LLVM supports applying conversion instructions to vectors of the same number of
elements (fptrunc, fptosi, etc.) but there had been no way for a Clang user to
cause such instructions to be generated when using builtin vector types.
C-style casting on vectors is already defined in terms of bitcasts, and so
cannot be used for these conversions as well (without leading to a very
confusing set of semantics). As a result, this adds a __builtin_convertvector
intrinsic (patterned after the OpenCL __builtin_astype intrinsic). This is
intended to aid the creation of vector intrinsic headers that create generic IR
instead of target-dependent intrinsics (in other words, this is a generic
_mm_cvtepi32_ps). As noted in the documentation, the action of
__builtin_convertvector is defined in terms of the action of a C-style cast on
each vector element.
llvm-svn: 190915
GCC ToT doesn't do this & it's worth about 3.2% on Clang's DWO file size
with Clang. Some or all of this may be due to things like r190715 which
could have source fixes/improvements, but it's not clear that's the case
and that doesn't help other source bases.
llvm-svn: 190716
This restores the sqrt -> llvm.sqrt mapping, but only in fast-math mode
(specifically, when the UnsafeFPMath or NoNaNsFPMath CodeGen options are
enabled). The @llvm.sqrt* intrinsics have slightly different semantics from the
libm call, specifically, they are undefined when given a non-zero negative
number (the libm calls will always return NaN for any negative number).
This mapping was removed in r100613, and replaced with a TODO, but at that time
the fast-math flags were not yet implemented. Now that we have these, restoring
this mapping is important because it will enable autovectorization of sqrt
calls in loops (at least in fast-math mode).
llvm-svn: 190646
The code in CGExpr was added back in 2012 (r165536) but not exercised in tests
until recently.
Detected on the MemorySanitizer bootstrap bot.
llvm-svn: 190521
Static locals requiring initialization are not thread safe on Windows.
Unfortunately, it's possible to create static locals that are actually
externally visible with inline functions and templates. As a result, we
have to implement an initialization guard scheme that is compatible with
TUs built by MSVC, which makes thread safety prohibitively difficult.
MSVC's scheme is that every function that requires a guard gets an i32
bitfield. Each static local is assigned a bit that indicates if it has
been initialized, up to 32 bits, at which point a new bitfield is
created. MSVC rejects inline functions with more than 32 static locals,
and the externally visible mangling (?_B) only allows for one guard
variable per function.
On Eli's recommendation, I used MangleNumberingContext to track which
bit each static corresponds to.
Implements PR16888.
Reviewers: rjmccall, eli.friedman
Differential Revision: http://llvm-reviews.chandlerc.com/D1416
llvm-svn: 190427
By removing the possibility of strange partial definitions with no
members that older GCC's produced for the otherwise unreferenced outer
types of referenced inner types, we can simplify debug info generation
and correct this bug. Newer (4.8.1 and ToT) GCC's don't produce this
quirky debug info, and instead produce the full definition for the outer
type (except in the case where that type is dynamic and its vtable is
not emitted in this TU).
During the creation of the context for a type, we may revisit that type
(due to the need to visit template parameters, among other things) and
used to end up visiting it first there. Then when we would reach the
original code attempting to define that type, we would lose debug info
by overwriting its members.
By avoiding the possibility of latent "defined with no members" types,
we can be sure than whenever we already have a type in a cache (either a
definition or declaration), we can just return that. In the case of a
full definition, our work is done. In the case of a partial definition,
we must already be in the process of completing it. And in the case of a
declaration, the completed/vtable/etc callbacks can handle converting it
to a definition.
llvm-svn: 190122
Debug info emission was tripping over an IRGen bug (fixed in r189996)
that was resulting in duplicate emission of static data members of class
templates in namespaces.
We could add more test coverage to debug info for this issue
specifically, but I think the underlying IRGen test is more targeted and
sufficient for the issue.
llvm-svn: 190001
A quirk of AST representation leads to class template static data member
definitions being visited twice during Clang IRGen resulting in
duplicate (benign) initializers.
Discovered while investigating a possibly-related debug info bug tickled
by the duplicate emission of these members & their associated debug
info.
With thanks to Richard Smith for help investigating, understanding, and
helping with the fix.
llvm-svn: 189996
I tried to implement this properly in r189051, but I didn't have enough
test coverage. Richard kindly provided more test cases than I could
possibly imagine and now we should have the correct condition.
llvm-svn: 189898
This fixes pr13124.
From the discussion at
http://lists.cs.uiuc.edu/pipermail/cfe-dev/2012-June/022606.html
we know that we cannot make funcions in a weak_odr vtable also weak_odr. They
should remain linkonce_odr.
The side effect is that we cannot emit a available_externally vtable unless we
also emit a copy of the function. This also has an issue: If codegen is going
to output a function, sema has to mark it used. Given llvm.org/pr9114, it looks
like sema cannot be more aggressive at marking functions used because
of vtables.
This leaves us with a few unpleasant options:
* Marking functions in vtables used if possible. This sounds a bit sloppy, so
we should avoid it.
* Producing available_externally vtables only when all the functions in it are
already used or weak_odr. This would cover cases like
--------------------
struct foo {
virtual ~foo();
};
struct bar : public foo {
virtual void zed();
};
void f() {
foo *x(new bar);
delete x;
}
void g(bar *x) {
x->~bar(); // force the destructor to be used
}
--------------------------
and
----------------------------------
template<typename T>
struct bar {
virtual ~bar();
};
template<typename T>
bar<T>::~bar() {
}
// make the destructor weak_odr instead of linkonce_odr
extern template class bar<int>;
void f() {
bar<int> *x(new bar<int>);
delete x;
}
----------------------------
These look like corner cases, so it is unclear if it is worth it.
* And finally: Just nuke this optimization. That is what this patch implements.
llvm-svn: 189852
We use CXX mangler to generate unique identifier for external C++ struct,
union, class and enum. Types with unique identifier are added to retained
types by DIBuilder.
Testing cases are updated to reflect the unique identifier generated for types.
The order of MDNodes is changed because of retained types and testing cases
are updated accordingly.
Testing case debug-info-uuid.cpp now emits error with Itanium mangler, since
uuid is not yet handled in Itanium mangler. And it will check for the error
message.
llvm-svn: 189622
We had further discussions on how to retain types, whether to do it in front end
or in DIBuilder. And we agree to do it in DIBuilder so front ends
generating unique identifier do not need to worry about retaining them.
llvm-svn: 189609
We use CXX mangler to generate unique identifier for external C++ struct,
union, class and enum. Types with unique identifier are added to RetainedTypes
to make sure they are treated as used even when all uses are replaced with
the identifiers.
A single type can be added to RetainedTypes multiple times. For example, both
createForwardDecl and createLimitedType can add the same type to RetainedTypes.
A set is used to avoid duplication when updating AllRetainTypes in DIBuilder.
Testing cases are updated to reflect the unique identifier generated for types.
The order of MDNodes is changed because of retained types and testing cases
are updated accordingly.
Testing case debug-info-uuid.cpp now emits error with Itanium mangler, since
uuid is not yet handled in Itanium mangler.
We choose to update RetainedTypes in clang, then at finalize(), we update
AllRetainTypes in DIBuilder. The other choice is to update AllRetainTypes
in DIBuilder when creating a DICompositeType with unique identifier. This
option requires using ValueHandle for AllRetainTypes in DIBuilder since
the created DICompositeType can be modified later on by setContainingType etc.
llvm-svn: 189600
Both functions will take a Type pointer instead of a Decl pointer. This helps
with follow-up type uniquing patches, which need the Type pointer to call
CXX mangler to generate unique identifiers.
No functionality change.
llvm-svn: 189519
In the transition from declaration (with some members) to definition, we
were overwriting the list of members with the empty list when attaching
template parameters.
The fix is in llvm::DICompositeType::addMember (along with asserts that
cause this bug to be covered by existing Clang test cases), including
adding some asserts to catch this sort of issue which found issues fixed
in this commit.
llvm-svn: 189494
These operations "vector add high-half narrow" actually correspond to the
sequence:
%sum = add <4 x i32> %lhs, %rhs
%high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16>
%res = trunc <4 x i32> %high to <4 x i16>
Now that LLVM can spot this, Clang should emit the corresponding LLVM IR.
llvm-svn: 189463
The NEON intrinsics vqdmlal and vqdmlsl are really just combinations of a
saturating-doubling-multiply (vqdmull) and a saturating add/sub, so now that
LLVM can spot those patterns Clang should emit them instead of specialised
intrinsics.
Feature already tested by existing ARM NEON intrinsics tests.
llvm-svn: 189462
This reverts commit r189320.
Alexey Samsonov and Dmitry Vyukov presented some arguments for keeping
these around - though it still seems like those tasks could be solved by
a tool just using the symbol table. In a very small number of cases,
thunks may be inlined & debug info might be able to save profilers &
similar tools from misclassifying those cases as part of the caller.
The extra changes here plumb through the VarDecl for various cases to
CodeGenFunction - this provides better fidelity through a few APIs but
generally just causes the CGF::StartFunction to fallback to using the
name of the IR function as the name in the debug info.
The changes to debug-info-global-ctor-dtor.cpp seem like goodness. The
two names that go missing (in favor of only emitting those names as
linkage names) are names that can be demangled - emitting them only as
the linkage name should encourage tools to do just that.
Again, thanks to Dinesh Dwivedi for investigation/work on this issue.
llvm-svn: 189421
Summary:
Makes functions with implicit calling convention compatible with
function types with a matching explicit calling convention. This fixes
things like calls to qsort(), which has an explicit __cdecl attribute on
the comparator in Windows headers.
Clang will now infer the calling convention from the declarator. There
are two cases when the CC must be adjusted during redeclaration:
1. When defining a non-inline static method.
2. When redeclaring a function with an implicit or mismatched
convention.
Fixes PR13457, and allows clang to compile CommandLine.cpp for the
Microsoft C++ ABI.
Excellent test cases provided by Alexander Zinenko!
Reviewers: rsmith
Differential Revision: http://llvm-reviews.chandlerc.com/D1231
llvm-svn: 189412
This was added in r166676 based on PR13942 on the basis that tools may
need debug information for any executable code/function for some fairly
broad/non-specific purposes. It seems to me (as noted in PR14569) that
the major/only purpose is in backtraces, which should generally not
apply to thunks as they won't appear in the stack themselves. By
removing them we fix PR14569 and reduce the size of Clang's debug info.
Strangely enough this doesn't seem to have a substantial impact on
Clang's self-hosted debug info (at least looking at DWO file size) size
at all. Not sure if I failed to test this correctly but I only observed
a 0.004% change in DWO file size over Clang+LLVM.
With thanks to Dinesh Dwivedi for work on this PR.
llvm-svn: 189320
CodeGenFunction is run on only one function - a new object is made for
each new function. I would add an assertion/flag to this effect, but
there's an exception: ObjC properties involve emitting helper functions
that are all emitted by the same CodeGenFunction object, so such a check
is not possible/correct.
llvm-svn: 189277
- __func__ or __FUNCTION__ returns captured statement's parent
function name, not the one compiler generated.
Differential Revision: http://llvm-reviews.chandlerc.com/D1491
Reviewed by bkramer
llvm-svn: 189219
