Emits the following code for the clause at the beginning of the outlined function for implicit threads:
if (<not a master thread>) {
...
<thread local copy of var> = <master thread local copy of var>;
...
}
<sync point>;
Checking for a non-master thread is performed by comparing of the address of the thread local variable with the address of the master's variable. Master thread always uses original variables, so you always know the address of the variable in the master thread.
Differential Revision: http://reviews.llvm.org/D9026
llvm-svn: 235075
#pragma omp for lastprivate(<var>)
for (i = a; i < b; ++b)
<BODY>;
This construct is translated into something like:
<last_iter> = alloca i32
<lastprivate_var> = alloca <type>
<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) {
<var> = <lastprivate_var> ; Update original variable with the lastprivate value.
}
call __kmpc_cancel_barrier() ; an implicit barrier to avoid possible data race.
Differential Revision: http://reviews.llvm.org/D8658
llvm-svn: 235074
More fallout from r228234; when looking up an identifier in a PCH that
imports the Cocoa module on Darwin, it was taking 2 to 5 seconds
because we were hammering the MapVector::erase() function, which is
O(n). For now, just clear() the contained SmallVector to get back to
0.25 - 0.5 seconds. This is probably not the long-term fix, because
without modules or without PCH the performance is more like 0.02
seconds.
llvm-svn: 234655
Added sema checks for forms of expressions/statements allowed under control of 'atomic capture' directive + generation of helper objects for future codegen.
llvm-svn: 233785
Adds atomic update codegen for the following forms of expressions:
x binop= expr;
x++;
++x;
x--;
--x;
x = x binop expr;
x = expr binop x;
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:
bb:
...
atomic load <x>
cont:
<expected> = phi [ <x>, label %bb ], [ <new_failed>, %cont ]
<desired> = <expected> binop <expr>
<res> = cmpxchg atomic &<x>, desired, expected
<new_failed> = <res>.field1;
br <res>field2, label %exit, label %cont
exit:
...
Differential Revision: http://reviews.llvm.org/D8536
llvm-svn: 233513
I again added the "reasonable" assertions and they again fired during
a modules self-host.
This hopefully will un-break the self-host build bot. No test case handy
and adding one seems to have little or no value really.
llvm-svn: 233426
if the merged definition is visible, and perform lookups into all merged copies
of the definition (not just for special members) so that we can complete the
redecl chains for members of the class.
llvm-svn: 233420
declaration name so that we mark declarations for emission in
a deterministic order (and in turn give them deterministic IDs).
This is the last for loop or data structure I can find by inspection of
the AST writer which doesn't use a deterministic order.
Found by inspection, no test case.
llvm-svn: 233348
order based on order of insertion.
This should cause both our warnings about these and the modules
serialization to be deterministic as a consequence.
Found by inspection.
llvm-svn: 233343
DeclIDs so that in addition to be grouped by file, the order of these
groups is stable.
Found by inspection, no test case. Not sure this can be observed without
a randomized seed for the hash table, but we shouldn't be relying on the
hash table layout under any circumstances.
llvm-svn: 233339
traversing the identifier table.
No easy test case as this table is somewhere between hard and impossible
to observe as non-deterministically ordered. The table is a hash table
but we hash the string contents and never remove entries from the table
so the growth pattern, etc, is all completely fixed. However, relying on
the hash function being deterministic is specifically against the
long-term direction of LLVM's hashing datastructures, which are intended
to provide *no* ordering guarantees. As such, this defends against these
things by sorting the identifiers. Sorting identifiers right before we
emit them to a serialized form seems a low cost for predictability here.
llvm-svn: 233332
logic removed.
This logic was both inserting all builtins into the identifier table and
ensuring they would get serialized. The first happens unconditionally
now, and we always write out the entire identifier table. This code can
simply go away.
llvm-svn: 233331
constructors in the current lexical context even though name lookup
found them via some other context merged into the redecl chain.
This can only happen for implicit constructors which can only have the
name of the type of the current context, so we can fix this by simply
*always* merging those names first. This also has the advantage of
removing the walk of the current lexical context from the common case
when this is the only constructor name we need to deal with (implicit or
otherwise).
I've enhanced the tests to cover this case (and uncovered an unrelated
bug which I fixed in r233325).
llvm-svn: 233327
deserializing an inherited constructor.
This is the exact same logic we use when deserializing method overrides
for the same reason: the canonical decl may end up pinned to a different
decl when we are improting modules, we need to re-pin to the canonical
one during reading.
My test case for this will come in a subsequent commit. I was trying to
test a more tricky bug fix and the test case happened to tickle this bug
as well.
llvm-svn: 233325
Clang was inserting these into a dense map. While it never iterated the
dense map during normal compilation, it did when emitting a module. Fix
this by using a standard MapVector to preserve the order in which we
encounter the late parsed templates.
I suspect this still isn't ideal, as we don't seem to remove things from
this map even when we mark the templates as no longer late parsed. But
I don't know enough about this particular extension to craft a nice,
subtle test case covering this. I've managed to get the stress test to
at least do some late parsing and demonstrate the core problem here.
This patch fixes the test and provides deterministic behavior which is
a strict improvement over the prior state.
I've cleaned up some of the code here as well to be explicit about
inserting when that is what is actually going on.
llvm-svn: 233264
deterministically.
This fixes a latent issue where even Clang's Sema (and diagnostics) were
non-deterministic in the face of this pragma. The fix is super simple --
just use a MapVector so we track the order in which these are parsed (or
imported). Especially considering how rare they are, this seems like the
perfect tradeoff. I've also simplified the client code with judicious
use of auto and range based for loops.
I've added some pretty hilarious code to my stress test which now
survives the binary diff without issue.
llvm-svn: 233261
updated decl contexts get emitted.
Since this code was added, we have newer vastly simpler code for
handling this. The code I'm removing was very expensive and also
generated unstable order of declarations which made module outputs
non-deterministic.
All of the tests continue to pass for me and I'm able to check the
difference between the .pcm files after merging modules together.
llvm-svn: 233251
non-visible definition, skip the new definition and make the old one visible
instead of trying to parse it again and failing horribly. C++'s ODR allows
us to assume that the two definitions are identical.
llvm-svn: 233250
decl context lookup tables.
The first attepmt at this caused problems. We had significantly more
sources of non-determinism that I realized at first, and my change
essentially turned them from non-deterministic output into
use-after-free. Except that they weren't necessarily caught by tools
because the data wasn't really freed.
The new approach is much simpler. The first big simplification is to
inline the "visit" code and handle this directly. That works much
better, and I'll try to go and clean up the other caller of the visit
logic similarly.
The second key to the entire approach is that we need to *only* collect
names into a stable order at first. We then need to issue all of the
actual 'lookup()' calls in the stable order of the names so that we load
external results in a stable order. Once we have loaded all the results,
the table of results will stop being invalidated and we can walk all of
the names again and use the cheap 'noload_lookup()' method to quickly
get the results and serialize them.
To handle constructors and conversion functions (whose names can't be
stably ordered) in this approach, what we do is record only the visible
constructor and conversion function names at first. Then, if we have
any, we walk the decls of the class and add those names in the order
they occur in the AST. The rest falls out naturally.
This actually ends up simpler than the previous approach and seems much
more robust.
It uncovered a latent issue where we were building on-disk hash tables
for lookup results when the context was a linkage spec! This happened to
dodge all of the assert by some miracle. Instead, add a proper predicate
to the DeclContext class and use that which tests both for function
contexts and linkage specs.
It also uncovered PR23030 where we are forming somewhat bizarre negative
lookup results. I've just worked around this with a FIXME in place
because fixing this particular Clang bug seems quite hard.
I've flipped the first part of the test case I added for stability back
on in this commit. I'm taking it gradually to try and make sure the
build bots are happy this time.
llvm-svn: 233249
More than 2x speedup on modules builds with large redecl chains.
Roughly 15-20% speedup on non-modules builds for very large TUs.
Between 2-3% cost in memory on large TUs.
llvm-svn: 233228
lookup tables, we need to establish a stable ordering for constructing
the hash table. This is trickier than it might seem.
Most of these cases are easily handled by sorting the lookup results
associated with a specific name that has an identifier. However for
constructors and conversion functions, the story is more complicated.
Here we need to merge all of the constructors or conversion functions
together and this merge needs to be stable. We don't have any stable
ordering for either constructors or conversion functions as both would
require a stable ordering across types.
Instead, when we have constructors or conversion functions in the
results, we reconstruct a stable order by walking the decl context in
lexical order and merging them in the order their particular declaration
names are encountered. This doesn't generalize as there might be found
declaration names which don't actually occur within the lexical context,
but for constructors and conversion functions it is safe. It does
require loading the entire decl context if necessary to establish the
ordering but there doesn't seem to be a meaningful way around that.
Many thanks to Richard for talking through all of the design choices
here. While I wrote the code, he guided all the actual decisions about
how to establish the order of things.
No test case yet because the test case I have doesn't pass yet -- there
are still more sources of non-determinism. However, this is complex
enough that I wanted it to go into its own commit in case it causes some
unforseen issue or needs to be reverted.
llvm-svn: 233156
There are two aspects of non-determinism fixed here, which was the
minimum required to cause at least an empty module to be deterministic.
First, the random number signature is only inserted into the module when
we are building modules implicitly. The use case for these random
signatures is to work around the very fact that modules are not
deterministic in their output when working with the implicitly built and
populated module cache. Eventually this should go away entirely when
we're confident that Clang is producing deterministic output.
Second, the on-disk hash table is populated based on the order of
iteration over a DenseMap. Instead, use a MapVector so that we can walk
it in insertion order.
I've added a test that an empty module, when built twice, produces the
same binary PCM file.
llvm-svn: 233115
Previously we'd deserialize the list of mem-initializers for a constructor when
we deserialized the declaration of the constructor. That could trigger a
significant amount of unnecessary work (pulling in all base classes
recursively, for a start) and was causing problems for the modules buildbot due
to cyclic deserializations. We now deserialize these on demand.
This creates a certain amount of duplication with the handling of
CXXBaseSpecifiers; I'll look into reducing that next.
llvm-svn: 233052
* Strength reduce a std::function to a function pointer,
* Factor out checking the AST file magic number,
* Add a brief doc comment to readAStFileSignature
Thanks to Chandler for spotting these oddities.
llvm-svn: 233050
If there is at least one 'copyprivate' clause is associated with the single directive, the following code is generated:
```
i32 did_it = 0; \\ for 'copyprivate' clause
if(__kmpc_single(ident_t *, gtid)) {
SingleOpGen();
__kmpc_end_single(ident_t *, gtid);
did_it = 1; \\ for 'copyprivate' clause
}
<copyprivate_list>[0] = &var0;
...
<copyprivate_list>[n] = &varn;
call __kmpc_copyprivate(ident_t *, gtid, <copyprivate_list_size>,
<copyprivate_list>, <copy_func>, did_it);
...
void<copy_func>(void *LHSArg, void *RHSArg) {
Dst = (void * [n])(LHSArg);
Src = (void * [n])(RHSArg);
Dst[0] = Src[0];
... Dst[n] = Src[n];
}
```
All list items from all 'copyprivate' clauses are gathered into single <copyprivate list> (<copyprivate_list_size> is a size in bytes of this list) and <copy_func> is used to propagate values of private or threadprivate variables from the 'single' region to other implicit threads from outer 'parallel' region.
Differential Revision: http://reviews.llvm.org/D8410
llvm-svn: 232932
for a DeclContext, and fix propagation of exception specifications along
redeclaration chains.
This reverts r232905, r232907, and r232907, which reverted r232793, r232853,
and r232853.
One additional change is present here to resolve issues with LLDB: distinguish
between whether lexical decls missing from the lookup table are local or are
provided by the external AST source, and still look in the external source if
that's where they came from.
llvm-svn: 232928
The linear variable is privatized (similar to 'private') and its
value on current iteration is calculated, similar to the loop
counter variables.
Differential revision: http://reviews.llvm.org/D8375
llvm-svn: 232890
give an exception specification to a declaration that didn't have an exception
specification in any of our imported modules, emit an update record ourselves.
Without this, code importing the current module would not see an exception
specification that we could see and might have relied on.
llvm-svn: 232870
When we need to build the lookup table for a DeclContext, we used to pull in
all lexical declarations for the context; instead, just build a lookup table
for the local lexical declarations. We previously didn't guarantee that the
imported declarations would be in the returned map, but in some cases we'd
happen to put them all in there regardless. Now we're even lazier about this.
This unnecessary work was papering over some other bugs:
- LookupVisibleDecls would use the DC for name lookups in the TU in C, and
this was not guaranteed to find all imported names (generally, the DC for
the TU in C is not a reliable place to perform lookups). We now use an
identifier-based lookup mechanism for this.
- We didn't actually load in the list of eagerly-deserialized declarations
when importing a module (so external definitions in a module wouldn't be
emitted by users of those modules unless they happened to be deserialized
by the user of the module).
llvm-svn: 232793
Now that SmallString is a first-class citizen, most SmallString::str()
calls are not required. This patch removes a whole bunch of them, yet
there are lots more.
There are two use cases where str() is really needed:
1) To use one of StringRef member functions which is not available in
SmallString.
2) To convert to std::string, as StringRef implicitly converts while
SmallString do not. We may wish to change this, but it may introduce
ambiguity.
llvm-svn: 232622
consumers of that module.
Previously, such a file would only be available if the module happened to
actually import something from that module.
llvm-svn: 232583
namespace to not merge properly.
We have an invariant here: after a declaration reads its canonical declaration,
it can assume the canonical declaration is fully merged. This invariant can be
violated if deserializing some declaration triggers the deserialization of a
later declaration, because that later declaration can in turn deserialize a
redeclaration of that first declaration before it is fully merged.
The anonymous namespace for a namespace gets stored with the first declaration
of that namespace, which may be before its parent namespace, so defer loading
it until after we've finished merging the surrounding namespace.
llvm-svn: 232455
building its redecl chains, make sure we pull in the redeclarations of those
canonical declarations.
It's pretty difficult to reach a situation where we can find more canonical
declarations of an entity while building its redecl chains; I think the
provided testcase (4 modules and 7 declarations) cannot be reduced further.
llvm-svn: 232411
with a subset of the existing target CPU features or mismatched CPU
names.
While we can't check that the CPU name used to build the module will end
up being able to codegen correctly for the translation unit, we actually
check that the imported features are a subset of the existing features.
While here, rewrite the code to use std::set_difference and have it
diagnose all of the differences found.
Test case added which walks the set relationships and ensures we
diagnose all the right cases and accept the others.
No functional change for implicit modules here, just better diagnostics.
llvm-svn: 232248
headers even if they arrived when merging non-system modules.
The idea of this code is that we don't want to warn the user about
macros defined multiple times by their system headers with slightly
different definitions. We should have this behavior if either the
macro comes from a system module, or the definition within the module
comes from a system header. Previously, we would warn on ambiguous
macros being merged when they came from a users modules even though they
only showed up via system headers.
By surviving this we can handle common system header macro differences
like differing 'const' qualification of pointers due to some headers
predating 'const' being valid in C code, even when those systems headers
are pre-built into a system module.
Differential Revision: http://reviews.llvm.org/D8310
llvm-svn: 232149
definition, be sure to update the definition data on all declarations, not just
the canonical one, since the pattern might not be in the list of pending
definitions (if it used to be canonical itself).
One-line fix by me; reduced testcase by Daniel Jasper!
llvm-svn: 231950
specification, update all prior declarations if the new one has an explicit
exception specification and the prior ones don't.
Patch by Vassil Vassilev! Some minor tweaking and test case by me.
llvm-svn: 231738
move the operator delete updating into a separate update record so we can cope
with updating another module's destructor's operator delete.
llvm-svn: 231735
of extern "C" declarations. This is simpler and vastly more efficient for
modules builds (we no longer need to load *all* extern "C" declarations to
determine if we have a redeclaration).
No functionality change intended.
llvm-svn: 231538
We used to save out and eagerly load a (potentially huge) table of merged
formerly-canonical declarations when we loaded each module. This was extremely
inefficient in the presence of large amounts of merging, and didn't actually
save any merging lookup work, because we still needed to perform name lookup to
check that our merged declaration lists were complete. This also resulted in a
loss of laziness -- even if we only needed an early declaration of an entity, we
would eagerly pull in all declarations that had been merged into it regardless.
We now store the relevant fragments of the table within the declarations
themselves. In detail:
* The first declaration of each entity within a module stores a list of first
declarations from imported modules that are merged into it.
* Loading that declaration pre-loads those other entities, so that they appear
earlier within the redeclaration chain.
* The name lookup tables list the most recent local lookup result, if there
is one, or all directly-imported lookup results if not.
llvm-svn: 231424
dynamic classes in the translation unit and check whether each one's key
function is defined when we got to the end of the TU (and when we got to the
end of each module). This is really terrible for modules performance, since it
causes unnecessary deserialization of every dynamic class in every compilation.
We now use a much simpler (and, in a modules build, vastly more efficient)
system: when we see an out-of-line definition of a virtual function, we check
whether that function was in fact its class's key function. (If so, we need to
emit the vtable.)
llvm-svn: 230830
undeserialized specializations (because we merged an imported declaration of
the same template since we last added one), don't bother reading in the
specializations themselves just so we can write out their IDs again.
llvm-svn: 230805
This is a necessary prerequisite for debugging with modules.
The .pcm files become containers that hold the serialized AST which allows
us to store debug information in the module file that can be shared by all
object files that were built importing the module.
This reapplies r230044 with a fixed configure+make build and updated
dependencies and testcase requirements. Over the last iteration this
version adds
- missing target requirements for testcases that specify an x86 triple,
- a missing clangCodeGen.a dependency to libClang.a in the make build.
rdar://problem/19104245
llvm-svn: 230423
This is a necessary prerequisite for debugging with modules.
The .pcm files become containers that hold the serialized AST which allows
us to store debug information in the module file that can be shared by all
object files that were built importing the module.
rdar://problem/19104245
This reapplies r230044 with a fixed configure+make build and updated
dependencies. Take 3.
llvm-svn: 230305
invalidate lookup_iterators and lookup_results for some name within a
DeclContext if the lookup results for a *different* name change.
llvm-svn: 230121
This is a necessary prerequisite for debugging with modules.
The .pcm files become containers that hold the serialized AST which allows
us to store debug information in the module file that can be shared by all
object files that were built importing the module.
rdar://problem/19104245
This reapplies r230044 with a fixed configure+make build and updated
dependencies. Take 2.
llvm-svn: 230089
This is a necessary prerequisite for debugging with modules.
The .pcm files become containers that hold the serialized AST which allows
us to store debug information in the module file that can be shared by all
object files that were built importing the module.
rdar://problem/19104245
This reapplies r230044 with a fixed configure+make build and updated
dependencies.
llvm-svn: 230067
There are two issues here:
1) It's too late to rebuild at this point, because we won't go through
removeModules and when we try to reload the new .pcm we'll get the old
one instead. We might be able to call removeModules after an OutOfDate
here, but I'm not yet confident that it is always safe to do so.
2) In practice, this check fails spuriously when the umbrella header
appears to change because of a VFS change that means it maps to a
different copy of the same file. Because of this, we just skip the
check for now.
llvm-svn: 230064
This is a necessary prerequisite for debugging with modules.
The .pcm files become containers that hold the serialized AST which allows
us to store debug information in the module file that can be shared by all
object files that were built importing the module.
rdar://problem/19104245
llvm-svn: 230044
the one in the current compiler invocation. If they differ reject the PCH.
This protects against the badness occurring from getting modules loaded from different module caches (see crashes).
rdar://19889860
llvm-svn: 229909
The member gets invalidated as elements are added to the dense set. Directly
access the underlying pointer instead. Not sure how to create a test case for
this :-(. Maybe Richard can help.
llvm-svn: 229673
entity, put the originally-canonical decl IDs in the right places in the redecl
chain rather than reordering them all to the start. If we don't ensure that the
redecl chain order is consistent with the topological module order, we can fail
to make a declaration visible if later declarations are in more IDNSs than
earlier ones (for instance, because the earlier decls are invisible friends).
llvm-svn: 228978
When mangling the module map path into a .pcm file name, also mangle the
IsSystem bit, which can also depend on the header search paths. For
example, the user may change from -I to -isystem. This can affect
diagnostics in the importing TU.
llvm-svn: 228966
context as anonymous for merging purposes. They can't be found by their names,
so we merge them based on their position within the surrounding context.
llvm-svn: 228485
of that entity, ensure that the redeclaration chain is reordered properly on
reload. Otherwise, the result of name lookup for that entity may point to an
entity that is too old; if that's an injected friend name or the like, that
can result in the name not being found at all.
llvm-svn: 228371
Richard Smith