ARM ABI specifies that all the libcalls use soft FP ABI
(even hard FP binaries). These days clang emits _mulsc3 / _muldc3
calls with default (C) calling convention which would be translated
into AAPCS_VFP LLVM calling and thus the result of complex
multiplication will be bogus.
Introduce a way for a target to specify explicitly calling
convention for libcalls. Right now this is temporary correctness
fix. Ultimately, we'll end with intrinsic for complex
multiplication and all calling convention decisions for libcalls
will be put into backend.
llvm-svn: 223123
Now that LLVM can count the registers needed to implement AAPCS rules, we don't
need to duplicate that logic here. This means we can drop the explicit padding
and also use more natural types in many cases (e.g. "struct { float arr[3]; }"
used to end up as "[2 x double]" to avoid holes on the stack.
The one wrinkle is that AAPCS va_arg was also using the register counting
machinery. But the local replacement isn't too bad.
llvm-svn: 222904
Cygwin and MinGW fail to conform to the underlying system's structure passing
ABI. Make the check more precise to ensure that we correctly generate code for
the itanium environment.
llvm-svn: 222626
"global-init", "global-init-src" and "global-init-type" were originally
used to blacklist entities in ASan init-order checker. However, they
were never documented, and later were replaced by "=init" category.
Old blacklist entries should be converted as follows:
* global-init:foo -> global:foo=init
* global-init-src:bar -> src:bar=init
* global-init-type:baz -> type:baz=init
llvm-svn: 222401
This reverts commit r222144. Commit r222142 is being reverted due to
a spec2006/gcc execution-time regression.
Update mips-varargs test as well.
llvm-svn: 222397
Summary:
With this patch, passing a va_list to another function and reading 10 int's from
it works correctly on a big-endian target.
Based on a pair of patches by David Chisnall, one of which I've reworked
for the current trunk.
Reviewers: theraven, atanasyan
Reviewed By: theraven, atanasyan
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D6248
llvm-svn: 222339
Summary:
This distinguishes between -fpic and -fPIC now, with the additions in LLVM for
PIC level support.
Test Plan: No regressions
Reviewers: echristo, rafael
Reviewed By: rafael
Subscribers: rnk, emaste, llvm-commits
Differential Revision: http://reviews.llvm.org/D5400
llvm-svn: 222227
used inside blocks. It fixes a crash in naming code
for __func__ etc. when used in a block declared globally.
It also brings back old naming convention for
predefined expression which was broken. rdar://18961148
llvm-svn: 222065
Summary:
Ok, here is somewhat addition to D6217 aiming to preserve old darwin behavior wrt the typedefed types. The actual change to SemaChecking turned out to be pretty gross, in particular:
1. We need to extract the typedef'ed type for proper diagnostics
2. We need to walk over paren expressions as well
Reviewers: chandlerc, rsmith
Reviewed By: rsmith
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D6256
llvm-svn: 222044
VSX makes the "vector long long" and "vector double" types available.
This patch enables the vec_perm interface for these types. The same
builtin is generated regardless of the specified type, so no
additional work or testing is needed in the back end. Tests are added
to ensure this builtin is generated by the front end.
llvm-svn: 221988
This patch adds builtin support for xvdivdp and xvdivsp, along with a
new test case. The builtins are accessed using vec_div in altivec.h.
Builtins are listed (mostly) alphabetically there, so inserting these
changed the line numbers for deprecation warnings tested in
test/Headers/altivec-intrin.c.
There is a companion patch for LLVM.
llvm-svn: 221984
Summary:
Consider the following nifty 1 liner: (0 ? csqrtl(2.0f) : sqrtl(2.0f)). One can easily obtain such code from e.g. tgmath. Right now it produces an assertion because we fail to do the promotion real => _Complex real.
The case was properly handled previously (old handleOtherComplexFloatConversion routine), but was forgotten in the current version. This seems to be about fallout from r219557
Reviewers: chandlerc, rsmith
Reviewed By: rsmith
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D6217
llvm-svn: 221821
This patch enables the vec_vsx_ld and vec_vsx_st intrinsics for
PowerPC, which provide programmer access to the lxvd2x, lxvw4x,
stxvd2x, and stxvw4x instructions.
New code in altivec.h defines these in terms of new builtins, which
are themselves defined in BuiltinsPPC.def. The builtins are converted
to LLVM intrinsics in CGBuiltin.cpp. Additional code is added to
builtins-ppc-vsx.c to verify the correct generation of the intrinsics.
Note that I moved the other VSX builtins so all VSX builtins will be
alphabetical in their own section in BuiltinsPPC.def.
There is a companion patch for LLVM.
llvm-svn: 221768
Summary: If we've added poisoned paddings to a type do not emit memcpy for operator=.
Test Plan: regression tests.
Reviewers: majnemer, rsmith
Reviewed By: rsmith
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D6160
llvm-svn: 221739
Summary:
This change makes CodeGenFunction::EmitCheck() take several
conditions that needs to be checked (all of them need to be true),
together with sanitizer kinds these checks are for. This would allow
to split one call into UBSan runtime into several calls in case
different sanitizer kinds would have different recoverability
settings.
Tests should be fixed accordingly, I'm working on it.
Test Plan: regression test suite.
Reviewers: rsmith
Reviewed By: rsmith
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D6219
llvm-svn: 221716
Homogeneous aggregates on AAPCS_VFP ARM need to be passed *without* being
flattened (e.g. [2 x float] rather than "float, float") for various weird ABI
reasons. However, this isn't the case for anything else; further, we know at
the ABIArgInfo::getDirect callsites whether this flattening is allowed.
So, we can get more unified ARM code, with a simpler Clang, by just using that
knowledge directly.
llvm-svn: 221559
mingw64's headers implement fabs by calling __builtin_fabs, so using the
library call results in an infinite loop. If the backend legalizes
@llvm.fabs as a call to fabs later, things should work out, as the crt
provides a definition.
llvm-svn: 221206
It turns out that MinGW never dllimports of exports inline functions.
This means that code compiled with Clang would fail to link with
MinGW-compiled libraries since we might try to import functions that
are not imported.
To fix this, make Clang never dllimport inline functions when targeting
MinGW.
llvm-svn: 221154
The most complex aspect of the convention is the handling of homogeneous
vector and floating point aggregates. Reuse the homogeneous aggregate
classification code that we use on PPC64 and ARM for this.
This convention also has a C mangling, and we apparently implement that
in both Clang and LLVM.
Reviewed By: majnemer
Differential Revision: http://reviews.llvm.org/D6063
llvm-svn: 221006
Now that we have initial support for VSX, we can begin adding
intrinsics for programmer access to VSX instructions. This patch
performs the necessary enablement in the front end, and tests it by
implementing intrinsics for minimum and maximum using the vector
double data type.
The main change in the front end is to no longer disallow "vector" and
"double" in the same declaration (lib/Sema/DeclSpec.cpp), but "vector"
and "long double" must still be disallowed. The new intrinsics are
accessed via vec_max and vec_min with changes in
lib/Headers/altivec.h. Note that for v4f32, we already access
corresponding VMX builtins, but with VSX enabled we should use the
forms that allow all 64 vector registers.
The new built-ins are defined in include/clang/Basic/BuiltinsPPC.def.
I've added a new test in test/CodeGen/builtins-ppc-vsx.c that is
similar to, but much smaller than, builtins-ppc-altivec.c. This
allows us to test VSX IR generation without duplicating CHECK lines
for the existing bazillion Altivec tests.
Since vector double is now legal when VSX is available, I've modified
the error message, and changed where we test for it and for vector
long double, since the target machine isn't visible in the old place.
This serendipitously removed a not-pertinent warning about 'long'
being deprecated when used with 'vector', when "vector long double" is
encountered and we just want to issue an error. The existing tests
test/Parser/altivec.c and test/Parser/cxx-altivec.cpp have been
updated accordingly, and I've added test/Parser/vsx.c to verify that
"vector double" is now legitimate with VSX enabled.
There is a companion patch for LLVM.
llvm-svn: 220989
Summary:
When we are adding field paddings for asan even an empty dtor has to remain in the code,
so we ignore -mconstructor-aliases if the paddings are going to be added.
Test Plan: added a test
Reviewers: rsmith, rnk, rafael
Reviewed By: rafael
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D6038
llvm-svn: 220986
Reuse the PPC64 HVA detection algorithm for ARM and AArch64. This is a
nice code deduplication, since they are roughly identical. A few virtual
method extension points are needed to understand how big an HVA can be
and what element types it can have for a given architecture.
Also make the record expansion code work in the presence of non-virtual
bases.
Reviewed By: uweigand, asl
Differential Revision: http://reviews.llvm.org/D6045
llvm-svn: 220972
The Windows NT SDK uses __readfsdword and declares it as a compiler provided
builtin (#pragma intrinsic(__readfsword). Because intrin.h is not referenced
by winnt.h, it is not possible to provide an out-of-line definition for the
intrinsic. Provide a proper compiler builtin definition.
llvm-svn: 220859
Following the NVVM IR specifications, arguments of aggregate type should be
passed on the stack without splitting (byval).
http://reviews.llvm.org/D6020
Patch by Jacques Pienaar.
llvm-svn: 220854
As discussed in bug 21398, PowerPC ABI code needs to consider C++ base
classes when classifying a class as homogeneous aggregate (or not) for
ABI purposes.
llvm-svn: 220852
An updated implemnentation of VLA types capturing based on previously committed solution for Lambdas.
This version captures the whole VLA type instead of particular variables which are part of VLA size expression and allows to use previusly calculated size of VLA type in captured regions. Required for OpenMP.
Differential Revision: http://reviews.llvm.org/D5099
llvm-svn: 220850
Summary:
We should avoid a tail padding not only if the last field
has zero size but also if the last field is a struct with a flexible array.
If/when http://reviews.llvm.org/D5478 is committed,
this will also handle the case of structs with zero-sized arrays.
Reviewers: majnemer, rsmith
Reviewed By: rsmith
Subscribers: cfe-commits
Differential Revision: http://reviews.llvm.org/D5924
llvm-svn: 220708
Wire it through everywhere we have support for fastcall, essentially.
This allows us to parse the MSVC "14" CTP headers, but we will
miscompile them because LLVM doesn't support __vectorcall yet.
Reviewed By: Aaron Ballman
Differential Revision: http://reviews.llvm.org/D5808
llvm-svn: 220573
Summary:
This allows us to easily identify them in the backend which in turn allows us
to handle them correctly for big-endian targets (where they must be shifted
into the upper bits of the register).
Depends on D5961
Reviewers: atanasyan
Reviewed By: atanasyan
Subscribers: cfe-commits, theraven
Differential Revision: http://reviews.llvm.org/D5962
llvm-svn: 220566
Summary:
Ensure all integral/enumeration types are appropriately annotated with
signext/zeroext. In particular, i32 now has these attributes when using the
N32/N64 ABI. This paves the way for accurately representing the way the
N32/N64 ABI's promotes integer arguments to i64.
Reviewers: atanasyan
Reviewed By: atanasyan
Subscribers: cfe-commits, theraven
Differential Revision: http://reviews.llvm.org/D5961
llvm-svn: 220563
When SanitizerBlacklist decides if the SourceLocation is blacklisted,
we need to first turn it into a SpellingLoc before fetching the filename
and scanning "src:" entries. Otherwise we will fail to fecth the
correct filename for function definitions coming from macro expansion.
llvm-svn: 220403
This reverts commit r220169 which reverted r220153. However, it also
contains additional changes:
- We may need to add padding *after* we've packed the struct. This
occurs when the aligned next field offset is greater than the new
field's offset. When this occurs, we make the struct packed.
*However*, once packed the next field offset might be less than the
new feild's offset. It is in this case that we might further pad the
struct.
- We would pad structs which were perfectly sized! This behavior is
immensely old. This behavior came from blindly subtracting
NextFieldOffsetInChars from RecordSize. This doesn't take into
account the fact that the struct might have a greater overall
alignment than the last field.
llvm-svn: 220175
This commit caused two tests in LNT to regress. I'm able to reproduce on
any platform and will send reproduction steps to the original commit
log. This should restore the LNT bots that have been failing.
llvm-svn: 220169
a NaN-test prior to the call to the library function.
This should automatically make fastmath (including just non-NaNs) able to avoid
the expensive libcalls and also open the door to more advanced folding in LLVM
based on the rules for complex math.
Two important notes to remember: first is that this isn't yet a proper
limited range mode, it's still just improving the unlimited range mode.
Also, it isn't really perfecet w.r.t. what an unlimited range mode
should be doing because it isn't quite handling the flags produced by
all the operations in the way desirable for that mode, but then neither
is compiler-rt's libcall. When the compiler-rt libcall is improved to
carefully manage flags, the code emitted here should be improved
correspondingly. And it is still a long-term desirable thing to add
a limited range mode to Clang that would be able to use direct math
without library calls here.
Special thanks to Steve Canon for the careful review on this patch and
teaching me about these issues. =D
Differential Revision: http://reviews.llvm.org/D5756
llvm-svn: 220167
Before, ConstStructBuilder::AppendBytes would check packed constraints
prior to padding being added before the field's offset. However, adding
this padding might force our struct to be packed. Because we wouldn't
check *after* adding padding, ConstStructBuilder would be in an
inconsistent state leading to a crash.
This fixes PR21300.
llvm-svn: 220153
This commit changes the way we blacklist global variables in ASan.
Now the global is excluded from instrumentation (either regular
bounds checking, or initialization-order checking) if:
1) Global is explicitly blacklisted by its mangled name.
This part is left unchanged.
2) SourceLocation of a global is in blacklisted source file.
This changes the old behavior, where instead of looking at the
SourceLocation of a variable we simply considered llvm::Module
identifier. This was wrong, as identifier may not correspond to
the file name, and we incorrectly disabled instrumentation
for globals coming from #include'd files.
3) Global is blacklisted by type.
Now we build the type of a global variable using Clang machinery
(QualType::getAsString()), instead of llvm::StructType::getName().
After this commit, the active users of ASan blacklist files
may have to revisit them (this is a backwards-incompatible change).
llvm-svn: 220097
This commit changes the way we blacklist functions in ASan, TSan,
MSan and UBSan. We used to treat function as "blacklisted"
and turned off instrumentation in it in two cases:
1) Function is explicitly blacklisted by its mangled name.
This part is not changed.
2) Function is located in llvm::Module, whose identifier is
contained in the list of blacklisted sources. This is completely
wrong, as llvm::Module may not correspond to the actual source
file function is defined in. Also, function can be defined in
a header, in which case user had to blacklist the .cpp file
this header was #include'd into, not the header itself.
Such functions could cause other problems - for instance, if the
header was included in multiple source files, compiled
separately and linked into a single executable, we could end up
with both instrumented and non-instrumented version of the same
function participating in the same link.
After this change we will make blacklisting decision based on
the SourceLocation of a function definition. If a function is
not explicitly defined in the source file, (for example, the
function is compiler-generated and responsible for
initialization/destruction of a global variable), then it will
be blacklisted if the corresponding global variable is defined
in blacklisted source file, and will be instrumented otherwise.
After this commit, the active users of blacklist files may have
to revisit them. This is a backwards-incompatible change, but
I don't think it's possible or makes sense to support the
old incorrect behavior.
I plan to make similar change for blacklisting GlobalVariables
(which is ASan-specific).
llvm-svn: 219997
Summary:
The general approach is to add extra paddings after every field
in AST/RecordLayoutBuilder.cpp, then add code to CTORs/DTORs that poisons the paddings
(CodeGen/CGClass.cpp).
Everything is done under the flag -fsanitize-address-field-padding.
The blacklist file (-fsanitize-blacklist) allows to avoid the transformation
for given classes or source files.
See also https://code.google.com/p/address-sanitizer/wiki/IntraObjectOverflow
Test Plan: run SPEC2006 and some of the Chromium tests with -fsanitize-address-field-padding
Reviewers: samsonov, rnk, rsmith
Reviewed By: rsmith
Subscribers: majnemer, cfe-commits
Differential Revision: http://reviews.llvm.org/D5687
llvm-svn: 219961
They cannot be written to, so marking them const makes sense and may improve
optimisation.
As a side-effect, SectionInfos has to be moved from Sema to ASTContext.
It also fixes this problem, that occurs when compiling ATL:
warning LNK4254: section 'ATL' (C0000040) merged into '.rdata' (40000040) with different attributes
The ATL headers are putting variables in a special section that's marked
read-only. However, Clang currently can't model that read-onlyness in the IR.
But, by making the variables const, the section does become read-only, and
the linker warning is avoided.
Differential Revision: http://reviews.llvm.org/D5812
llvm-svn: 219960
CodeGen wouldn't mark the aliasee as thread_local if the aliasee was a
tentative definition.
Even if the definition was already emitted, it would never mark the
alias as thread_local.
This fixes PR21288.
llvm-svn: 219859
Thumb1 has legitimate reasons for preferring 32-bit alignment of types
i1/i8/i16, since the 16-bit encoding of "add rD, sp, #imm" requires #imm to be
a multiple of 4. However, this is a trade-off betweem code size and RAM usage;
the DataLayout string is not the best place to represent it even if desired.
So this patch removes the extra Thumb requirements, hopefully making ARM and
Thumb completely compatible in this respect.
llvm-svn: 219735
Before, ARM and Thumb mode code had different preferred alignments, which could
lead to some rather unexpected results. There's justification for reducing it
from the default 64-bits (wasted space), but I don't think there is for going
below 32-bits.
There's no actual ABI change here, just to reassure people.
llvm-svn: 219720
This addresses a regression introduced with SVN r219393. A block may be
contained within another block. In such a scenario, we would end up within a
BlockDecl, which is not a NamedDecl (as the names are synthesised). The cast to
a NamedDecl of the DeclContext would then assert as the types are unrelated.
Restore the mangling behaviour to that prior to SVN r219393. If the current
block is contained within a BlockDecl, walk up to the parent DeclContext,
recursively, until we have a non-BlockDecl. This is expected to be a NamedDecl.
Add in a couple of asserts to ensure that the assumption that we only encounter
a block within a NamedDecl or a BlockDecl.
llvm-svn: 219696
Previously loop hints such as #pragma loop vectorize_width(#) required a constant. This patch allows a constant expression to be used as well. Such as a non-type template parameter or an expression (2 * c + 1).
Reviewed by Richard Smith
llvm-svn: 219589
and !=) to support mixed complex and real operand types.
This requires removing an assert from SemaChecking, and adding support
both to the constant evaluator and the code generator to synthesize the
imaginary part when needed. This seemed somewhat cleaner than having
just the comparison operators force real-to-complex conversions.
I've added test cases for these operations. I'm really terrified that
there were *no* tests in-tree which exercised this.
This turned up when trying to build R after my change to the complex
type lowering.
llvm-svn: 219570
for complex math.
This should fix the windows build bots that started having trouble here
and generally fix complex libcall emission on targets which use sret for
complex data types. It also makes the code a bit simpler (despite
calling into a much more complex bucket of code).
llvm-svn: 219565
operators where one type is a C complex type, and to emit both the
efficient and correct implementation for complex arithmetic according to
C11 Annex G using this extra information.
For both multiply and divide the old code was writing a long-hand
reduced version of the math without any of the special handling of inf
and NaN recommended by the standard here. Instead of putting more
complexity here, this change does what GCC does which is to emit
a libcall for the fully general case.
However, the old code also failed to do the proper minimization of the
set of operations when there was a mixed complex and real operation. In
those cases, C provides a spec for much more minimal operations that are
valid. Clang now emits the exact suggested operations. This change isn't
*just* about performance though, without minimizing these operations, we
again lose the correct handling of infinities and NaNs. It is critical
that this happen in the frontend based on assymetric type operands to
complex math operations.
The performance implications of this change aren't trivial either. I've
run a set of benchmarks in Eigen, an open source mathematics library
that makes heavy use of complex. While a few have slowed down due to the
libcall being introduce, most sped up and some by a huge amount: up to
100% and 140%.
In order to make all of this work, also match the algorithm in the
constant evaluator to the one in the runtime library. Currently it is
a broken port of the simplifications from C's Annex G to the long-hand
formulation of the algorithm.
Splitting this patch up is very hard because none of this works without
the AST change to preserve non-complex operands. Sorry for the enormous
change.
Follow-up changes will include support for sinking the libcalls onto
cold paths in common cases and fastmath improvements to allow more
aggressive backend folding.
Differential Revision: http://reviews.llvm.org/D5698
llvm-svn: 219557
Make it possible to pass NULL through variadic functions on 64-bit
Windows targets. The Visual C++ headers define NULL to 0, when they
should define it to 0LL on Win64 so that NULL is a pointer-sized
integer.
Fixes PR20949.
Reviewers: thakis, rsmith
Differential Revision: http://reviews.llvm.org/D5480
llvm-svn: 219456
We already add the align parameter attribute for function parameters that have
the align_value attribute (or those with a typedef type having that attribute),
which is an important special case, but does not handle pointers with value
alignment assumptions that come into scope in any other way. To handle the
general case, emit an @llvm.assume-based alignment assumption whenever we load
the pointer-typed lvalue of an align_value-attributed variable (except for
function parameters, which we already deal with at entry).
I'll also note that this is more general than Intel's described support in:
https://software.intel.com/en-us/articles/data-alignment-to-assist-vectorization
which states that the compiler inserts __assume_aligned directives in response
to align_value-attributed variables only for function parameters and for the
initializers of local variables. I think that we can make the optimizer deal
with this more-general scheme (which could lead to a lot of calls to
@llvm.assume inside of loop bodies, for example), but if not, I'll rework this
to be less aggressive.
llvm-svn: 219052
This reverts commit r218917, effectively reapplying r218913. Original
commit message follows.
--
Update debug info testcases for an LLVM metadata schema change to fold
metadata constant operands into a single `MDString`.
Part of PR17891.
llvm-svn: 219011
This test includes stdint.h (via stdatomic.h), which might include system
headers (and that might not work, depending on the system configuration).
Attempting to fix llvm-clang-lld-x86_64-debian-fast.
llvm-svn: 218960
Adds a Clang-specific implementation of C11's stdatomic.h header. On systems,
such as FreeBSD, where a stdatomic.h header is already provided, we defer to
that header instead (using our __has_include_next technology). Otherwise, we
provide an implementation in terms of our __c11_atomic_* intrinsics (that were
created for this purpose).
C11 7.1.4p1 requires function declarations for atomic_thread_fence,
atomic_signal_fence, atomic_flag_test_and_set,
atomic_flag_test_and_set_explicit, and atomic_flag_clear, and requires that
they have external linkage. Accordingly, we provide these declarations, but if
a user elides the shadowing macros and uses them, then they must have a libc
(or similar) that actually provides definitions.
atomic_flag is implemented using _Bool as the underlying type. This is
consistent with the implementation provided by FreeBSD and also GCC 4.9 (at
least when __GCC_ATOMIC_TEST_AND_SET_TRUEVAL == 1).
Patch by Richard Smith (rebased and slightly edited by me -- Richard said I
should drive at this point).
llvm-svn: 218957
Update debug info testcases for an LLVM metadata schema change to fold
metadata constant operands into a single `MDString`.
Part of PR17891.
llvm-svn: 218913
This adds support for the align_value attribute. This attribute is supported by
Intel's compiler (versions 14.0+), and several of my HPC users have requested
support in Clang. It specifies an alignment assumption on the values to which a
pointer points, and is used by numerical libraries to encourage efficient
generation of vector code.
Of course, we already have an aligned attribute that can specify enhanced
alignment for a type, so why is this additional attribute important? The
problem is that if you want to specify that an input array of T is, say,
64-byte aligned, you could try this:
typedef double aligned_double attribute((aligned(64)));
void foo(aligned_double *P) {
double x = P[0]; // This is fine.
double y = P[1]; // What alignment did those doubles have again?
}
the access here to P[1] causes problems. P was specified as a pointer to type
aligned_double, and any object of type aligned_double must be 64-byte aligned.
But if P[0] is 64-byte aligned, then P[1] cannot be, and this access causes
undefined behavior. Getting round this problem requires a lot of awkward
casting and hand-unrolling of loops, all of which is bad.
With the align_value attribute, we can accomplish what we'd like in a well
defined way:
typedef double *aligned_double_ptr attribute((align_value(64)));
void foo(aligned_double_ptr P) {
double x = P[0]; // This is fine.
double y = P[1]; // This is fine too.
}
This attribute does not create a new type (and so it not part of the type
system), and so will only "propagate" through templates, auto, etc. by
optimizer deduction after inlining. This seems consistent with Intel's
implementation (thanks to Alexey for confirming the various Intel-compiler
behaviors).
As a final note, I would have chosen to call this aligned_value, not
align_value, for better naming consistency with the aligned attribute, but I
think it would be more useful to users to adopt Intel's name.
llvm-svn: 218910
Anton Korobeynikov