[cuda] treat file scope __asm as __host__ and ignore it during device-side compilation.
Currently clang emits file-scope asm during *both* host and device
compilation modes which is usually a wrong thing to do.
There's no way to attach any attribute to an __asm statement, so
there's no way to differentiate between host-side and device-side
file-scope asm. This patch makes clang to match nvcc behavior and
emit file-scope-asm only during host-side compilation.
Paul Robinson [Mon, 27 Apr 2015 18:14:32 +0000 (18:14 +0000)]
Support generating NMake/Jom-style depfiles.
NMake is a Make-like builder that comes with Microsoft Visual Studio.
Jom (https://wiki.qt.io/Jom) is an NMake-compatible build tool.
Dependency files for NMake/Jom need to use double-quotes to wrap
filespecs containing special characters, instead of the backslash
escapes that GNU Make wants.
Adds the -MV option, which specifies to use double-quotes as needed
instead of backslash escapes when writing the dependency file.
Emit the following code for 'taskwait' directive within tied task:
call i32 @__kmpc_omp_taskwait(<loc>, i32 <thread_id>);
Differential Revision: http://reviews.llvm.org/D9245
...
void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]};
switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), RedList, reduce_func, &<lock>)) {
case 1:
<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]);
...
<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]);
__kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>);
break;
case 2:
Atomic(<LHSExprs>[0] = ReductionOperation0(*<LHSExprs>[0], *<RHSExprs>[0]));
...
Atomic(<LHSExprs>[<n>-1] = ReductionOperation<n>-1(*<LHSExprs>[<n>-1], *<RHSExprs>[<n>-1]));
break;
default:;
}
Reduction variables are a kind of a private variables, they have private copies, but initial values are chosen in accordance with the reduction operation.
If sections directive has only single section, then original shared variables are used instead with barrier at the end of the directive.
Differential Revision: http://reviews.llvm.org/D9242
[OPENMP] Codegen for 'lastprivate' clause in 'sections' directive.
#pragma omp sections lastprivate(<var>)
<BODY>;
This construct is translated into something like:
<last_iter> = alloca i32
<init for lastprivates>;
<last_iter> = 0
; No initializer for simple variables or a default constructor is called for objects.
; For arrays perform element by element initialization by the call of the default constructor.
...
OMP_FOR_START(...,<last_iter>, ..); sets <last_iter> to 1 if this is the last iteration.
<BODY>
...
OMP_FOR_END
if (<last_iter> != 0) {
<final copy for lastprivate>; Update original variable with the lastprivate value.
}
call __kmpc_cancel_barrier() ; an implicit barrier to avoid possible data race.
If there is only one section, there is no special code generation, original shared variables are used + barrier is emitted at the end of the directive.
Differential Revision: http://reviews.llvm.org/D9240
[OPENMP] Codegen for 'private' clause in 'sections' directive.
If there are 2 or more sections in a 'section' directive the following code is generated:
<default init for privates>
@__kmpc_for_static_init_4();
<BODY for sections directive>
@__kmpc_for_static_fini()
If there is only one section, the following code is generated:
if (@__kmpc_single()) {
<default init for privates>
@__kmpc_end_single();
}
Differential Revision: http://reviews.llvm.org/D9239
David Majnemer [Sun, 26 Apr 2015 04:58:18 +0000 (04:58 +0000)]
[Sema] Don't allow unverified bitfields in FieldDecls
VerifyBitField must be called if we are to form a bitfield FieldDecl.
We will not verify the bitfield if the decl is known to be malformed in
other ways; pretend that we don't have a bitfield if this happens.
A more elegant fix would preclude this case by defining the
rules such that zero-size classes are always formally empty.
I believe the only extensions which create zero-size classes
right now are flexible arrays and zero-length arrays; it's
not abstractly unreasonable to say that those don't count
as members for the purposes of emptiness, just as zero-width
bitfields don't count. But that's an ABI-affecting change
and requires further discussion; in the meantime, let's not
assert / miscompile.
InstrProf: Fix coverage maps for conditional operators
This fixes a crash when we're emitting coverage and a macro appears
between two binary conditional operators, ie, "foo ?: MACRO ?: bar",
and fixes the interaction of macros and conditional operators in
general.
Before this patch, passing a non-existent absolute path to clang-cl would cause
stat'ing of impossible paths. For example, `clang-cl -c d:\adsfasdf.txt` would
cause a stat of
C:\Program Files (x86)\Microsoft Visual Studio 12.0\VC\LIBd:\asdfadsf.cc
[OPENMP] Codegen for 'firstprivate' clause in 'single' directive.
Emit the following code for 'single' directive with 'firtstprivate' clause:
if (@__kmpc_single()) {
<init for firstprivates>
@__kmpc_end_single();
}
@__kmpc_cancel_barrier(); // To avoid data race in firstprivate init
Differential Revision: http://reviews.llvm.org/D9223
[OPENMP] Codegen for 'firstprivate' clause in 'sections' directive.
If there are 2 or more sections in a 'section' directive the following code is generated:
<init for firstprivates>
@__kmpc_cancel_barrier();// To avoid data race in firstprivate init
@__kmpc_for_static_init_4();
<BODY for sections directive>
@__kmpc_for_static_fini()
If there is only one section, the following code is generated:
if (@__kmpc_single()) {
<init for firstprivates>
@__kmpc_end_single();
}
@__kmpc_cancel_barrier(); // To avoid data race in firstprivate init
Differential Revision: http://reviews.llvm.org/D9214
InstrProf: Stop using RegionCounter outside of CodeGenPGO (NFC)
The RegionCounter type does a lot of legwork, but most of it is only
meaningful within the implementation of CodeGenPGO. The uses elsewhere
in CodeGen generally just want to increment or read counters, so do
that directly.
Richard Smith [Thu, 23 Apr 2015 18:18:26 +0000 (18:18 +0000)]
[modules] Determine the set of macros exported by a submodule at the end of that submodule.
Previously we'd defer this determination until writing the AST, which doesn't
allow us to use this information when building other submodules of the same
module. This change also allows us to use a uniform mechanism for writing
module macro records, independent of whether they are local or imported.
In r235553, Clang started emitting lifetime markers more often. This
caused false negative in MSan, because MSan only poisons all allocas
once at function entry. Eventually, MSan should poison allocas at
lifetime start and probably also lifetime end, but until then, let's not
emit markers that aren't going to be useful.
Adds codegen for 'atomic capture' constructs with the following forms of expressions/statements:
v = x binop= expr;
v = x++;
v = ++x;
v = x--;
v = --x;
v = x = x binop expr;
v = x = expr binop x;
{v = x; x = binop= expr;}
{v = x; x++;}
{v = x; ++x;}
{v = x; x--;}
{v = x; --x;}
{x = x binop expr; v = x;}
{x binop= expr; v = x;}
{x++; v = x;}
{++x; v = x;}
{x--; v = x;}
{--x; v = x;}
{x = x binop expr; v = x;}
{x = expr binop x; v = x;}
{v = x; x = expr;}
If x and expr are integer and binop is associative or x is a LHS in a RHS of the assignment expression, and atomics are allowed for type of x on the target platform atomicrmw instruction is emitted.
Otherwise compare-and-swap sequence is emitted.
Update of 'v' is not required to be be atomic with respect to the read or write of the 'x'.
David Majnemer [Thu, 23 Apr 2015 05:21:20 +0000 (05:21 +0000)]
[MS ABI] Treat ConstantArrayType like IncompleteArrayType in args
Type backreferences for arguments use the DecayedType's original type.
Because of this, arguments with the same canonical type with the same
mangling would not backreference each other if one was a
ConstantArrayType while the other was an IncompleteArrayType. Solve
this by canonicalizing the ConstantArrayType to a suitable
IncompleteArrayType.
Unify the way we report overflow in increment/decrement operator.
Summary:
Make sure signed overflow in "x--" is checked with
llvm.ssub.with.overflow intrinsic and is reported as:
"-2147483648 - 1 cannot be represented in type 'int'"
instead of:
"-2147483648 + -1 cannot be represented in type 'int'"
, like we do for unsigned overflow.
Test Plan: clang + compiler-rt regression test suite
InstrProf: Fix a shadowing error that would break length of profile names
We try to use the member variable "FuncName" here, but we've also used
that name as a parameter. This ends with us getting the length of the
function name wrong when we generate the coverage data.
[WinEH] Don't emit an exceptional cleanup for llvm.eh.endcatch
These extra endcatch markers aren't helping identify regions to outline,
so let's get rid of them. LLVM outlines (more or less) from begincatch
to endcatch. Any unwind edge from an enclosed invoke is a transition to
a new exception handler, which has it's own outlining markers.
David Majnemer [Wed, 22 Apr 2015 21:38:15 +0000 (21:38 +0000)]
Revert "Revert r234581, it might have caused a few miscompiles in Chromium."
This reverts commit r234700. It turns out that the lifetime markers
were not the cause of Chromium failing but a bug which was uncovered by
optimizations exposed by the markers.
[OPENMP] Codegen for 'if' clause in 'task' directive.
If condition evaluates to true, the code executes task by calling @__kmpc_omp_task() runtime function.
If condition evaluates to false, the code executes serial version of the code by executing the following code:
call void @__kmpc_omp_task_begin_if0(<loc>, <threadid>, <task_t_ptr, returned by @__kmpc_omp_task_alloc()>);
proxy_task_entry(<gtid>, <task_t_ptr, returned by @__kmpc_omp_task_alloc()>);
call void @__kmpc_omp_task_complete_if0(<loc>, <threadid>, <task_t_ptr, returned by @__kmpc_omp_task_alloc()>);
Also it checks if the condition is constant and if it is constant it evaluates its value and then generates either parallel version of the code (if the condition evaluates to true), or the serial version of the code (if the condition evaluates to false).
Differential Revision: http://reviews.llvm.org/D9143
[OPENMP] Codegen for 'private' clause in 'for' directive.
This patch generates helper variables which used as a private copies of the corresponding original variables inside an OpenMP 'for' directive. These generated variables are initialized by default (with the default constructor, if any). In OpenMP region references to original variables are replaced by the references to these private helper variables.
Differential Revision: http://reviews.llvm.org/D9106
[OPENMP] Fix use of unsigned counters in loops with zero trip count.
Patch fixes bugs in codegen for loops with unsigned counters and zero trip count. Previously preconditions for all loops were built using logic (Upper - Lower) > 0. But if the loop is a loop with zero trip count, then Upper - Lower is < 0 only for signed integer, for unsigned we're running into an underflow situation.
In this patch we're using original Lower<Upper condition to check that loop body can be executed at least once. Also this allows to skip code generation for loops, if it is known that preconditions for the loop are always false.
Differential Revision: http://reviews.llvm.org/D9103
__kmpc_ordered(ident_t *, gtid);
<associated statement>;
__kmpc_end_ordered(ident_t *, gtid);
Also for 'for' directives with the dynamic scheduling and an 'ordered' clause added a call to '__kmpc_dispatch_fini_(4|8)[u]()' function after increment expression for loop control variable:
Hans Wennborg [Wed, 22 Apr 2015 04:05:17 +0000 (04:05 +0000)]
Don't dllimport/export class members with internal linkage (PR23308)
For example, a function taking a parameter with internal linkage will
itself have internal linkage since it cannot be called outside the
translation unit.
Richard Smith [Wed, 22 Apr 2015 00:26:11 +0000 (00:26 +0000)]
[modules] Build a DAG of module macros for each identifier.
This graph will be used to determine the current set of active macros. This is
foundation work for getting macro visibility correct across submodules of the
current module. No functionality change for now.
[cuda] Allow using integral non-type template parameters as launch_bounds attribute arguments.
- Changed CUDALaunchBounds arguments from integers to Expr* so they can
be saved in AST for instantiation.
- Added support for template instantiation of launch_bounds attrubute.
- Moved evaluation of launch_bounds arguments to NVPTXTargetCodeGenInfo::
SetTargetAttributes() where it can be done after template instantiation.
- Added a warning on negative launch_bounds arguments.
- Amended test cases.
Added cuda_builtin_vars.h which implements built-in CUDA variables
using __declattr(property).
Fields of built-in variables (except for warpSize) are implemented
using __declattr(property) which replaces read/write of a member field
with a call to a getter/setter member function, in this case with
appropriate NVPTX builtin.
Added a test case to check diagnostics on attempt to construct or
improperly access a built-in variable.
Richard Smith [Tue, 21 Apr 2015 21:46:32 +0000 (21:46 +0000)]
[modules] Move list of exported module macros from IdentifierInfo lookup table to separate storage, adjacent to the macro directive history.
This is substantially simpler, provides better space usage accounting in bcanalyzer,
and gives a more compact representation. No functionality change intended.
DebugInfo: Prepare for removal of DIArray and DITypeArray typedefs
An upcoming LLVM commit will remove the `DIArray` and `DITypeArray`
typedefs that shadow `DebugNodeArray` and `MDTypeRefArray`,
respectively. Use those types directly.
Added cuda_builtin_vars.h which implements built-in CUDA variables
using __declattr(property).
Fields of built-in variables (except for warpSize) are implemented
using __declattr(property) which replaces read/write of a member field
with a call to a getter/setter member function, in this case with
appropriate NVPTX builtin.
Added a test case to check diagnostics on attempt to construct or
improperly access a built-in variable.
Implement target-specific __attribute__((aligned)) value
The GCC construct __attribute__((aligned)) is defined to set alignment
to "the default alignment for the target architecture" according to
the GCC documentation:
The default alignment is sufficient for all scalar types, but may not be
enough for all vector types on a target that supports vector operations.
The default alignment is fixed for a particular target ABI.
clang currently hard-coded an alignment of 16 bytes for that construct,
which is correct on some platforms (including X86), but wrong on others
(including SystemZ). Since this value is ABI-relevant, it is important
to get correct for compatibility purposes.
This patch adds a new TargetInfo member "DefaultAlignForAttributeAligned"
that targets can set to the appropriate default __attribute__((aligned))
value.
Note that I'm deliberately *not* using the existing "SuitableAlign"
value, which is used to set the pre-defined macro __BIGGEST_ALIGNMENT__,
since those two values may not be the same on all platforms. In fact,
on X86, __attribute__((aligned)) always uses 16-byte alignment, while
__BIGGEST_ALIGNMENT__ may be larger if AVX-2 or AVX-512 are supported.
(This is actually not yet correctly implemented in clang either.)
The patch provides a value for DefaultAlignForAttributeAligned only for
SystemZ, and leaves the default for all other targets at 16, which means
no visible change in behavior on all other targets. (The value is still
wrong for some other targets, but I'd prefer to leave it to the target
maintainers for those platforms to fix.)
Code in CodeGenModule::GetOrCreateLLVMGlobal that sets up GlobalValue
object for LLVM external symbols has this comment:
// FIXME: This code is overly simple and should be merged with other global
// handling.
One part does seems to be "overly simple" currently is that this code
never sets any alignment info on the GlobalValue, so that the emitted
IR does not have any align attribute on external globals. This can
lead to unnecessarily inefficient code generation.
This patch adds a GV->setAlignment call to set alignment info.
SystemZ prefers to align all global variables to two bytes, which is
implemented by setting the TargetInfo member MinGlobalAlign.
However, for compatibility with existing compilers this should *not*
change the ABI alignment value as retrieved via __alignof__, which
it currently does.
This patch fixes the issue by having ASTContext::getDeclAlign ignore
the MinGlobalAlign setting in the ForAlignof case.
Since SystemZ is the only platform setting MinGlobalAlign, this should
cause no change for any other target.
DebugInfo: Prepare for deletion of subclasses of DIType
Subclasses of (the already deleted) `DIType` will be deleted by an
upcoming LLVM commit. Remove references.
While `DICompositeType` wraps `MDCompositeTypeBase` and `DIDerivedType`
wraps `MDDerivedTypeBase`, most uses of each really meant the more
specific `MDCompositeType` and `MDDerivedType`. I updated accordingly.
Put statement expression decls in the enclosing code DeclContext
We already check that statement expressions are in a function or block,
but we didn't do anything with that information. Now we use that
DeclContext for the duration of the statement expression. Otherwise,
we'd treat statement expression locals as static data members and go
into the weeds.
David Majnemer [Sun, 19 Apr 2015 07:53:29 +0000 (07:53 +0000)]
[MS ABI] Correctly associate align attrs before the class-key
__declspec(align(...)) is unlike all other attributes in that it is not
applied to a variable if it appears before the class-key. If the
tag in question isn't part of a variable declaration, it is not ignored.
Instead, the alignment attribute is applied to the tag.
Pete Cooper [Fri, 17 Apr 2015 22:16:24 +0000 (22:16 +0000)]
Change AArch64 i128 returns to use [2 x i64] when possible.
Something like { void*, void * } would be passed to a function as a [2 x i64], but returned as an i128. This patch unifies the 2 behaviours so that we also return it as a [2 x i64].
This is better for the quality of the IR, and the size of the final LLVM binary as we tend to want to insert/extract values from these types and do so with the insert/extract instructions is less IR than shifting, truncating, and or'ing values.
projects / clang / log