Add front-end infrastructure now address space casts are in LLVM IR.
With the introduction of explicit address space casts into LLVM, there's
a need to provide a new cast kind the front-end can create for C/OpenCL/CUDA
and code to produce address space casts from those kinds when appropriate.
Teach Clang to provide ::max_align_t in C11 and C++11 modes.
This definition is not chosen idly. There is an unfortunate reality with
max_align_t -- the specific nature of its definition leaks into the ABI
almost immediately. Because it is part of C11 and C++11 it becomes
essential for it to match with other systems on that ABI. There is an
effort to discourage any further use of this construct as a consequence
-- using max_align_t introduces an immediate ABI problem. We can never
update it to have larger alignment even as the microarchitecture changes
to necessitate higher alignment. =/
The particular definition here exactly matches the ABI of GCC's chosen
::max_align_t definition, for better or worse. This was written with the
help of Richard Smith who was decoding the exact ABI implications of the
selected definition in GCC. Notably, in-register arguments are impacted
by the particular definition chosen. =/
No one is under the illusion that this is a "good" or "useful"
definition of max_align_t, and we are working with the standards
committee to specify a more useful interface to address this need.
Fix an logic error in the clang driver preventing crtfastmath.o from linking when -Ofast is used without -ffast-math
In gcc using -Ofast forces linking of crtfastmath.o.
In the current clang crtfastmath.o is only linked when -ffast-math/-funsafe-math-optimizations passed. It can lead to performance issues, when using only -Ofast without explicit -ffast-math (I faced with it).
My patch fixes inconsistency with gcc behaviour and also introduces few tests on it.
[PowerPC] Make -pg generate calls to _mcount not mcount
At least on REL6 (Linux/glibc 2.12), the proper symbol for generating gprof
data is _mcount, not mcount. Prior to this change, compiling with -pg would
generate linking errors (because of unresolved references to mcount), after
this change -pg seems at least minimally functional.
Change the color of comment nodes from bright yellow to blue. Bright yellow on
a white background is difficult to read. Also include a chart showing which
colors are used by which elements in the AST dump.
For -Wconsumed, walk the namespaces to find if the top most namespace is "std"
to determine if a move function is the std::move function. This allows functions
like std::__1::move to also be treated a the move function.
Temporary fix for PR18473: Don't try to evaluate the initializer for a
type-dependent variable, even if the initializer isn't value-dependent. This
happens for ParenListExprs composed of non-value-dependent subexpressions, for
instance.
We should really give ParenListExprs (and InitListExprs) the type of the
initialized entity if they're used to represent a dependent initialization (and
if so, set them to be type-, value- and instantiation-dependent).
Once this happens, we would believe that D with a nested namespace
specifier of A was a variable that was being declared. We would go on
to complain that there was an extraneous 'template <>' on their variable
declaration.
Crashes would happen when 'A' gets defined as 'enum class A {}' as
various asserts would fire.
Instead, we should skip up until the semicolon if we see that we are in
the middle of a definition and the current token is a ':'
Make Sema::BuildCXXNestedNameSpecifier correctly clear the previous
CXXScopeSpec when necessary while performing typo correction. This fixes
the crash reported in PR18213 (the problem existed since r185487, and
r193020 made it easier to hit).
------------------------------------------------------------------------
Bill Wendling [Mon, 16 Dec 2013 02:32:55 +0000 (02:32 +0000)]
Merging r196212:
------------------------------------------------------------------------
r196212 | alp | 2013-12-02 22:13:01 -0800 (Mon, 02 Dec 2013) | 22 lines
Emit an extension warning when changing system header tokens
clang converts keywords to identifiers for compatibility with various system
headers such as GNU libc.
Implement a -Wkeyword-compat extension warning to diagnose those cases. The
warning is on by default but will generally be ignored in system headers. It
can however be enabled globally to aid standards conformance testing.
This also changes the __uptr keyword avoidance from r195710 to no longer
special-case system headers, bringing it in line with other similar workarounds
in clang.
Implementation returns bool for symmetry with token annotation functions.
Some examples:
warning: keyword '__is_pod' will be treated as an identifier for the remainder of the translation unit [-Wkeyword-compat]
struct __is_pod
warning: keyword '__uptr' will be treated as an identifier here [-Wkeyword-compat]
union w *__uptr;
------------------------------------------------------------------------
PR18246: When performing template argument deduction to decide which template
is specialized by an explicit specialization, start from the first declaration
in case we've got a member of a class template (redeclarations might not number
the template parameters the same way).
When performing an array new of a multidimensional array with an initializer
list, each element of the initializer list may provide more than one of the
base elements of the array. Be sure to initialize the right type and bump the
array pointer by the right amount.
[-cxx-abi microsoft] Mangle large integral constants correctly
Testing has revealed that large integral constants (i.e. > INT64_MAX)
are always mangled as-if they are negative, even in places where it
would not make sense for them to be negative (like non-type template
parameters of type unsigned long long).
To address this, we change the way we model number mangling: always
mangle as-if our number is an int64_t. This should result in correct
results when we have large unsigned numbers.
N.B. Bizarrely, things that are 32-bit displacements like vbptr offsets
are mangled as-if they are unsigned 32-bit numbers. This is a pretty
egregious waste of space, it would be a 4x savings if we could mangle it
like a signed 32-bit number. Instead, we explicitly cast these
displacements to uint32_t and let the mangler proceed.
This is another regression fixed by reverting r189090.
In this case, the problem is not live variables but the approach that was taken in r189090. This regression was caused by explicitly binding "true" to the condition when we take the true branch. Normally that's okay, but in this case we're planning to reuse that condition as the value of the expression.
------------------------------------------------------------------------
The original patch introduced regressions (see the added live-variables.* tests). The patch depends on the correctness of live variable analyses, which are not computed correctly. I've opened PR18159 to track the proper resolution to this problem.
The patch was a stepping block to r189746. This is why part of the patch reverts temporary destructor tests that started crashing. The temporary destructors feature is disabled by default.
------------------------------------------------------------------------
We already support using "r" on 64-bit values (a GPRPair is
allocated), but Sema doesn't know this yet so issues a warning. This
should fix it.
------------------------------------------------------------------------
Bill Wendling [Mon, 9 Dec 2013 02:00:10 +0000 (02:00 +0000)]
Merging r196712:
------------------------------------------------------------------------
r196712 | rafael | 2013-12-07 17:13:22 -0800 (Sat, 07 Dec 2013) | 12 lines
Fix pr18174.
Clang outputs LLVM one top level decl at a time. This combined with the
visibility computation code looking for the newest NamespaceDecl would cause
it to produce different results for nested namespaces.
The two options for producing consistent results are
* Delay codegen of anything inside a namespace until the end of the file.
* Don't look for the newest NamespaceDecl.
This patch implements the second option.
This matches the gcc behavior too.
------------------------------------------------------------------------
AArch64: add missing ACLE intrinsics mapping to general arithmetic operation from VFP instructions.
E.g. float64x1_t vadd_f64(float64x1_t a, float64x1_t b) -> FADD Dd, Dn, Dm.
PR17983: Fix crasher bug in C++1y mode when performing a non-global array
delete on a class which has no array cookie and has no class-specific operator
new.
CodeGen: Don't emit linkage on thunks that aren't emitted because they're vararg.
This can happen when we're trying to emit a thunk with available_externally
linkage with optimization enabled but bail because it doesn't make sense
for vararg functions.
Bill Wendling [Thu, 5 Dec 2013 18:35:37 +0000 (18:35 +0000)]
Merging r196387:
------------------------------------------------------------------------
r196387 | aaronballman | 2013-12-04 07:32:26 -0800 (Wed, 04 Dec 2013) | 1 line
When parsing ignored attribute arguments, presuming the first argument is an unresolved identifier the same way that we do for unknown arguments. This resolves PR18075, where we regressed the handling of OpenBSD's bounded attribute.
------------------------------------------------------------------------
For an init capture, process the initialization expression
right away. For lambda init-captures such as the following:
const int x = 10;
auto L = [i = x+1](int a) {
return [j = x+2,
&k = x](char b) { };
};
keep in mind that each lambda init-capture has to have:
- its initialization expression executed in the context
of the enclosing/parent decl-context.
- but the variable itself has to be 'injected' into the
decl-context of its lambda's call-operator (which has
not yet been created).
Each init-expression is a full-expression that has to get
Sema-analyzed (for capturing etc.) before its lambda's
call-operator's decl-context, scope & scopeinfo are pushed on their
respective stacks. Thus if any variable is odr-used in the init-capture
it will correctly get captured in the enclosing lambda, if one exists.
The init-variables above are created later once the lambdascope and
call-operators decl-context is pushed onto its respective stack.
Since the lambda init-capture's initializer expression occurs in the
context of the enclosing function or lambda, therefore we can not wait
till a lambda scope has been pushed on before deciding whether the
variable needs to be captured. We also need to process all
lvalue-to-rvalue conversions and discarded-value conversions,
so that we can avoid capturing certain constant variables.
For e.g.,
void test() {
const int x = 10;
auto L = [&z = x](char a) { <-- don't capture by the current lambda
return [y = x](int i) { <-- don't capture by enclosing lambda
return y;
}
};
If x was not const, the second use would require 'L' to capture, and
that would be an error.
Make sure TranformLambdaExpr is also aware of this.
Patch approved by Richard (Thanks!!)
http://llvm-reviews.chandlerc.com/D2092
------------------------------------------------------------------------
Fix for PR18052 - Lambdas within NSDMI's and default arguments in Nested classes.
Clang currently croaks on the following:
struct X1 {
struct X2 {
int L = ([] (int i) { return i; })(2);
};
};
asserting that the containing lexical context of the lambda is not Sema's cur context, when pushing the lambda's decl context on.
This occurs because (prior to this patch) getContainingDC always returns the non-nested class for functions at class scope (even for inline member functions of nested classes (to account for delayed parsing of their bodies)). The patch addresses this by having getContainingDC always return the lexical DC for a lambda's call operator.
Link to the bug: http://llvm.org/bugs/show_bug.cgi?id=18052
Link to Richard Smith's feedback on phabricator: http://llvm-reviews.chandlerc.com/D2331
[SystemZ] Fix handling of pass-by-pointer arguments
I'd misunderstood getIndirect() to mean that the argument should be passed
as a pointer at the ABI level, with the ByVal argument choosing caller-copy
semantics over no-caller-copy (callee-copy-on-write) semantics. But
getIndirect(x) actually means that x is passed by pointer at the IR
level but (at least on all other targets I looked at) directly at the
ABI level. getIndirect(x, false) selects a pointer to a caller-made
copy, which is what SystemZ was aiming for.
This fixes a miscompilation of c-index-test. Structure arguments were being
passed by pointer, but no copy was being made, so a write in the callee
stomped over a caller's local variable.
[AArch64] Implemented vcopy_lane patterns using scalar DUP instruction.
Patch by Ana Pazos!
------------------------------------------------------------------------
[Mips] Take in account the -mfp64 command line option when build paths
to the crt*.o files, libraries and headers for the MIPS FSFS toolchain.
------------------------------------------------------------------------
projects / clang / log