterInfo is a bit buggy in C++. I plan to actually make use of
this Prefix/Suffix info in the near future, but I like small
single-purpose changes. Especially when those changes are hard to
actually test...
At the moment, some of our C++-specific CodeGen pretends that ABIs can
only add arguments to the beginning of a function call. This isn't quite
correct: args can be appended to the end, as well. It hasn't mattered
much until now, since we seem to only use this "number of arguments
added" data when calculating the ExtParameterInfo to use when making a
CGFunctionInfo. Said ExtParameterInfo is currently only used for
ParameterABIs (Swift) and ns_consumed (ObjC).
So, this patch allows ABIs to indicate whether args they added were at
the beginning or end of an argument list. We can use this information to
emit ExtParameterInfos more correctly, though like said, that bit is
coming soon.
No tests since this is theoretically a nop.
llvm-svn: 295870
The following code would crash clang:
void foo(unsigned *const __attribute__((pass_object_size(0))));
void bar(unsigned *i) { foo(i); }
This is because we were always selecting the version of
`@llvm.objectsize` that takes an i8* in CodeGen. Passing an i32* as an
i8* makes LLVM very unhappy.
(Yes, I'm surprised that this remained uncaught for so long, too. :) )
As an added bonus, we'll now also use the appropriate address space when
emitting @llvm.objectsize calls.
llvm-svn: 295805
declaration declared using class template argument deduction.
Patch by Eric Fiselier (who is busy and asked me to commit this on his behalf)!
Differential Revision: https://reviews.llvm.org/D30082
llvm-svn: 295794
Summary: AddDiscriminator pass is only useful for sample pgo. This patch restricts AddDiscriminator to -fdebug-info-for-profiling so that it does not introduce unecessary debug size increases for non-sample-pgo builds.
Reviewers: dblaikie, aprantl
Reviewed By: dblaikie
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D30220
llvm-svn: 295764
This patch teaches ubsan to insert exactly one null check for the 'this'
pointer per method/lambda.
Previously, given a load of a member variable from an instance method
('this->x'), ubsan would insert a null check for 'this', and another
null check for '&this->x', before allowing the load to occur.
Similarly, given a call to a method from another method bound to the
same instance ('this->foo()'), ubsan would a redundant null check for
'this'. There is also a redundant null check in the case where the
object pointer is a reference ('Ref.foo()').
This patch teaches ubsan to remove the redundant null checks identified
above.
Testing: check-clang, check-ubsan, and a stage2 ubsan build.
I also compiled X86FastISel.cpp with -fsanitize=null using
patched/unpatched clangs based on r293572. Here are the number of null
checks emitted:
-------------------------------------
| Setup | # of null checks |
-------------------------------------
| unpatched, -O0 | 21767 |
| patched, -O0 | 10758 |
-------------------------------------
Changes since the initial commit:
- Don't introduce any unintentional object-size or alignment checks.
- Don't rely on IRGen of C labels in the test.
Differential Revision: https://reviews.llvm.org/D29530
llvm-svn: 295515
CodeGenFunction::EmitTypeCheck accepts a bool flag which controls
whether or not null checks are emitted. Make this a bit more flexible by
changing the bool to a SanitizerSet.
Needed for an upcoming change which deals with a scenario in which we
only want to emit null checks.
llvm-svn: 295514
This reverts commit r295401. It breaks the ubsan self-host. It inserts
object size checks once per C++ method which fire when the structure is
empty.
llvm-svn: 295494
With tasks, the cancel may happen in another task. This has a different
region info which means that we can't find it here.
Differential Revision: https://reviews.llvm.org/D30091
llvm-svn: 295474
This resolves a deadlock with the cancel directive when there is no explicit
cancellation point. In that case, the implicit barrier acts as cancellation
point. After removing the barrier after cancel, the now unmatched barrier for
the explicit cancellation point has to go as well.
This has probably worked before rL255992: With the calls for the explicit
barrier, it was sure that all threads passed a barrier before exiting.
Reported by Simon Convent and Joachim Protze!
Differential Revision: https://reviews.llvm.org/D30088
llvm-svn: 295473
This patch teaches ubsan to insert exactly one null check for the 'this'
pointer per method/lambda.
Previously, given a load of a member variable from an instance method
('this->x'), ubsan would insert a null check for 'this', and another
null check for '&this->x', before allowing the load to occur.
Similarly, given a call to a method from another method bound to the
same instance ('this->foo()'), ubsan would a redundant null check for
'this'. There is also a redundant null check in the case where the
object pointer is a reference ('Ref.foo()').
This patch teaches ubsan to remove the redundant null checks identified
above.
Testing: check-clang and check-ubsan. I also compiled X86FastISel.cpp
with -fsanitize=null using patched/unpatched clangs based on r293572.
Here are the number of null checks emitted:
-------------------------------------
| Setup | # of null checks |
-------------------------------------
| unpatched, -O0 | 21767 |
| patched, -O0 | 10758 |
-------------------------------------
Changes since the initial commit: don't rely on IRGen of C labels in the
test.
Differential Revision: https://reviews.llvm.org/D29530
llvm-svn: 295401
This patch teaches ubsan to insert exactly one null check for the 'this'
pointer per method/lambda.
Previously, given a load of a member variable from an instance method
('this->x'), ubsan would insert a null check for 'this', and another
null check for '&this->x', before allowing the load to occur.
Similarly, given a call to a method from another method bound to the
same instance ('this->foo()'), ubsan would a redundant null check for
'this'. There is also a redundant null check in the case where the
object pointer is a reference ('Ref.foo()').
This patch teaches ubsan to remove the redundant null checks identified
above.
Testing: check-clang and check-ubsan. I also compiled X86FastISel.cpp
with -fsanitize=null using patched/unpatched clangs based on r293572.
Here are the number of null checks emitted:
-------------------------------------
| Setup | # of null checks |
-------------------------------------
| unpatched, -O0 | 21767 |
| patched, -O0 | 10758 |
-------------------------------------
Differential Revision: https://reviews.llvm.org/D29530
llvm-svn: 295391
This patch implements codegen for the reduction clause on
any teams construct for elementary data types. It builds
on parallel reductions on the GPU. Subsequently,
the team master writes to a unique location in a global
memory scratchpad. The last team to do so loads and
reduces this array to calculate the final result.
This patch emits two helper functions that are used by
the OpenMP runtime on the GPU to perform reductions across
teams.
Patch by Tian Jin in collaboration with Arpith Jacob
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29879
llvm-svn: 295335
This patch implements codegen for the reduction clause on
any parallel construct for elementary data types. An efficient
implementation requires hierarchical reduction within a
warp and a threadblock. It is complicated by the fact that
variables declared in the stack of a CUDA thread cannot be
shared with other threads.
The patch creates a struct to hold reduction variables and
a number of helper functions. The OpenMP runtime on the GPU
implements reduction algorithms that uses these helper
functions to perform reductions within a team. Variables are
shared between CUDA threads using shuffle intrinsics.
An implementation of reductions on the NVPTX device is
substantially different to that of CPUs. However, this patch
is written so that there are minimal changes to the rest of
OpenMP codegen.
The implemented design allows the compiler and runtime to be
decoupled, i.e., the runtime does not need to know of the
reduction operation(s), the type of the reduction variable(s),
or the number of reductions. The design also allows reuse of
host codegen, with appropriate specialization for the NVPTX
device.
While the patch does introduce a number of abstractions, the
expected use case calls for inlining of the GPU OpenMP runtime.
After inlining and optimizations in LLVM, these abstractions
are unwound and performance of OpenMP reductions is comparable
to CUDA-canonical code.
Patch by Tian Jin in collaboration with Arpith Jacob
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29758
llvm-svn: 295333
This patch implements codegen for the reduction clause on
any parallel construct for elementary data types. An efficient
implementation requires hierarchical reduction within a
warp and a threadblock. It is complicated by the fact that
variables declared in the stack of a CUDA thread cannot be
shared with other threads.
The patch creates a struct to hold reduction variables and
a number of helper functions. The OpenMP runtime on the GPU
implements reduction algorithms that uses these helper
functions to perform reductions within a team. Variables are
shared between CUDA threads using shuffle intrinsics.
An implementation of reductions on the NVPTX device is
substantially different to that of CPUs. However, this patch
is written so that there are minimal changes to the rest of
OpenMP codegen.
The implemented design allows the compiler and runtime to be
decoupled, i.e., the runtime does not need to know of the
reduction operation(s), the type of the reduction variable(s),
or the number of reductions. The design also allows reuse of
host codegen, with appropriate specialization for the NVPTX
device.
While the patch does introduce a number of abstractions, the
expected use case calls for inlining of the GPU OpenMP runtime.
After inlining and optimizations in LLVM, these abstractions
are unwound and performance of OpenMP reductions is comparable
to CUDA-canonical code.
Patch by Tian Jin in collaboration with Arpith Jacob
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29758
llvm-svn: 295319
Removed ndrange_t as Clang builtin type and added
as a struct type in the OpenCL header.
Use type name to do the Sema checking in enqueue_kernel
and modify IR generation accordingly.
Review: D28058
Patch by Dmitry Borisenkov!
llvm-svn: 295311
Destructor references are not modelled explicitly in the AST. This adds
checks for destructor calls due to variable definitions and temporaries.
If a dllimport function references a non-dllimport destructor, it must
not be emitted available_externally, as the referenced destructor might
live across the DLL boundary and isn't exported.
llvm-svn: 295258
The function is used to check whether a type is a class with
non-dllimport destructor. It needs to look through typedefs and array
types.
llvm-svn: 295257
block or lambda.
This is a follow-up to r281682, which fixed a bug in computeBlockInfo
where the captured VarDecl's type, rather than the captured field type
of the enclosing lambda or block, was used to compute the layout of a
block.
This commit makes similar changes to enterBlockScope. This is necessary
to correctly determine whether a block capture requires cleanup.
rdar://problem/30388124
llvm-svn: 295034
This bypasses integer sanitization checks which are redundant on the expression since it's been checked by Sema. Fixes a clang codegen assertion on "void test() { new int[0+1]{0}; }" when building with -fsanitize=signed-integer-overflow.
llvm-svn: 295006
Use # as the comment leader for AArch64 auto-release elision marker.
This is to keep it in sync with the value used in swift. When building
libdispatch for Linux AArch64, the auto-release elision marker was
emitted. However, ELF uses # as the comment leader while MachO accepts
both ; and #. Use the common marker for it instead.
llvm-svn: 294877
Certain ARC runtime functions have an ABI contract of being forwarding.
Annotate the functions with the appropriate `returned` attribute on the
arguments. This hoists some of the runtime ABI contract information
into the frontend rather than the backend transformations.
The test adjustments are to mark the returned function parameter as
such. The minor change to the IR output is due to the fact that the
returned reference of the object causes it to extend the lifetime of the
object by returning an autoreleased return value. The result is that
the explicit objc_autorelease call is no longer formed, as autorelease
elision is now possible on the return.
llvm-svn: 294872
__fastfail terminates the process immediately with a special system
call. It does not run any process shutdown code or exception recovery
logic.
Fixes PR31854
llvm-svn: 294606
Sometimes the MS ABI needs to emit thunks for declarations that don't
have bodies. Destructor thunks make calls to inlinable functions, so
they need line info or LLVM will complain.
Fixes PR31893
llvm-svn: 294465
Summary:
This teaches clang how to parse and lower the 'interrupt' and 'naked'
attributes.
This allows interrupt signal handlers to be written.
Reviewers: aaron.ballman
Subscribers: malcolm.parsons, cfe-commits
Differential Revision: https://reviews.llvm.org/D28451
llvm-svn: 294402
Guard against a null pointer dereference that caused Clang to crash
when processing a class containing an _Atomic qualified data member,
and that is tagged with 'dllexport'.
Differential Revision: https://reviews.llvm.org/D29208
llvm-svn: 293911
clang-cl would evaluate the arguments right-to-left (see PR), and for
non-Windows targets I suppose we only got it because we were already
emitting left-to-right in CodeGenFunction::EmitCallArgs.
Differential Revision: https://reviews.llvm.org/D29350
llvm-svn: 293732
This re-applies r293343 (reverts commit r293475) with a fix for an
assertion failure caused by a missing integer cast. I tested this patch
by using the built compiler to compile X86FastISel.cpp.o with ubsan.
Original commit message:
Ubsan does not report UB shifts in some cases where the shift exponent
needs to be truncated to match the type of the shift base. We perform a
range check on the truncated shift amount, leading to false negatives.
Fix the issue (PR27271) by performing the range check on the original
shift amount.
Differential Revision: https://reviews.llvm.org/D29234
llvm-svn: 293572
Accounts for a case that caused an assertion failure by attempting to
query for the vtable linkage of a non-dynamic type.t
This reverts commit r292801.
llvm-svn: 293462
First pass at generating weak definitions of inline functions from module files
(& skipping (-O0) or emitting available_externally (optimizations)
definitions where those modules are used).
External functions defined in modules are emitted into the modular
object file as well (this may turn an existing ODR violation (if that
module were imported into multiple translations) into valid/linkable
code).
Internal symbols (static functions, for example) are not correctly
supported yet. The symbol will be produced, internal, in the modular
object - unreferenceable from the users.
Reviewers: rsmith
Differential Revision: https://reviews.llvm.org/D28845
llvm-svn: 293456
Support for CUDA printf is exploited to support printf for
an NVPTX OpenMP device.
To reflect the support of both programming models, the file
CGCUDABuiltin.cpp has been renamed to CGGPUBuiltin.cpp, and
the call EmitCUDADevicePrintfCallExpr has been renamed to
EmitGPUDevicePrintfCallExpr.
Reviewers: jlebar
Differential Revision: https://reviews.llvm.org/D17890
llvm-svn: 293444
Ubsan does not report UB shifts in some cases where the shift exponent
needs to be truncated to match the type of the shift base. We perform a
range check on the truncated shift amount, leading to false negatives.
Fix the issue (PR27271) by performing the range check on the original
shift amount.
Differential Revision: https://reviews.llvm.org/D29234
llvm-svn: 293343
Modify ObjC blocks impl wrt address spaces as follows:
- keep default private address space for blocks generated
as local variables (with captures);
- add global address space for global block literals (no captures);
- make the block invoke function and enqueue_kernel prototype with
the generic AS block pointer parameter to accommodate both
private and global AS cases from above;
- add block handling into default AS because it's implemented as
a special pointer type (BlockPointer) in the frontend and therefore
it is used as a pointer everywhere. This is also needed to accommodate
both private and global AS blocks for the two cases above.
- removes ObjC RT specific symbols (NSConcreteStackBlock and
NSConcreteGlobalBlock) in the OpenCL mode.
Review: https://reviews.llvm.org/D28814
llvm-svn: 293286
This change adds a new type node, DeducedTemplateSpecializationType, to
represent a type template name that has been used as a type. This is modeled
around AutoType, and shares a common base class for representing a deduced
placeholder type.
We allow deduced class template types in a few more places than the standard
does: in conditions and for-range-declarators, and in new-type-ids. This is
consistent with GCC and with discussion on the core reflector. This patch
does not yet support deduced class template types being named in typename
specifiers.
llvm-svn: 293207
This is a simple patch to teach OpenMP codegen to emit the construct
in Generic mode.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29143
llvm-svn: 293183
The handler that deals with IR passed/missed/analysis remarks is extended to
also handle the corresponding MIR remarks.
The more thorough testing in done via llc (rL293113, rL293121). Here we just
make sure that the functionality is accessible through clang.
llvm-svn: 293146
in the current lexical scope.
clang currently emits the lifetime.start marker of a variable when the
variable comes into scope even though a variable's lifetime starts at
the entry of the block with which it is associated, according to the C
standard. This normally doesn't cause any problems, but in the rare case
where a goto jumps backwards past the variable declaration to an earlier
point in the block (see the test case added to lifetime2.c), it can
cause mis-compilation.
To prevent such mis-compiles, this commit conservatively disables
emitting lifetime variables when a label has been seen in the current
block.
This problem was discussed on cfe-dev here:
http://lists.llvm.org/pipermail/cfe-dev/2016-July/050066.html
rdar://problem/30153946
Differential Revision: https://reviews.llvm.org/D27680
llvm-svn: 293106
Summary:
Now when you ask clang to link in a bitcode module, you can tell it to
set attributes on that module's functions to match what we would have
set if we'd emitted those functions ourselves.
This is particularly important for fast-math attributes in CUDA
compilations.
Each CUDA compilation links in libdevice, a bitcode library provided by
nvidia as part of the CUDA distribution. Without this patch, if we have
a user-function F that is compiled with -ffast-math that calls a
function G from libdevice, F will have the unsafe-fp-math=true (etc.)
attributes, but G will have no attributes.
Since F calls G, the inliner will merge G's attributes into F's. It
considers the lack of an unsafe-fp-math=true attribute on G to be
tantamount to unsafe-fp-math=false, so it "merges" these by setting
unsafe-fp-math=false on F.
This then continues up the call graph, until every function that
(transitively) calls something in libdevice gets unsafe-fp-math=false
set, thus disabling fastmath in almost all CUDA code.
Reviewers: echristo
Subscribers: hfinkel, llvm-commits, mehdi_amini
Differential Revision: https://reviews.llvm.org/D28538
llvm-svn: 293097
This patch adds support for the proc_bind clause on the Spmd construct
'target parallel' on the NVPTX device. Since the parallel region is created
upon kernel launch, this clause can be safely ignored on the NVPTX device at
codegen time for level 0 parallelism.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29128
llvm-svn: 293069
This patch adds support for codegen of 'target teams' on the host.
This combined directive has two captured statements, one for the
'teams' region, and the other for the 'parallel'.
This target teams region is offloaded using the __tgt_target_teams()
call. The patch sets the number of teams as an argument to
this call.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29084
llvm-svn: 293005
This patch adds support for codegen of 'target teams' on the host.
This combined directive has two captured statements, one for the
'teams' region, and the other for the 'parallel'.
This target teams region is offloaded using the __tgt_target_teams()
call. The patch sets the number of teams as an argument to
this call.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29084
llvm-svn: 293001
This patch adds support for the Spmd construct 'target parallel' on the
NVPTX device. This involves ignoring the num_threads clause on the device
since the number of threads in this combined construct is already set on
the host through the call to __tgt_target_teams().
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29083
llvm-svn: 292999
The num_threads-clause on the combined directive applies to the
'parallel' region of this construct. We modify the NumThreadsClause
class to capture the clause expression within the 'target' region.
The offload runtime call for 'target parallel' is changed to
__tgt_target_teams() with 1 team and the number of threads set by
this clause or a default if none.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D29082
llvm-svn: 292997
Summary:
The LibFunc::Func enum holds enumerators named for libc functions.
Unfortunately, there are real situations, including libc implementations, where
function names are actually macros (musl uses "#define fopen64 fopen", for
example; any other transitively visible macro would have similar effects).
Strictly speaking, a conforming C++ Standard Library should provide any such
macros as functions instead (via <cstdio>). However, there are some "library"
functions which are not part of the standard, and thus not subject to this
rule (fopen64, for example). So, in order to be both portable and consistent,
the enum should not use the bare function names.
The old enum naming used a namespace LibFunc and an enum Func, with bare
enumerators. This patch changes LibFunc to be an enum with enumerators prefixed
with "LF_". (Unfortunately, a scoped enum is not sufficient to override macros.)
These changes are for clang. See https://reviews.llvm.org/D28476 for LLVM.
Reviewers: rsmith
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D28477
llvm-svn: 292849
Summary:
This reverts commit r292662.
This change broke internal builds. Will provide a reproducer internally.
Subscribers: pcc, mehdi_amini, cfe-commits, mgorny
Differential Revision: https://reviews.llvm.org/D29025
llvm-svn: 292791
To ensure optimization level doesn't pessimize the -fstandalone-debug
vtable debug info optimization (where class definitions are only emitted
where the vtable is emitted - reducing redundant debug info) ensure the
debug info class definition is still omitted when an
available_externally vtable definition is emitted for optimization
purposes.
llvm-svn: 292768
with SEH and openmp
In some cituations (during codegen for Windows SEH constructs)
CodeGenFunction instance may have CurFn equal to nullptr. OpenMP related
code does not expect such situation during cleanup.
llvm-svn: 292590
by providing a memchr builtin that returns char* instead of void*.
Also add a __has_feature flag to indicate the presence of constexpr forms of
the relevant <string> functions.
llvm-svn: 292555
Summary:
SamplePGO uses profile with debug info to collect profile. Unlike the traditional debugging purpose, sample pgo needs more accurate debug info to represent the profile. We add -femit-accurate-debug-info for this purpose. It can be combined with all debugging modes (-g, -gmlt, etc). It makes sure that the following pieces of info is always emitted:
* start line of all subprograms
* linkage name of all subprograms
* standalone subprograms (functions that has neither inlined nor been inlined)
The impact on speccpu2006 binary size (size increase comparing with -g0 binary, also includes data for -g binary, which does not change with this patch):
-gmlt(orig) -gmlt(patched) -g
433.milc 4.68% 5.40% 19.73%
444.namd 8.45% 8.93% 45.99%
447.dealII 97.43% 115.21% 374.89%
450.soplex 27.75% 31.88% 126.04%
453.povray 21.81% 26.16% 92.03%
470.lbm 0.60% 0.67% 1.96%
482.sphinx3 5.77% 6.47% 26.17%
400.perlbench 17.81% 19.43% 73.08%
401.bzip2 3.73% 3.92% 12.18%
403.gcc 31.75% 34.48% 122.75%
429.mcf 0.78% 0.88% 3.89%
445.gobmk 6.08% 7.92% 42.27%
456.hmmer 10.36% 11.25% 35.23%
458.sjeng 5.08% 5.42% 14.36%
462.libquantum 1.71% 1.96% 6.36%
464.h264ref 15.61% 16.56% 43.92%
471.omnetpp 11.93% 15.84% 60.09%
473.astar 3.11% 3.69% 14.18%
483.xalancbmk 56.29% 81.63% 353.22%
geomean 15.60% 18.30% 57.81%
Debug info size change for -gmlt binary with this patch:
433.milc 13.46%
444.namd 5.35%
447.dealII 18.21%
450.soplex 14.68%
453.povray 19.65%
470.lbm 6.03%
482.sphinx3 11.21%
400.perlbench 8.91%
401.bzip2 4.41%
403.gcc 8.56%
429.mcf 8.24%
445.gobmk 29.47%
456.hmmer 8.19%
458.sjeng 6.05%
462.libquantum 11.23%
464.h264ref 5.93%
471.omnetpp 31.89%
473.astar 16.20%
483.xalancbmk 44.62%
geomean 16.83%
Reviewers: davidxl, andreadb, rob.lougher, dblaikie, echristo
Reviewed By: dblaikie, echristo
Subscribers: hfinkel, rob.lougher, andreadb, gbedwell, cfe-commits, probinson, llvm-commits, mehdi_amini
Differential Revision: https://reviews.llvm.org/D25435
llvm-svn: 292458
In ThinLTO mode, type metadata will require the module to be written as a
multi-module bitcode file, which is currently incompatible with the Darwin
linker. It is also useful to be able to enable or disable multi-module bitcode
for testing purposes. This introduces a cc1-level flag, -f{,no-}lto-unit,
which is used by the driver to enable multi-module bitcode on all but
Darwin+ThinLTO, and can also be used to enable/disable the feature manually.
Differential Revision: https://reviews.llvm.org/D28877
llvm-svn: 292448
Introduced in r181561 - it may've been subsumed by work done to allow
emission of declarations for vtable types while still emitting some of
their member functions correctly for those declarations. Whatever the
reason, the tests pass without this code now.
llvm-svn: 292439
The if-clause on the combined directive potentially applies to both the
'target' and the 'parallel' regions. Codegen'ing the if-clause on the
combined directive requires additional support because the expression in
the clause must be captured by the 'target' capture statement but not
the 'parallel' capture statement. Note that this situation arises for
other clauses such as num_threads.
The OMPIfClause class inherits OMPClauseWithPreInit to support capturing
of expressions in the clause. A member CaptureRegion is added to
OMPClauseWithPreInit to indicate which captured statement (in this case
'target' but not 'parallel') captures these expressions.
To ensure correct codegen of captured expressions in the presence of
combined 'target' directives, OMPParallelScope was added to 'parallel'
codegen.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28781
llvm-svn: 292437
This patch adds codegen for the 'target parallel' directive on the NVPTX
device. We term offload OpenMP directives such as 'target parallel' and
'target teams distribute parallel for' as SPMD constructs. SPMD constructs,
in contrast to Generic ones like the plain 'target', can never contain
a serial region.
SPMD constructs can be handled more efficiently on the GPU and do not
require the Warp Loop of the Generic codegen scheme. This patch adds
SPMD codegen support for 'target parallel' on the NVPTX device and can
be reused for other SPMD constructs.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28755
llvm-svn: 292428
This patch adds support for codegen of 'target parallel' on the host.
It is also the first combined directive that requires two or more
captured statements. Support for this functionality is included in
the patch.
A combined directive such as 'target parallel' has two captured
statements, one for the 'target' and the other for the 'parallel'
region. Two captured statements are required because each has
different implicit parameters (see SemaOpenMP.cpp). For example,
the 'parallel' has 'global_tid' and 'bound_tid' while the 'target'
does not. The patch adds support for handling multiple captured
statements based on the combined directive.
When codegen'ing the 'target parallel' directive, the 'target'
outlined function is created using the outer captured statement
and the 'parallel' outlined function is created using the inner
captured statement.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28753
llvm-svn: 292419
This patch adds support for codegen of 'target parallel' on the host.
It is also the first combined directive that requires two or more
captured statements. Support for this functionality is included in
the patch.
A combined directive such as 'target parallel' has two captured
statements, one for the 'target' and the other for the 'parallel'
region. Two captured statements are required because each has
different implicit parameters (see SemaOpenMP.cpp). For example,
the 'parallel' has 'global_tid' and 'bound_tid' while the 'target'
does not. The patch adds support for handling multiple captured
statements based on the combined directive.
When codegen'ing the 'target parallel' directive, the 'target'
outlined function is created using the outer captured statement
and the 'parallel' outlined function is created using the inner
captured statement.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28753
llvm-svn: 292374
This patch refactors code that calls codegen for target regions. Currently
the codebase only supports the 'target' directive. The patch pulls out
common target processing code into a static function that can be called
by codegen for any target directive.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28752
llvm-svn: 292134
This comes up in V8, which has a Handle template class that wraps a
typed pointer, and is frequently passed by value. The pointer is stored
in the base, HandleBase. This change allows us to pass the struct as a
pointer instead of using byval. This avoids creating tons of temporary
allocas that we copy from during call lowering.
Eventually, it would be good to use FCAs here instead.
llvm-svn: 291917
Summary: LTO backend will not invoke SampleProfileLoader pass even if -fprofile-sample-use is specified. This patch passes the flag down so that pass manager can add the SampleProfileLoader pass correctly.
Reviewers: mehdi_amini, tejohnson
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D28588
llvm-svn: 291870
There is a synchronization point between the reference count of a block dropping to zero and it's destruction, which TSan does not observe. Do not report errors in the compiler-emitted block destroy method and everything called from it.
This is similar to https://reviews.llvm.org/D25857
Differential Revision: https://reviews.llvm.org/D28387
llvm-svn: 291868
* Do not initialize these variables when initializing the rest of the
thread_locals in the TU; they have unordered initialization so they can be
initialized by themselves.
This fixes a rejects-valid bug: we would make the per-variable initializer
function internal, but put it in a comdat keyed off the variable, resulting
in link errors when the comdat is selected from a different TU (as the per
TU TLS init function tries to call an init function that does not exist).
* On Darwin, when we decide that we're not going to emit a thread wrapper
function at all, demote its linkage to External. Fixes a verifier failure
on explicit instantiation of a thread_local variable on Darwin.
llvm-svn: 291865
Summary: LTO backend will not invoke SampleProfileLoader pass even if -fprofile-sample-use is specified. This patch passes the flag down so that pass manager can add the SampleProfileLoader pass correctly.
Reviewers: mehdi_amini, tejohnson
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D28588
llvm-svn: 291774
a header of that same module.
This fixes a regression caused by r280409.
rdar://problem/29930553
This is an updated version for r291628 (which was reverted in r291688).
llvm-svn: 291689
This patch is to implement sema and parsing for 'target teams distribute simd’ pragma.
Differential Revision: https://reviews.llvm.org/D28252
llvm-svn: 291579
Summary:
This patch introduces support for the execution of parallel constructs in a target
region on the NVPTX device. Parallel regions must be in the lexical scope of the
target directive.
The master thread in the master warp signals parallel work for worker threads in worker
warps on encountering a parallel region.
Note: The patch does not yet support capture of arguments in a parallel region so
the test cases are simple.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28145
llvm-svn: 291565
Summary:
In order to simplify distributed build system integration, where actions
may be scheduled before the Thin Link which determines the list of
objects selected by the linker. The gold plugin currently will emit
0-sized index files for objects not selected by the link, to enable
checking for expected output files by the build system. If the build
system then schedules a backend action for these bitcode files, we want
to be able to fall back to normal compilation instead of failing.
Fallback is enabled under an option in LLVM (D28410), in which case a
nullptr is returned from llvm::getModuleSummaryIndexForFile. Clang can
just proceed with non-ThinLTO compilation in that case.
I am investigating whether this can be addressed in our build system,
but that is a longer term fix and so this enables a workaround in the
meantime.
Reviewers: mehdi_amini
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D28362
llvm-svn: 291303
Summary:
This patch makes the type_mismatch static data 7 bytes smaller (and it
ends up being 16 bytes smaller due to alignment restrictions, at least
on some x86-64 environments).
It revs up the type_mismatch handler version since we're breaking binary
compatibility. I will soon post a patch for the compiler-rt side.
Reviewers: rsmith, kcc, vitalybuka, pgousseau, gbedwell
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D28242
llvm-svn: 291236
Add builtins for the functions and custom codegen mapping the builtins to their
corresponding intrinsics and handling the endian related swapping.
https://reviews.llvm.org/D26546
llvm-svn: 291179
Summary:
This patch adds two fields to the offload entry descriptor. One field is meant to signal Ctors/Dtors and `link` global variables, and the other is reserved for runtime library use.
Currently, these fields are only filled with zeros in the current code generation, but that will change when `declare target` is added.
The reason, we are adding these fields now is to make the code generation consistent with the runtime library proposal under review in https://reviews.llvm.org/D14031.
Reviewers: ABataev, hfinkel, carlo.bertolli, kkwli0, arpith-jacob, Hahnfeld
Subscribers: cfe-commits, caomhin, jholewinski
Differential Revision: https://reviews.llvm.org/D28298
llvm-svn: 291124
inline assembly may use the `.include` directive to include other
content into the file. Without the integrated assembler, the `-I` group
gets passed to the assembler. Emulate this by collecting the header
search paths and passing them to the IAS.
Resolves PR24811!
llvm-svn: 291123
This patch includes updates for codegen of the target region for the NVPTX
device. It moves initializers from the compiler to the runtime and updates
the worker loop to assume parallel work is retrieved from the runtime. A
subsequent patch will update the codegen to retrieve the parallel work using
calls to the runtime. It includes the removal of the inline attribute
for the worker loop and disabling debug info in it.
This allows codegen for a target directive and serial execution on the
NVPTX device.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28125
llvm-svn: 291121
Front end component (back end changes are D27392). The vectorcall
calling convention was broken subtly in two cases. First,
it didn't properly handle homogeneous vector aggregates (HVAs).
Second, the vectorcall specification requires that only the
first 6 parameters be eligible for register assignment.
This patch fixes both issues.
Differential Revision: https://reviews.llvm.org/D27529
llvm-svn: 291041
This patch includes updates for codegen of the target region for the NVPTX
device. It moves initializers from the compiler to the runtime and updates
the worker loop to assume parallel work is retrieved from the runtime. A
subsequent patch will update the codegen to retrieve the parallel work using
calls to the runtime. It includes the removal of the inline attribute
for the worker loop and disabling debug info in it.
This allows codegen for a target directive and serial execution on the
NVPTX device.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28125
llvm-svn: 290983
in non-void functions that fall off at the end without returning a value when
compiling C++.
Clang uses the new compiler flag to determine when it should treat control flow
paths that fall off the end of a non-void function as unreachable. If
-fno-strict-return is on, the code generator emits the ureachable and trap
IR only when the function returns either a record type with a non-trivial
destructor or another non-trivially copyable type.
The primary goal of this flag is to avoid treating falling off the end of a
non-void function as undefined behaviour. The burden of undefined behaviour
is placed on the caller instead: if the caller ignores the returned value then
the undefined behaviour is avoided. This kind of behaviour is useful in
several cases, e.g. when compiling C code in C++ mode.
rdar://13102603
Differential Revision: https://reviews.llvm.org/D27163
llvm-svn: 290960
The special case to widen the integer literal zero when passed to
variadic function calls should only apply to variadic functions, not
unprototyped functions. This is consistent with what MSVC does. In this
test case, MSVC uses a 4-byte store to pass the 5th argument to 'kr' and
an 8-byte store to pass the zero to 'v':
void v(int, ...);
void kr();
void f(void) {
v(1, 2, 3, 4, 0);
kr(1, 2, 3, 4, 0);
}
Aaron Ballman discovered this issue in https://reviews.llvm.org/D28166
llvm-svn: 290906
This patch cleans up private methods for NVPTX OpenMP codegen. It converts private
members to static functions to follow the coding style of CGOpenMPRuntime.cpp and
declutter the header file.
Reviewers: ABataev
Differential Revision: https://reviews.llvm.org/D28124
llvm-svn: 290904
https://reviews.llvm.org/D17840
This patch enables private, firstprivate, and lastprivate clauses for the OpenMP distribute directive.
Regression tests differ from the similar case of the same clauses on the for directive, by removing a reference to two global variables g and g1. This is necessary because: 1. a distribute pragma is only allowed inside a target region; 2. referring a global variable (e.g. g and g1) in a target region requires the program to enclose the variable in a "declare target" region; 3. declare target pragmas, which are used to define a declare target region, are currently unavailable in clang (patch being prepared).
For this reason, I moved the global declarations into local variables.
llvm-svn: 290898
This patch is to implement sema and parsing for 'target teams distribute parallel for simd’ pragma.
Differential Revision: https://reviews.llvm.org/D28202
llvm-svn: 290862
The MS ABI RTTI has a reserved field which is used as a cache for the
demangled name. It must be zero-initialized, which is used as a hint by
the runtime to say that the cache has not been populated. Since this
field is populated at runtime, the RTTI structures must be placed in the
.data section rather than .rdata. NFC
llvm-svn: 290799
This patch is to implement sema and parsing for 'target teams distribute parallel for’ pragma.
Differential Revision: https://reviews.llvm.org/D28160
llvm-svn: 290725
Summary:
We can simply import all external values with summaries included in
the individual index file created for the distributed backend job,
as only those are added to the individual index file created by the
WriteIndexesThinBackend (in addition to summaries for the original
module, which are skipped here).
While computing the cross module imports on this index would come to
the same conclusion as the original thin link import logic, it is
unnecessary work. And when tuning, it avoids the need to pass the
same function importing parameters (e.g. -import-instr-limit) to
both the thin link and the backends (otherwise they won't make the
same decisions).
Reviewers: mehdi_amini, pcc
Subscribers: cfe-commits
Differential Revision: https://reviews.llvm.org/D28139
llvm-svn: 290674
Our newly aggressive constant folding logic makes it possible for
CGExprConstant to see the same CompoundLiteralExpr more than once. So,
emitting a new GlobalVariable every time we see a CompoundLiteral is no
longer correct.
We had a similar issue with BlockExprs that was caught while testing
said aggressive folding, so I applied the same style of fix (see D26410)
here. If we find yet another case where this needs to happen, we should
probably refactor this so we don't have a third DenseMap+getter+setter.
As a design note: getAddrOfConstantCompoundLiteralIfEmitted is really
only intended to be called by ConstExprEmitter::EmitLValue. So,
returning a GlobalVariable* instead of a ConstantAddress costs us
effectively nothing, and saves us either a few bytes per entry in our
map or a bit of code duplication.
llvm-svn: 290661
John McCall