move-constructors and move-assignment operators, use memcpy to copy adjacent
POD members.
Previously, classes with one or more Non-POD members would fall back on
element-wise copies for all members, including POD members. This often
generated a lot of IR. Without padding metadata, it wasn't often possible
for the LLVM optimizers to turn the element-wise copies into a memcpy.
This code hasn't yet received any serious tuning. I didn't see any serious
regressions on a self-hosted clang build, or any of the nightly tests, but
I think it's important to get this out in the wild to get more testing.
Insights, feedback and comments welcome.
Many thanks to David Blaikie, Richard Smith, and especially John McCall for
their help and feedback on this work.
llvm-svn: 174919
This is to ensure that GlobalOpt in LLVM does not attempt to look through a
selector reference to a method var name at compile time.
I also added a test/updated old tests that need to recognize the new keyword.
rdar://12580965.
llvm-svn: 174461
This can yield dramatic speedups of dynamic_cast for simple inheritance trees,
at least with libsupc++. Neither libcxxabi nor libcxxrt make use of this
hint currently, it was never implemented because clang didn't support it.
There was some concern about the number of class hierarchy walks this change
introduces. If it turns out to be an issue we can add caching either at the cast
pair level or even deeper, but we also do a lot of walks in Sema so this
codepath is probably fairly optimized already.
llvm-svn: 174293
r173593 made us a little too eager to associate all code at the end of a
function with the user-written 'return' line. This caused problems with
breakpoints as they'd be set in exception handling code preceeding the
actual non-exception return handling code, leading to the breakpoint never
being hit in non-exceptional execution.
This change restores the pre-r173593 exception handling line information where
the cleanup code is associated with the '}' not the return line.
llvm-svn: 174206
In cooperation with the LLVM patch, this should implement all scalar front-end
parts of the C and C++ ABIs for AArch64.
This patch excludes the NEON support also reviewed due to an outbreak of
batshit insanity in our legal department. That will be committed soon bringing
the changes to precisely what has been approved.
Further reviews would be gratefully received.
llvm-svn: 174055
MSan instrumentation is driven by the original code. We take every
incoming instruction and emit another instruction (or ten) next to
it, operating on the shadow values (but sometimes on the real values,
too). Two programs in one, essentially. There can be any kinds of
redundancies in the second one, so we just run whatever is normally
run at -O2, and then exclude some passes that do not help much with
benchmarks.
llvm-svn: 174049
Several places were still treating the Attribute object as respresenting
multiple attributes. Those places now use the AttributeSet to represent
multiple attributes.
llvm-svn: 174004
implementation; this is much more inline with the original implementation
(i.e., pre-ubsan) and does not require run-time library support.
The trapping implementation can be invoked using either '-fcatch-undefined-behavior'
or '-fsanitize=undefined-trap -fsanitize-undefined-trap-on-error', with the latter
being preferred. Eventually, the -fcatch-undefined-behavior' flag will be removed.
llvm-svn: 173848
In the future, AttributeWithIndex won't be used anymore. Besides, it exposes the
internals of the AttributeSet to outside users, which isn't goodness.
llvm-svn: 173605
One of the gotchas (see changes to CodeGenFunction) was due to the fix in
r139416 (for PR10829). This only worked previously because the top level
lexical block would set the location to the end of the function, the debug
location would be updated (as per r139416), the location would be set to
the end of the function again (but that would no-op, since it was the same
as the previous location), then the return instruction would be emitted using
the debug location.
Once the top level lexical block was no longer emitted, the end-of-function
location change was causing the debug loc to be updated, regressing that bug.
llvm-svn: 173593
Title: [PR9027] volatile struct bug: member is not loaded at -O;
This is caused by last flag passed to @llvm.memcpy being false,
not honoring that aggregate has at least one 'volatile' data member
(even though aggregate itself has not been qualified as 'volatile'.
As a result, optimization optimizes away the memcpy altogether.
Patch review by John MaCall (I still need to fix up a test though).
llvm-svn: 173535
never key functions. We did not implement that rule for the
iOS ABI, which was driven by what was implemented in gcc-4.2.
However, implement it now for other ARM-based platforms.
llvm-svn: 173515
Collections of attributes are handled via the AttributeSet class now. This
finally frees us up to make significant changes to how attributes are structured.
llvm-svn: 173229
We were previously hard-coding a particular field index. This was
fine before (because we were obviously guaranteed the presence
of a copy/dispose member) except for (1) alignment padding and
(2) future extensions adding extra members to the header, such
as the extended-layout pointer.
Note that we only introduce the extended-layout pointer in the
presence of structs. (We also seem to be introducing it even
in the presence of an all-non-object layout, but that's a
different potential issue.)
llvm-svn: 173122
This is more code to isolate the use of the Attribute class to that of just
holding one attribute instead of a collection of attributes.
llvm-svn: 173095
Looks like r161368 fixed this for one case but not all. This change generalizes
the solution over all the unwrapping cases. Now that preserving the qualifiers
is done independent of the particular type being unwrapped I won't bother
adding test cases for each one but at least demonstrate that this change was
necessary & sufficient to fix the bug.
llvm-svn: 173002
Adding the pseudo first parameter to a member function pointer's function type
and mark it as artificial.
Combined with a fix to GDB ( http://sourceware.org/bugzilla/show_bug.cgi?id=14998 )
this fixes gdb.cp/member-ptr.exp with Clang.
llvm-svn: 172911
Manually fix the order of UnwrappedLineParser.cpp as that one didn't
have its associated header as the first header.
This also uncovered a subtle inclusion order dependency as CLog.h didn't
include LLVM.h to pick up using declarations it relied upon.
llvm-svn: 172892
it apart from [[gnu::noreturn]] / __attribute__((noreturn)), since their
semantics are not equivalent (for instance, we treat [[gnu::noreturn]] as
affecting the function type, whereas [[noreturn]] does not).
llvm-svn: 172691
!0 = metadata !{metadata !"-lautolink"}
!1 = metadata !{metadata !"-framework", metadata !"autolink_framework"}
referenced from llvm.module.linkoptions, e.g.,
!llvm.module.linkoptions = !{!0, !1, !2, !3}
This conceptually moves the logic for figuring out the syntax the
linker will accept from LLVM into Clang. Moreover, it makes it easier
to support MSVC's
#pragma comment(linker, "some option")
in the future, should anyone care to do so.
llvm-svn: 172441
the 64-bit PowerPC ELF ABI.
The ABI requires that the real and imaginary parts of a complex argument
each occupy their own doubleword. Arguments smaller than 8 bytes are
right-adjusted within the doubleword.
Clang expects EmitVAARG() to return a pointer to a structure in which
the real and imaginary parts are packed adjacently in memory. To accomplish
this, we generate code to load the code appropriately from the varargs
location and pack the values into a temporary variable in the form Clang
expects, returning a pointer to that structure.
The test case demonstrates correct code generation for all "small" complex
types on PPC64: int, short, char, and float.
llvm-svn: 172438
metadata for linking against the libraries/frameworks for imported
modules.
The module map language is extended with a new "link" directive that
specifies what library or framework to link against when a module is
imported, e.g.,
link "clangAST"
or
link framework "MyFramework"
Importing the corresponding module (or any of its submodules) will
eventually link against the named library/framework.
For now, I've added some placeholder global metadata that encodes the
imported libraries/frameworks, so that we can test that this
information gets through to the IR. The format of the data is still
under discussion.
llvm-svn: 172437
In ArrayRef<T>(X), X should not be temporary value. It could be rewritten more redundantly;
llvm::Type *XTy = X->getType();
ArrayRef<llvm::Type *> Ty(XTy);
llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, Ty);
Since it is safe if both XTy and Ty are temporary value in one statement, it could be shorten;
llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, ArrayRef<llvm::Type*>(X->getType()));
ArrayRef<T> has an implicit constructor to create uni-entry of T;
llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType());
MSVC-generated clang.exe crashed.
llvm-svn: 172352
storage and thus is implicitly zero-initialized, no need to
do C++11 memory model. This patch unconditionally detects
such condition and zeroinitializer's the variable.
Patch has been commented on and OKed by Doug off-line.
// rdar://12897704
llvm-svn: 172144
difference between type widths of a vector and the width of one of its elements
in the case of vector shifts. Use correct witdth in the vector case.
llvm-svn: 172047
This is in preparation for using this to construct the function type for
pointers to member functions to include the implicit/artificial 'this'
parameter in that case as well. (feedback from GDB indicates that this might be
all that's necessary to get it to behave well with Clang's pointer-to-member
function debug output)
llvm-svn: 171809
with respect to the lower "left-hand-side bitwidth" bits, even when negative);
see OpenCL spec 6.3j. This patch both implements this behaviour in the code
generator and "constant folding" bits of Sema, and also prevents tests
to detect undefinedness in terms of the weaker C99 or C++ specifications
from being applied.
llvm-svn: 171755
Using added LLVM functionality in r171698. This works in GDB for member
variable pointers but not member function pointers. See the LLVM commit and
GDB bug 14998 for details.
Un-xfailing cases in the GDB 7.5 test suite will follow.
llvm-svn: 171699
Catch some cases I'd missed in r171605 related to unnamed parameters of record
type. This resolves all remaining cases of PR14573 suppression in the GDB 7.5
test suite. Fix to the test suite to follow.
llvm-svn: 171633
__strong __block variables, perform objc_storeStrong on
source and destination instead of direct move. This
is done with -O0 and to improve some analysis.
// rdar://12530881
llvm-svn: 171555
The line information was changed when emitting debug information for all the
DeclRefExprs and we should change it back to get ready for PopClenupBlocks
called from FinishFunction.
rdar://11562117
llvm-svn: 171493
directly.
This is in preparation for removing the use of the 'Attribute' class as a
collection of attributes. That will shift to the AttributeSet class instead.
llvm-svn: 171254
/// \param array - a value of type elementType*
/// \param destructionKind - the kind of destruction required
/// \param initializedElementCount - a value of type size_t* holding the number of successfully-constructed elements
llvm-svn: 171013
When we are visiting the extern declaration of 'i' in
static int i = 99;
int foo() {
extern int i;
return i;
}
We should not try to handle it as if it was an function static. That is, we
must consider the written storage class.
Fixing this then exposes that the assert in EmitGlobalVarDeclLValue and the
if leading to its call are not completely accurate. They were passing before
because the second decl was marked as having external storage. I changed them
to check the linkage, which I find easier to understand.
Last but not least, there is something strange going on with cuda and opencl.
My guess is that the linkage computation for these languages needs to be
audited, but I didn't want to change that in this patch so I just updated
the storage classes to keep the current behavior.
Thanks to Reed Kotler for reporting this.
llvm-svn: 170827
This does limit these typedefs to being sequences, but no current usage
requires them to be contiguous (we could expand this to a more general
iterator pair range concept at some point).
Also, it'd be nice if SmallVector were constructible directly from an ArrayRef
but this is a bit tricky since ArrayRef depends on SmallVectorBaseImpl for the
inverse conversion. (& generalizing over all range-like things, while nice,
would require some nontrivial SFINAE I haven't thought about yet)
llvm-svn: 170482
PR 14529 was opened because neither Clang or LLVM was expanding
calls to creal* or cimag* into instructions that just load the
respective complex field. After some discussion, it was not
considered realistic to do this in LLVM because of the platform
specific way complex types are expanded. Thus a way to solve
this in Clang was pursued. GCC does a similar expansion.
This patch adds the feature to Clang by making the creal* and
cimag* functions library builtins and modifying the builtin code
generator to look for the new builtin types.
llvm-svn: 170455
incompatibility with how complex values are returned. It is sufficient
to flag all complex types as direct rather than indirect.
A new test case is provided that checks correct IR generation for the
various supported flavors of _Complex.
llvm-svn: 170302
I wasn't sure where to put the test case for this, but this seemed like as good
a place as any. I had to reorder the tests here to make them legible while
still matching the order of metadata output in the IR file (for some reason
making it virtual changed the ordering).
Relevant commit to fix up LLVM to actually respect 'artificial' member
variables is coming once I write up a test case for it.
llvm-svn: 170154
My variadics patch, r169588, changed these calls to typically be
bitcasts rather than calls to a supposedly variadic function.
This totally subverted a hack where we intentionally dropped
excess arguments from such calls in order to appease the inliner
and a "warning" from the optimizer. This patch extends the hack
to also work with bitcasts, as well as teaching it to rewrite
invokes.
llvm-svn: 170034
call sites as tail calls unconditionally. While it's theoretically true that
this is just an optimization, it's an optimization that we very much want to
happen even at -O0, or else ARC applications become substantially harder to
debug. See r169796 for the llvm/fast-isel side of things.
rdar://12553082
llvm-svn: 169996
Note that there is no test suite update. This was found by a couple of
tests failing when the test suite was run on a powerpc64 host (thanks
Roman!). The tests don't specify a triple, which might seem surprising
for a codegen test. But in fact, these tests don't even inspect their
output. Not at all. I could add a bunch of triples to these tests so
that we'd get the test coverage for normal builds, but really someone
needs to go through and add actual *tests* to these tests. =[ The ones
in question are:
test/CodeGen/bitfield-init.c
test/CodeGen/union.c
llvm-svn: 169694
This was an egregious bug due to the several iterations of refactorings
that took place. Size no longer meant what it original did by the time
I finished, but this line of code never got updated. Unfortunately we
had essentially zero tests for this in the regression test suite. =[
I've added a PPC64 run over the bitfield test case I've been primarily
using. I'm still looking at adding more tests and making sure this is
the *correct* bitfield access code on PPC64 linux, but it looks pretty
close to me, and it is *worlds* better than before this patch as it no
longer asserts! =] More commits to follow with at least additional tests
and maybe more fixes.
Sorry for the long breakage due to this....
llvm-svn: 169691
We were emitting calls to blocks as if all arguments were
required --- i.e. with signature (A,B,C,D,...) rather than
(A,B,...). This patch fixes that and accounts for the
implicit block-context argument as a required argument.
In addition, this patch changes the function type under which
we call unprototyped functions on platforms like x86-64 that
guarantee compatibility of variadic functions with unprototyped
function types; previously we would always call such functions
under the LLVM type T (...)*, but now we will call them under
the type T (A,B,C,D,...)*. This last change should have no
material effect except for making the type conventions more
explicit; it was a side-effect of the most convenient implementation.
llvm-svn: 169588
generally support the C++11 memory model requirements for bitfield
accesses by relying more heavily on LLVM's memory model.
The primary change this introduces is to move from a manually aligned
and strided access pattern across the bits of the bitfield to a much
simpler lump access of all bits in the bitfield followed by math to
extract the bits relevant for the particular field.
This simplifies the code significantly, but relies on LLVM to
intelligently lowering these integers.
I have tested LLVM's lowering both synthetically and in benchmarks. The
lowering appears to be functional, and there are no really significant
performance regressions. Different code patterns accessing bitfields
will vary in how this impacts them. The only real regressions I'm seeing
are a few patterns where the LLVM code generation for loads that feed
directly into a mask operation don't take advantage of the x86 ability
to do a smaller load and a cheap zero-extension. This doesn't regress
any benchmark in the nightly test suite on my box past the noise
threshold, but my box is quite noisy. I'll be watching the LNT numbers,
and will look into further improvements to the LLVM lowering as needed.
llvm-svn: 169489
The count attribute is more accurate with regards to the size of an array. It
also obviates the upper bound attribute in the subrange. We can also better
handle an unbound array by setting the count to -1 instead of the lower bound to
1 and upper bound to 0.
llvm-svn: 169311
uncovered.
This required manually correcting all of the incorrect main-module
headers I could find, and running the new llvm/utils/sort_includes.py
script over the files.
I also manually added quite a few missing headers that were uncovered by
shuffling the order or moving headers up to be main-module-headers.
llvm-svn: 169237
The count field is necessary because there isn't a difference between the 'lo'
and 'hi' attributes for a one-element array and a zero-element array. When the
count is '0', we know that this is a zero-element array. When it's >=1, then
it's a normal constant sized array. When it's -1, then the array is unbounded.
llvm-svn: 169219
1) init-order sanitizer: initialization-order checker.
Status: usable, but may produce false positives w/o proper blacklisting.
2) use-after-return sanitizer
Status: implemented, but heavily understed.
Should be optional, as it significanlty slows program down.
3) use-after-scope sanitizer
Status: in progress.
llvm-svn: 168950
Among other differences, GCC accepts
typedef int IA[];
typedef int A10[10];
static A10 *f(void);
static IA *f(void);
void g(void) {
(void)sizeof(*f());
}
but clang used to reject it with:
invalid application of 'sizeof' to an incomplete type 'IA' (aka 'int []')
The intention of c99's 6.2.7 seems to be that we should use the composite type
and accept as gcc does.
Doing the type merging required some extra fixes:
* Use the type from the function type in initializations, even if an parameter
is available.
* Fix the merging of the noreturn attribute in function types.
* Make CodeGen handle the fact that an parameter type can be different from
the corresponding type in the function type.
llvm-svn: 168895
in deciding a copy/dispose field is needed in a byref structure
and when generating the copy/dispose helpers. In certain
cases, these fields were being added but no copy/dispose was
being generated. This was uncovered in ARC, but not in MRR.
// rdar://12759433
llvm-svn: 168825
the original parameter or return type.
Since we do not accurately represent the data fields of a union, we should not
directly load or store a union type.
As an exmple, if we have i8,i8, i32, i32 as one field type and i32,i32 as
another field type, the first field type will be chosen to represent the union.
If we load with the union's type, the 3rd byte and the 4th byte will be skipped.
rdar://12723368
llvm-svn: 168820
objc_loadWeak. This retains and autorelease the weakly-refereced
object. This hidden autorelease sometimes makes __weak variable alive even
after the weak reference is erased, because the object is still referenced
by an autorelease pool. This patch overcomes this behavior by loading a
weak object via call to objc_loadWeakRetained(), followng it by objc_release
at appropriate place, thereby removing the hidden autorelease. // rdar://10849570
llvm-svn: 168740
ELF ABI.
Complex values are to be passed in registers as though the real and
imaginary parts were passed as separate parameters. Prior to this
patch, complex values were passed as byval aggregates. It turns out
that specifying getDirect() for all complex types when classifying the
argument type results in the desired behavior.
The new Clang test case verifies that the correct LLVM IR is generated
for caller and callee for each of the underlying types for _Complex.
llvm-svn: 168673
Separate out the notions of 'has a trivial special member' and 'has a
non-trivial special member', and use them appropriately. These are not
opposites of one another (there might be no special member, or in C++11 there
might be a trivial one and a non-trivial one). The CXXRecordDecl predicates
continue to produce incorrect results, but do so in fewer cases now, and
they document the cases where they might be wrong.
No functionality changes are intended here (they will come when the predicates
start producing the right answers...).
llvm-svn: 168119
more sense anyway - it determines how expressions are codegen'd. It also ensures
that -ffp-contract=fast has the intended effect when compiling LLVM IR.
llvm-svn: 168027
type as written from the ParmVarDecl; it's unclear whether the standard
(C99 6.9.1p10) requires this, but we're following the precedent set by gcc,
and hopefully nobody will ever ask about this again.
PR9559 / <rdar://problem/12621983>.
llvm-svn: 167985
If we have a type 'int a[1]' and a type 'int b[0]', the generated DWARF is the
same for both of them because we use the 'upper_bound' attribute. Instead use
the 'count' attrbute, which gives the correct number of elements in the array.
<rdar://problem/12566646>
llvm-svn: 167807
at whether the *selected* constructor would be trivial rather than considering
whether the array's element type has *any* non-trivial constructors of the
relevant kind.
llvm-svn: 167562
This is useful because unnamed bitfields can have effects on the
offsets which are not otherwise reflected in the DWARF information.
<rdar://problem/12629719>
llvm-svn: 167503
disabling byval, we set realign to true.
It will perform an aligned alloca, and call memcpy to copy the byval
argument to the local variable.
Change the size threshold back to 64 bytes.
rdar://12596507
llvm-svn: 167440
token. This is important because the first token could actually be
after an #include that triggers a module import, which might use
either Sema or the AST consumer before it would have been initialized.
llvm-svn: 167423
the type alignment of the byval argument. This patch will disable byval in this case,
it also increases the size threshold for turning on byval.
A backend fix will be attempted.
rdar://12596507
llvm-svn: 167416
checks to enable. Remove frontend support for -fcatch-undefined-behavior,
-faddress-sanitizer and -fthread-sanitizer now that they don't do anything.
llvm-svn: 167413
John McCall