MS inline asm: Allow __asm blocks to set a return value
If control falls off the end of a function after an __asm block, MSVC
assumes that the inline assembly filled the EAX and possibly EDX
registers with an appropriate return value. This functionality is used
in inline functions returning 64-bit integers in system headers, so we
need some amount of compatibility.
This is implemented in Clang by adding extra output constraints to every
inline asm block, and storing the resulting output registers into the
return value slot. If we see an asm block somewhere in the function
body, we emit a normal epilogue instead of marking the end of the
function with a return type unreachable.
Normal returns in functions not using this functionality will overwrite
the return value slot, and in most cases LLVM should be able to
eliminate the dead stores.
David Blaikie [Thu, 4 Sep 2014 16:01:24 +0000 (16:01 +0000)]
Fix the clang -Werror build after r217152 by flagging a polymorphically used but statically owned hierarchy with a protected base dtor and final classes to satisfy -Wnon-virtual-dtor
Daniel Sanders [Thu, 4 Sep 2014 15:05:39 +0000 (15:05 +0000)]
[mips] Mark aggregates returned in registers with the 'inreg' attribute.
Summary:
This allows us to easily find them in the backend after the aggregates have
been lowered to other types. This is important on big-endian targets using
the N32/N64 ABI's since these ABI's must shift small structures into the
upper bits of the register.
Refactor VariantMatcher::MatcherOps to reduce the amount of generated code.
Summary:
Refactor VariantMatcher::MatcherOps to reduce the amount of generated code.
- Make some code type agnostic and move it to the cpp file.
- Return a DynTypedMatcher instead of storing the object in MatcherOps.
This change reduces the number of symbols generated in Registry.cpp by
~19%, the object byte size by ~17% and the compilation time (in non-release mode) by ~20%.
Daniel Sanders [Thu, 4 Sep 2014 13:28:14 +0000 (13:28 +0000)]
[mips] Zero-sized structs cannot be ignored in MipsABIInfo::classifyReturnType() for O32
Summary:
They are returned indirectly which causes the other arguments to move to
the next argument slot.
With this, utils/ABITest does not discover any failing cases in the first
500 attempts on big/little endian for O32. Previously some of these failed.
Also tested N32/N64 little endian (big endian has other known issues) with
no issues.
Oliver Stannard [Thu, 4 Sep 2014 10:38:53 +0000 (10:38 +0000)]
ARM: Default to apcs-gnu ABI for NetBSD
r216662 changed the default ABI for 32-bit ARM targets to be "aapcs"
when no environment is given in the triple, however NetBSD requires it
to be "apcs-gnu".
Richard Smith [Wed, 3 Sep 2014 23:11:22 +0000 (23:11 +0000)]
[modules] Make NamespaceAliasDecl redeclarable, as it should be. This fixes
merging of namespace aliases across modules and improves source fidelity.
Incidentally also fixes PR20816.
Anton Yartsev [Wed, 3 Sep 2014 17:59:21 +0000 (17:59 +0000)]
Enhance the 'Vectors and Extended Vectors' section.
Added cast operations to the table of vector operations. Supported status 'no' means that there are no tests in the Clang test suite for the given cast.
Eli Bendersky [Wed, 3 Sep 2014 15:27:03 +0000 (15:27 +0000)]
Split off CUDA-specific Sema parts to a new file
In line with SemaOpenMP.cpp, etc. CUDA-specific semantic analysis code goes into
a separate file. This is in anticipation of adding extra functionality here in
the near future.
Tom Stellard [Wed, 3 Sep 2014 15:24:29 +0000 (15:24 +0000)]
CGBuiltin: Use @llvm.fabs rather than fabs libcall when emitting builtins
Using the intrinsic allows the SelectionDAGBuilder to turn this call
into the FABS Node and also the intrinsic is something the vectorizer knows
how to vectorize.
This patch also sets the readnone attribute on this call, which should
enable additional optmizations.
Benjamin Kramer [Wed, 3 Sep 2014 12:08:14 +0000 (12:08 +0000)]
ASTMatchers: Add a matcher to detect whether a decl or stmt is inside a template instantiation.
This is hoisted from clang-tidy where it's used everywhere. The implementation
is not particularly efficient right now, but there is no easy fix for that.
Ed Schouten [Wed, 3 Sep 2014 06:00:11 +0000 (06:00 +0000)]
Allow a scoped lockable object to acquire/release multiple locks.
Scoped lockable objects (mutex guards) are implemented as if it is a
lock itself that is acquired upon construction and unlocked upon
destruction. As it if course needs to be used to actually lock down
something else (a mutex), it keeps track of this knowledge through its
underlying mutex field in its FactEntry.
The problem with this approach is that this only allows us to lock down
a single mutex, so extend the code to use a vector of underlying
mutexes. This, however, makes the code a bit more complex than
necessary, so subclass FactEntry into LockableFactEntry and
ScopedLockableFactEntry and move all the logic that differs between
regular locks and scoped lockables into member functions.
Richard Smith [Wed, 3 Sep 2014 02:33:22 +0000 (02:33 +0000)]
[modules] Use DeclContext::equals rather than == on DeclContext* when
determining whether a declaration is out of line, instead of assuming
that the semantic and lexical DeclContext will be the same declaration
whenever they're the same entity.
This fixes behavior of declarations within merged classes and enums.
Ed Schouten [Tue, 2 Sep 2014 20:59:13 +0000 (20:59 +0000)]
Use /usr/bin/env python instead of /usr/bin/python.
On operating systems like the BSDs, it is typically the case that
/usr/bin/python does not exist. We should therefore use /usr/bin/env
instead. This is also done in various other scripts in tools/.
NAKAMURA Takumi [Sun, 31 Aug 2014 12:21:50 +0000 (12:21 +0000)]
clang/test/CXX/drs/dr5xx.cpp: Suppress dr547 for targeting MSVC x86 for now, due to incompatibility of attribute(thiscall).
With targeting i686-win32,
error: 'error' diagnostics seen but not expected:
File clang/test/CXX/drs/dr5xx.cpp Line 521: implicit instantiation of undefined template 'dr547::X<void () __attribute__((thiscall)) const>'
File clang/test/CXX/drs/dr5xx.cpp Line 518: implicit instantiation of undefined template 'dr547::X<void () __attribute__((thiscall)) const>'
File clang/test/CXX/drs/dr5xx.cpp Line 518: implicit instantiation of undefined template 'dr547::X<void () __attribute__((thiscall)) const>'
error: 'note' diagnostics seen but not expected:
File clang/test/CXX/drs/dr5xx.cpp Line 516: template is declared here
File clang/test/CXX/drs/dr5xx.cpp Line 521: in instantiation of function template specialization 'dr547::f<void () __attribute__((thiscall)) const, dr547::S>' requested here
File clang/test/CXX/drs/dr5xx.cpp Line 516: template is declared here
File clang/test/CXX/drs/dr5xx.cpp Line 516: template is declared here
7 errors generated.
Craig Topper [Sat, 30 Aug 2014 16:55:39 +0000 (16:55 +0000)]
Use llvm::makeArrayRef instead of explicitly calling ArrayRef constructor and mentioning the type. This works now that we have a conversion from ArrayRef<T*> to ArrayRef<const T*>.
Richard Smith [Sat, 30 Aug 2014 00:04:23 +0000 (00:04 +0000)]
[modules] Fix deserialization cycle when loading a tag declaration with a typedef name for linkage purposes. When loading the type, delay loading its typedef until we've finished loading and merging the type. In its place, save out the name of the typedef, which we need for merging purposes.
David Blaikie [Fri, 29 Aug 2014 22:04:40 +0000 (22:04 +0000)]
unique_ptrify some parameters to PTHManager::PTHManager
A couple of these arguments were passed by void* as a rather extreme
example of pimpling. Adjusting this to a more classic form of the idiom
(involving forward declarations) makes this more legible and allows
explicit passing of ownership via std::unique_ptr.
Objective-C [qoi]. If property is going to be implemented
in the super class, do not issue the warning about property
in current class's protocol will not be auto synthesized.
// rdar://18179833
Anna Zaks [Fri, 29 Aug 2014 20:01:38 +0000 (20:01 +0000)]
Add an option to silence all analyzer warnings.
People have been incorrectly using "-analyzer-disable-checker" to
silence analyzer warnings on a file, when analyzing a project. Add
the "-analyzer-disable-all-checks" option, which would allow the
suppression and suggest it as part of the error message for
"-analyzer-disable-checker". The idea here is to compose this with
"--analyze" so that users can selectively opt out specific files from
static analysis.
David Blaikie [Fri, 29 Aug 2014 19:57:52 +0000 (19:57 +0000)]
unique_ptrify thep passing of BugReports to BugReportEquivClass
I suspect llvm::ilist should take elements by unique_ptr, since it does
take ownership of the element (by stitching it into the linked list) -
one day.
Objective-C. Tweak diagnosing properties that are not auto-synthesized.
Do not warn when property declared in class's protocol will be auto-synthesized
by its uper class implementation because super class has also declared this
property while this class has not. Continue to warn if current class
has declared the property also (because this declaration will not result
in a 2nd synthesis).
rdar://18152478
Again, if shared ownership is the right model here (I assume it is,
given graph algorithms & such) this could be tidied up (the 'release'
call removed in favor of something safer) by having
IntrunsiveRefCntPointer constructible from a unique_ptr.
(& honestly I'd probably favor taking a page out of shared_ptr's book,
allowing implicit construction from a unique_ptr rvalue, and only allow
explicit from a raw pointer - currently IntrusiveRefCntPointer can
implicitly own from a raw pointer, which seems unsafe)
David Blaikie [Fri, 29 Aug 2014 16:53:14 +0000 (16:53 +0000)]
Provide a BuryPointer for unique_ptrs.
In theory, it'd be nice if we could move to a case where all buried
pointers were buried via unique_ptr to demonstrate that the program had
finished with the value (that we could really have cleanly deallocated
it) but instead chose to bury it.
I think the main reason that's not possible right now is the various
IntrusiveRefCntPtrs in the Frontend, sharing ownership for a variety of
compiler bits (see the various similar
"CompilerInstance::releaseAndLeak*" functions). I have yet to figure out
their correct ownership semantics - but perhaps, even if the
intrusiveness can be removed, the shared ownership may yet remain and
that would lead to a non-unique burying as is there today. (though we
could model that a little better - by passing in a shared_ptr, etc -
rather than needing the two step that's currently used in those other
releaseAndLeak* functions)
This might be a bit more robust if BuryPointer took the boolean:
BuryPointer(bool, unique_ptr<T>)
and the choice to bury was made internally - that way, even when
DisableFree was not set, the unique_ptr would still be null in the
caller and there'd be no chance of accidentally having a different
codepath where the value is used after burial in !DisableFree, but it
becomes null only in DisableFree, etc...
Benjamin Kramer [Fri, 29 Aug 2014 11:22:47 +0000 (11:22 +0000)]
ASTMatchers: Replace some copies of the bound nodes tree builder with moves.
But don't move if all we do is clearing the thing. The move method is too large
to be inlined and performs a ton of unnecessary checking when the RHS is empty.
James Molloy [Fri, 29 Aug 2014 10:17:52 +0000 (10:17 +0000)]
Use store size instead of alloc size when coercing.
Previously, EnterStructPointerForCoercedAccess used Alloc size when determining how to convert. This was problematic, because there were situations were the alloc size was larger than the store size. For example, if the first element of a structure were i24 and the destination type were i32, the old code would generate a GEP and a load i24. The code should compare store sizes to ensure the whole object is loaded. I have attached a test case.
This patch modifies the output of arm64-be-bitfield.c test case, but the new IR seems to be equivalent, and after -O3, the compiler generates identical ARM assembly. (asr x0, x0, #54)
Summary:
We did a great job getting this wrong:
- We messed up which LLVM IR types to use for arguments and return values.
The optimized libcalls use integer types for values.
Clang attempted to use the IR type which corresponds to the value
passed in instead of using an appropriately sized integer type. This
would result in violations of the ABI for, as an example, floating
point types.
- We didn't bother recording the result of the atomic libcall in the
destination memory.
Instead, call the functions with arguments matching the type of the
libcall prototype's parameters.
David Blaikie [Fri, 29 Aug 2014 07:25:23 +0000 (07:25 +0000)]
unique_ptrify Driver Action handling
It hits a limit when we reach ActionList, which is used for dynamic
conditional ownership, so we lose type safety there.
This did expose at least one caller "lying" about ownership (passing
ownership to an Action, then updating the Action to specify that it
doesn't actually own the thing that was passed) - changing this to
unique_ptr just makes that oddity more obvious.
David Blaikie [Fri, 29 Aug 2014 06:34:53 +0000 (06:34 +0000)]
Improve unique_ptr-y ownership in ASTUnit::ComputePreamble
Rather than having a pair of pairs and a reference out parameter, build
a structure with everything together and named. A raw pointer and a
unique_ptr, rather than a raw pointer and a boolean, are used to
communicate ownership transfer.
It's possible one day we'll end up with a conditional pointer (probably
represented by a raw pointer and a boolean) abstraction to use in places
like this. Conditional ownership seems to be coming up more often than
I'd hoped...
Craig Topper [Fri, 29 Aug 2014 06:05:01 +0000 (06:05 +0000)]
Use llvm::makeArrayRef instead of explicitly calling ArrayRef constructor and mentioning the type. This works now that we have a conversion from ArrayRef<T*> to ArrayRef<const T*>.
call __asan_load_cxx_array_cookie when loading array cookie in asan mode.
Summary:
The current implementation of asan cookie is incorrect:
we add nosanitize metadata to the cookie load, but the metadata may be lost
and we will instrument the load from poisoned memory.
This change replaces the load with a call to __asan_load_cxx_array_cookie (r216692)
Hans Wennborg [Fri, 29 Aug 2014 00:16:06 +0000 (00:16 +0000)]
Better codegen support for DLL attributes being dropped after the first declaration (PR20792)
For the following code:
__declspec(dllimport) int f(int x);
int user(int x) {
return f(x);
}
int f(int x) { return 1; }
Clang will drop the dllimport attribute in the AST, but CodeGen would have
already put it on the LLVM::Function, and that would never get updated.
(The same thing happens for global variables.)
This makes Clang check dropped DLL attribute case each time the LLVM object
is referenced.
This isn't perfect, because we will still get it wrong if the function is
never referenced by codegen after the attribute is dropped, but this handles
the common cases and makes us not fail in the verifier.
Alexey Samsonov [Thu, 28 Aug 2014 23:51:01 +0000 (23:51 +0000)]
Don't create a null reference in ASTUnit::LoadFromCommandLine.
This change is the last in the pack of five commits
(also see r216691, r216694, r216695, and r216696) that reduces the number
of test failures in "check-clang" invocation in UBSan bootstrap
from 2443 down to 5.
Rafael Espindola [Thu, 28 Aug 2014 21:23:05 +0000 (21:23 +0000)]
Call powerpc-darwin external tools with -arch ppc.
With this patch we call external tools for powerpc-darwin with "-arch ppc"
instead of "-arch powerpc", so as to be compatible with the cctools assembler
and ld64 linker.