Anton Afanasyev [Mon, 19 Aug 2019 22:58:26 +0000 (22:58 +0000)]
[Support][Time profiler] Make FE codegen blocks to be inside frontend blocks
Summary:
Add `Frontend` time trace entry to `HandleTranslationUnit()` function.
Add test to check all codegen blocks are inside frontend blocks.
Also, change `--time-trace-granularity` option a bit to make sure very small
time blocks are outputed to json-file when using `--time-trace-granularity=0`.
Aaron Ballman [Mon, 19 Aug 2019 13:37:41 +0000 (13:37 +0000)]
Don't keep stale pointers to LoopInfos.
CGLoopInfo was keeping pointers to parent loop LoopInfos, but when the loop info vector grew, it reallocated the storage and invalidated all of the parent pointers, causing use-after-free. Manage the lifetimes of the LoopInfos separately so that the pointers aren't stale.
[OpenCL] Add generic type handling for builtin functions
Generic types are an abstraction of type sets. It mimics the way
functions are defined in the OpenCL specification. For example,
floatN can abstract all the vector sizes of the float type.
This allows to
* stick more closely to the specification, which uses generic types;
* factorize definitions of functions with numerous prototypes in the
tablegen file; and
* reduce the memory impact of functions with many overloads.
[OpenCL] Fix addr space deduction for pointers/references to arrays.
Rewrite the logic for detecting if we are deducing addr space of
a pointee type to take into account special logic for arrays. For
pointers/references to arrays we can have any number of parentheses
expressions as well as nested pointers.
Troy A. Johnson [Sat, 17 Aug 2019 04:20:24 +0000 (04:20 +0000)]
[X86] Support -mlong-double-80
Add an option group for all of the -mlong-double-* options and make
-mlong-double-80 restore the default long double behavior for X86. The
motivations are that GNU accepts the -mlong-double-80 option and that complex
Makefiles often need a way of undoing earlier options. Prior to this commit, if
one chooses 64-bit or 128-bit long double for X86, there is no way to undo that
choice and restore the 80-bit behavior.
Troy A. Johnson [Fri, 16 Aug 2019 21:00:22 +0000 (21:00 +0000)]
[X86] Support -mlong-double-80
Add an option group for all of the -mlong-double-* options and make
-mlong-double-80 restore the default long double behavior for X86. The
motivations are that GNU accepts the -mlong-double-80 option and that complex
Makefiles often need a way of undoing earlier options. Prior to this commit, if
one chooses 64-bit or 128-bit long double for X86, there is no way to undo that
choice and restore the 80-bit behavior.
Alexey Bataev [Fri, 16 Aug 2019 20:15:02 +0000 (20:15 +0000)]
[OPENMP5.0]Diagnose global variables in lambda not marked as declare
target.
According to OpenMP 5.0, if a lambda declaration and definition appears between a declare target directive and the matching end declare target directive, all variables that are captured by the lambda expression must also appear in a to clause.
Balazs Keri [Fri, 16 Aug 2019 12:10:03 +0000 (12:10 +0000)]
[ASTImporter] Import ctor initializers after setting flags.
Summary:
Code to import "ctor initializers" at import of functions
is moved to be after the flags in the newly created function
are imported. This fixes an error when the already created but
incomplete (flags are not set) function declaration is accessed.
Csaba Dabis [Fri, 16 Aug 2019 01:53:14 +0000 (01:53 +0000)]
[analyzer] Analysis: Silence checkers
Summary:
This patch introduces a new `analyzer-config` configuration:
`-analyzer-config silence-checkers`
which could be used to silence the given checkers.
It accepts a semicolon separated list, packed into quotation marks, e.g:
`-analyzer-config silence-checkers="core.DivideZero;core.NullDereference"`
It could be used to "disable" core checkers, so they model the analysis as
before, just if some of them are too noisy it prevents to emit reports.
This patch also adds support for that new option to the scan-build.
Passing the option `-disable-checker core.DivideZero` to the scan-build
will be transferred to `-analyzer-config silence-checkers=core.DivideZero`.
Joel E. Denny [Thu, 15 Aug 2019 21:17:48 +0000 (21:17 +0000)]
[Rewrite][NFC] Add FIXMEs and tests for RemoveLineIfEmpty bug
I'd like to add these comments to warn others of problems I
encountered when trying to use `RemoveLineIfEmpty`. I originally
tried to fix the problem, but I realized I could implement the
functionality more easily and efficiently in my calling code where I
can make the simplifying assumption that there are no prior edits to
the line from which text is being removed. While I've lost the
motivation to write a fix, which doesn't look easy, I figure a warning
to others is better than silence.
I've added a unit test to demonstrate the problem. I don't know how
to mark it as an expected failure, so I just marked it disabled.
Reid Kleckner [Thu, 15 Aug 2019 19:45:28 +0000 (19:45 +0000)]
[Sema] Implement DR2386 for C++17 structured binding
Allow implementations to provide complete definitions of
std::tuple_size<T>, but to omit the 'value' member to signal that T is
not tuple-like. The Microsoft standard library implements
std::tuple_size<const T> this way.
If the value member exists, clang still validates that it is an ICE, but
if it does not, then the type is considered to not be tuple-like.
Guanzhong Chen [Thu, 15 Aug 2019 19:33:36 +0000 (19:33 +0000)]
[WebAssembly] Correctly handle va_arg of zero-sized structures
Summary:
D66168 passes size 0 structs indirectly, while the wasm backend expects it to
be passed directly. This causes subsequent variadic arguments to be read
incorrectly.
This diff changes it so that size 0 structs are passed directly.
Aaron Ballman [Thu, 15 Aug 2019 18:35:44 +0000 (18:35 +0000)]
Allow standards-based attributes to have leading and trailing underscores.
This gives library implementers a way to use standards-based attributes that do not conflict with user-defined macros of the same name. Attributes in C2x require this behavior normatively (C2x 6.7.11p4), but there's no reason to not have the same behavior in C++, especially given that such attributes may be used by a C library consumed by a C++ compilation.
Alexey Bataev [Thu, 15 Aug 2019 17:15:35 +0000 (17:15 +0000)]
[Driver][Bundler] Improve bundling of object files.
Summary:
Previously, object files were bundled using partial linking. It resulted
in the following structure of the bundled objects:
```
<host_code>
clang-offload-bundle
__CLANG_OFFLOAD_BUNDLE__<target>
<target_code>
```
But when we tried to unbundle object files, it worked correctly only for
the target objects. The host object remains bundled. It produced a lot of
junk sections in the host object files and in some cases may caused
incorrect linking.
Patch improves bundling of the object files. After this patch the
bundled object looks like this:
With this structure we are able to unbundle the host object files too so
that after unbundling they are the same as were before.
The host section is bundled twice. The bundled section is used to
unbundle the original host section.
Kristof Umann [Thu, 15 Aug 2019 08:53:16 +0000 (08:53 +0000)]
[analyzer] Warn about -analyzer-configs being meant for development purposes only
This is more of a temporary fix, long term, we should convert AnalyzerOptions.def
into the universally beloved (*coughs*) TableGen format, where they can more
easily be separated into developer-only, alpha, and user-facing configs.
New pragma "vectorize_predicate(enable)" now implies "vectorize(enable)",
and it is ignored when vectorization is disabled with e.g.
"vectorize(disable) vectorize_predicate(enable)".
David L. Jones [Thu, 15 Aug 2019 04:10:11 +0000 (04:10 +0000)]
[Tooling] Add a hack to work around issues with matcher binding in r368681.
The change in r368681 contains a (probably unintentional) behavioral change for
rewrite rules with a single matcher. Previously, the single matcher would not
need to be bound (`joinCaseMatchers` returned it directly), even though a final
DynTypeMatcher was created and bound by `buildMatcher`. With the new change, a
single matcher will be bound, in addition to the final binding (which is now in
`buildMatchers`, but happens roughly at the same point in the overall flow).
This patch simply duplicates the "final matcher" trick: it creates an extra
DynTypedMatcher for each rewrite rule case matcher, and unconditionally makes it
bindable. This is probably not the right long-term fix, but it does allow
existing code to continue to work with this interface.
Alex Langford [Thu, 15 Aug 2019 00:58:51 +0000 (00:58 +0000)]
[NFCI] Always initialize BugReport const fields
Summary:
Some compilers require that const fields of an object must be explicitly
initialized by the constructor. I ran into this issue building with clang
3.8 on Ubuntu 16.04.
[Clang] Migrate llvm::make_unique to std::make_unique
Now that we've moved to C++14, we no longer need the llvm::make_unique
implementation from STLExtras.h. This patch is a mechanical replacement
of (hopefully) all the llvm::make_unique instances across the monorepo.
Richard Smith [Wed, 14 Aug 2019 22:57:50 +0000 (22:57 +0000)]
Fix handling of class member access into a vector type.
When handling a member access into a non-class, non-ObjC-object type, we
would perform a lookup into the surrounding scope as if for an
unqualified lookup. If the member access was followed by a '<' and this
lookup (or the typo-correction for it) found a template name, we'd treat
the member access as naming that template.
Now we treat such accesses as never naming a template if the type of the
object expression is of vector type, so that vector component accesses
are never misinterpreted as naming something else. This is not entirely
correct, since it is in fact valid to name a template from the enclosing
scope in this context, when invoking a pseudo-destructor for the vector
type via an alias template, but that's very much a corner case, and this
change leaves that case only as broken as the corresponding case for
Objective-C types is.
This incidentally adds support for dr2292, which permits a 'template'
keyword at the start of a member access naming a pseudo-destructor.
Alexey Bataev [Wed, 14 Aug 2019 19:30:06 +0000 (19:30 +0000)]
[OPENMP]Support for non-rectangular loops.
Added basic support for non-rectangular loops. It requires an additional
analysis of min/max boundaries for non-rectangular loops. Since only
linear dependency is allowed, we can do this analysis.
Kristof Umann [Wed, 14 Aug 2019 17:05:55 +0000 (17:05 +0000)]
[CFG] Introduce CFGElementRef, a wrapper that knows it's position in a CFGBlock
Previously, collecting CFGElements in a set was practially impossible, because
both CFGBlock::operator[] and both the iterators returned it by value. One
workaround would be to collect the iterators instead, but they don't really
capture the concept of an element, and elements from different iterator types are incomparable.
This patch introduces CFGElementRef, a wrapper around a (CFGBlock, Index) pair,
and a variety of new iterators and iterator ranges to solve this problem.
I guess you could say that this patch took a couple iterations to get right :^)
This patch simply moves code that already exists into a new function.
Specifically I think it will make the BuildActions code for building a clang
job pipeline easier to read and work with.
Erik Pilkington [Wed, 14 Aug 2019 16:57:11 +0000 (16:57 +0000)]
[Sema][ObjC] Fix a -Wformat false positive with localizedStringForKey
Only honour format_arg attributes on -[NSBundle localizedStringForKey] when its
argument has a format specifier in it, otherwise its likely to just be a key to
fetch localized strings.
Kristof Umann [Wed, 14 Aug 2019 12:20:08 +0000 (12:20 +0000)]
[analyzer][CFG] Don't track the condition of asserts
Well, what is says on the tin I guess!
Some more changes:
* Move isInevitablySinking() from BugReporter.cpp to CFGBlock's interface
* Rename and move findBlockForNode() from BugReporter.cpp to
ExplodedNode::getCFGBlock()
Balazs Keri [Wed, 14 Aug 2019 09:41:39 +0000 (09:41 +0000)]
[ASTImporter] Import default expression of param before creating the param.
Summary:
The default expression of a parameter variable should be imported before
the parameter variable object is created. Otherwise the function is created
with an incomplete parameter variable (default argument is nullptr) and in
this intermediary state the expression is imported. This import can have
a reference to the incomplete parameter variable that causes crash.
Taewook Oh [Wed, 14 Aug 2019 07:11:09 +0000 (07:11 +0000)]
[NewPM][PassInstrumentation] IR printing support from clang driver
Summary: https://reviews.llvm.org/D50923 enabled the IR printing support for the new pass manager, but only for the case when `opt` tool is used as a driver. This patch is to enable the IR printing when `clang` is used as a driver.
Richard Smith [Wed, 14 Aug 2019 02:30:11 +0000 (02:30 +0000)]
Add __has_builtin support for builtin function-like type traits.
Summary:
Previously __has_builtin(__builtin_*) would return false for
__builtin_*s that we modeled as keywords rather than as functions
(because they take type arguments). With this patch, all builtins
that are called with function-call-like syntax return true from
__has_builtin (covering __builtin_* and also the __is_* and __has_* type
traits and the handful of similar builtins without such a prefix).
Update the documentation on __has_builtin and on type traits to match.
While doing this I noticed the type trait documentation was out of date
and incomplete; that's fixed here too.
Summary:
Explicitly deleting the copy constructor makes compiling the function
`ento::registerGenericTaintChecker` difficult with some compilers. When we
construct an `llvm::Optional<TaintConfig>`, the optional is constructed with a
const TaintConfig reference which it then uses to invoke the deleted TaintConfig
copy constructor.
I've observered this failing with clang 3.8 on Ubuntu 16.04.
Kristof Umann [Wed, 14 Aug 2019 00:48:57 +0000 (00:48 +0000)]
[analyzer][NFC] Prepare visitors for different tracking kinds
When we're tracking a variable that is responsible for a null pointer
dereference or some other sinister programming error, we of course would like to
gather as much information why we think that the variable has that specific
value as possible. However, the newly introduced condition tracking shows that
tracking all values this thoroughly could easily cause an intolerable growth in
the bug report's length.
There are a variety of heuristics we discussed on the mailing list[1] to combat
this, all of them requiring to differentiate in between tracking a "regular
value" and a "condition".
This patch introduces the new `bugreporter::TrackingKind` enum, adds it to
several visitors as a non-optional argument, and moves some functions around to
make the code a little more coherent.
Kristof Umann [Tue, 13 Aug 2019 23:48:10 +0000 (23:48 +0000)]
[analyzer] Track the right hand side of the last store regardless of its value
Summary:
The following code snippet taken from D64271#1572188 has an issue: namely,
because `flag`'s value isn't undef or a concrete int, it isn't being tracked.
int flag;
bool coin();
void foo() {
flag = coin();
}
void test() {
int *x = 0;
int local_flag;
flag = 1;
foo();
local_flag = flag;
if (local_flag)
x = new int;
foo();
local_flag = flag;
if (local_flag)
*x = 5;
}
This, in my opinion, makes no sense, other values may be interesting too.
Originally added by rC185608.
Kristof Umann [Tue, 13 Aug 2019 23:22:33 +0000 (23:22 +0000)]
[analyzer] Prune calls to functions with linear CFGs that return a non-zero constrained value
During the evaluation of D62883, I noticed a bunch of totally
meaningless notes with the pattern of "Calling 'A'" -> "Returning value"
-> "Returning from 'A'", which added no value to the report at all.
This patch (not only affecting tracked conditions mind you) prunes
diagnostic messages to functions that return a value not constrained to
be 0, and are also linear.
When -trim-egraph is unavailable (say, when you're debugging a crash on
a real-world code that takes too long to reduce), it makes sense to view
the untrimmed graph up to the crashing node's predecessor, then dump the ID
(or a pointer) of the node in the attached debugger, and then trim
the dumped graph in order to keep only paths from the root to the node.
Artem Dergachev [Tue, 13 Aug 2019 23:04:50 +0000 (23:04 +0000)]
[analyzer] exploded-graph-rewriter: NFC: Refactor explorers into trimmers.
Explorers aren't the right abstraction. For the purposes of displaying svg files
we don't care in which order do we explore the nodes. We may care about this for
other analyses, but we're not there yet.
The function of cutting out chunks of the graph is performed poorly by
the explorers, because querying predecessors/successors on the explored nodes
yields original successors/predecessors even if they aren't being explored.
Introduce a new entity, "trimmers", that do one thing but to it right: cut out
chunks of the graph. Trimmers mutate the graph, so stale edges aren't even
visible to their consumers in the pipeline. Additionally, trimmers are
intrinsically composable: multiple trimmers can be applied to the graph
sequentially.
Refactor the single-path explorer into the single-path trimmer.
Rename the test file for consistency.
Artem Dergachev [Tue, 13 Aug 2019 23:04:47 +0000 (23:04 +0000)]
[analyzer] exploded-graph-rewriter: Open the converted graph immediately.
Change the default behavior: the tool no longer dumps the rewritten .dot file
to stdout, but instead it automatically converts it into an .html file
(which essentially wraps an .svg file) and immediately opens it with
the default web browser.
This means that the tool should now be fairly easy to use:
$ exploded-graph-rewriter.py /tmp/ExprEngine.dot
The benefits of wrapping the .svg file into an .html file are:
- It'll open in a web browser, which is the intended behavior.
An .svg file would be open with an image viewer/editor instead.
- It avoids the white background around the otherwise dark svg area
in dark mode.
The feature can be turned off by passing a flag '--rewrite-only'.
The LIT substitution is updated to enforce the old mode because
we don't want web browsers opening on our buildbots.
Kristof Umann [Tue, 13 Aug 2019 22:03:08 +0000 (22:03 +0000)]
[analyzer][NFC] Make sure that the BugReport is not modified during the construction of non-visitor pieces
I feel this is kinda important, because in a followup patch I'm adding different
kinds of interestingness, and propagating the correct kind in BugReporter.cpp is
just one less thing to worry about.
Kristof Umann [Tue, 13 Aug 2019 21:48:17 +0000 (21:48 +0000)]
[analyzer][NFC] Refactoring BugReporter.cpp P6.: Completely get rid of interestingness propagation
Apparently this does literally nothing.
When you think about this, it makes sense. If something is really important,
we're tracking it anyways, and that system is sophisticated enough to mark
actually interesting statements as such. I wouldn't say that it's even likely
that subexpressions are also interesting (array[10 - x + x]), so I guess even
if this produced any effects, its probably undesirable.
Guanzhong Chen [Tue, 13 Aug 2019 21:41:11 +0000 (21:41 +0000)]
[WebAssembly] Make clang emit correct va_arg code for structs
Summary:
In the WebAssembly backend, when lowering variadic function calls, non-single
member aggregate type arguments are always passed by pointer.
However, when emitting va_arg code in clang, the arguments are instead read as
if they are passed directly. This results in the pointer being read as the
actual structure.
Alexey Bataev [Tue, 13 Aug 2019 19:32:36 +0000 (19:32 +0000)]
Don't use std::errc
Summary:
As noted on Errc.h:
// * std::errc is just marked with is_error_condition_enum. This means that
// common patters like AnErrorCode == errc::no_such_file_or_directory take
// 4 virtual calls instead of two comparisons.
And on some libstdc++ those virtual functions conclude that
projects / clang / log