Chris Bieneman [Fri, 15 Jan 2016 21:21:12 +0000 (21:21 +0000)]
[CMake] Support generation of linker order files using dtrace
Summary:
This patch extends the lit-based perf-training tooling supplied for PGO data generation to also generate linker order files using dtrace.
This patch should work on any system that has dtrace. If CMake can find the dtrace tool it will generate a target 'generate-order-file' which will run the per-training tests wrapped by dtrace to capture function entries. There are several algorithms implemented for sorting the order files which can be experimented with for best performance. The dtrace wrapper also supports bot oneshot and pid probes.
The perf-helper.py changes to support order file construction are ported from internal changes by ddunbar; he gets all the credit for the hard work here, I just copy and pasted.
Note: I've tested these patches on FreeBSD and OS X 10.10.
Artem Dergachev [Fri, 15 Jan 2016 15:22:05 +0000 (15:22 +0000)]
[analyzer] Provide .def-files and visitors for SVal/SymExpr/MemRegion, v2.
Provide separate visitor templates for the three hierarchies, and also
the `FullSValVisitor' class, which is a union of all three visitors.
Additionally, add a particular example visitor, `SValExplainer', in order to
test the visitor templates. This visitor is capable of explaining the SVal,
SymExpr, or MemRegion in a natural language.
Compared to the reverted r257605, this fixes the test that used to fail
on some triples, and fixes build failure under -fmodules.
Richard Trieu [Fri, 15 Jan 2016 05:48:38 +0000 (05:48 +0000)]
Fixing more issues with template type diffing
1) Print qualifiers for templates with zero arguments
2) Add a few more tests for the template type diffing refactoring.
Specifically, PR24587 has been fixed and has a test case from
http://reviews.llvm.org/D15384
3) Adds asserts to check the DiffTree is in correct state when moving nodes
4) Rename the field FromType and ToType since it is heavily used within
member functions.
Richard Trieu [Fri, 15 Jan 2016 05:01:53 +0000 (05:01 +0000)]
Add new diff modes to template type diffing.
Remove an old assertion that does not hold. It is possible for a template
argument to be a declaration in one instantiation and an integer in another.
Create two new diff kinds for these (decl vs int and int vs decl).
Alexey Bataev [Fri, 15 Jan 2016 04:36:32 +0000 (04:36 +0000)]
PR26111: segmentation fault with __attribute__((mode(QI))) on function declaration, by Denis Zobnin
Allow "mode" attribute to be applied to VarDecl, not ValueDecl (which includes FunctionDecl and EnumConstantDecl), emit an error if this attribute is used with function declarations and enum constants.
Differential Revision: http://reviews.llvm.org/D16112
Alexey Bataev [Fri, 15 Jan 2016 04:06:31 +0000 (04:06 +0000)]
[X86] Support 'interrupt' attribute for x86
This attribute may be attached to a function definition and instructs the backend to generate appropriate function entry/exit code so that
it can be used directly as an interrupt handler.
The IRET instruction, instead of the RET instruction, is used to return from interrupt or exception handlers. All registers, except for the EFLAGS register which is restored by the IRET instruction, are preserved by the compiler.
Any interruptible-without-stack-switch code must be compiled with -mno-red-zone since interrupt handlers can and will, because of the hardware design, touch
the red zone.
interrupt handler must be declared with a mandatory pointer argument:
struct interrupt_frame;
__attribute__ ((interrupt))
void f (struct interrupt_frame *frame) {
...
}
and user must properly define the structure the pointer pointing to.
exception handler:
The exception handler is very similar to the interrupt handler with a different mandatory function signature:
#ifdef __x86_64__
typedef unsigned long long int uword_t;
#else
typedef unsigned int uword_t;
#endif
struct interrupt_frame;
__attribute__ ((interrupt))
void f (struct interrupt_frame *frame, uword_t error_code) {
...
}
and compiler pops the error code off stack before the IRET instruction.
The exception handler should only be used for exceptions which push an error code and all other exceptions must use the interrupt handler.
The system will crash if the wrong handler is used.
Differential Revision: http://reviews.llvm.org/D15709
Akira Hatanaka [Fri, 15 Jan 2016 03:34:06 +0000 (03:34 +0000)]
[CodeGen] Attach attributes to thread local wrapper function.
This commit is a follow-up to r251734, r251476, and r249735, which fixes
a bug where function attributes were not attached to thread local
wrapper functions.
Richard Smith [Fri, 15 Jan 2016 03:24:18 +0000 (03:24 +0000)]
When copying whitespace flags from the token naming a macro argument onto the
first token of the expansion, don't forget to copy the "is at the start of a
line" token (which is always false, as newlines cannot appear within a macro
body); otherwise, stringizing the result can insert spurious whitespace.
Richard Trieu [Fri, 15 Jan 2016 02:55:17 +0000 (02:55 +0000)]
Save the integer type for integral template arguments.
Save the integer type when diffing integers in template type diffing. When
integers are different sizes, print out the type along with the integer value.
Also with the type information, print true and false instead of 1 and 0 for
boolean values.
Richard Trieu [Fri, 15 Jan 2016 01:08:56 +0000 (01:08 +0000)]
Make template type diffing use the new desguared iterator.
If available, use the canonical template argument to fill in information for
template type diffing instead of attempting to special case and evaluate Expr's
for the value. Since those are the values used in template instantiation,
we don't have to worry about difference between our evaluator and theirs. Also
move the nullptr template arguments from DiffKind::Expression to
DiffKind::Declaration and allow DiffKind::Declaration to set an Expr. The only
effect that should result is that a named nullptr will show up as
'ptr aka nullptr' in diagnostics.
Justin Lebar [Thu, 14 Jan 2016 23:31:30 +0000 (23:31 +0000)]
[CUDA] Warn undeclared identifiers in CUDA kernel calls
Value, type, and instantiation dependence were not being handled
correctly for CUDAKernelCallExpr AST nodes. As a result, if an
undeclared identifier was used in the triple-angle-bracket kernel call
configuration, there would be no error during parsing, and there would
be a crash during code gen. This patch makes sure that an error will be
issued during parsing in this case, just as there would be for any other
use of an undeclared identifier in C++.
Richard Trieu [Thu, 14 Jan 2016 23:30:12 +0000 (23:30 +0000)]
Change the TSTiterator in Template Type Diffing.
Modify the TSTiterator to have two internal iterators, which will walk
the provided sugared type and the desugared type. This will provide better
access to the template argument information. No functional changes.
Richard Trieu [Thu, 14 Jan 2016 22:56:39 +0000 (22:56 +0000)]
Refactor template type diffing
1) Instead of using pairs of From/To* fields, combine fields into a struct
TemplateArgInfo and have two in each DiffNode.
2) Use default initialization in DiffNode so that the constructor shows the
only field that is initialized differently on construction.
3) Use Set and Get functions per each DiffKind to make sure all fields for the
diff is set. In one case, the Expr fields were not set.
4) Don't print boolean literals for boolean template arguments. This prevents
printing 'false aka 0'
Only #3 has a functional change, which is reflected in the test change.
Justin Lebar [Thu, 14 Jan 2016 21:41:27 +0000 (21:41 +0000)]
[CUDA] Invoke ptxas and fatbinary during compilation.
Summary:
Previously we compiled CUDA device code to PTX assembly and embedded
that asm as text in our host binary. Now we compile to PTX assembly and
then invoke ptxas to assemble the PTX into a cubin file. We gather the
ptx and cubin files for each of our --cuda-gpu-archs and combine them
using fatbinary, and then embed that into the host binary.
Adds two new command-line flags, -Xcuda_ptxas and -Xcuda_fatbinary,
which pass args down to the external tools.
Andrey Bokhanko [Thu, 14 Jan 2016 10:41:16 +0000 (10:41 +0000)]
PR25910: clang allows two var definitions with the same mangled name
Proper diagnostic and resolution of mangled names' conflicts in variables.
When there is a declaration and a definition using the same name but different
types, we emit what is in the definition. When there are two conflicting
definitions, we issue an error.
Daniel Jasper [Wed, 13 Jan 2016 16:41:34 +0000 (16:41 +0000)]
clang-format: [ObjC+JS] Allow bin-packing of array literals.
After reading the style guides again, they don't actually say how to
pack or not pack array literals. Based on some user reports, array
initializers can unnecessarily get quite long if they contain many
small elements. Array literals with trailing commas are still formatted
one per line so that users have a way to opt out of the packing.
Artem Dergachev [Wed, 13 Jan 2016 15:13:48 +0000 (15:13 +0000)]
[analyzer] Provide .def-files and visitors for SVal/SymExpr/MemRegion.
Provide separate visitor templates for the three hierarchies, and also
the `FullSValVisitor' class, which is a union of all three visitors.
Additionally, add a particular example visitor, `SValExplainer', in order to
test the visitor templates. This visitor is capable of explaining the SVal,
SymExpr, or MemRegion in a natural language.
2. MemSpaceRegion used to be inconsistently used as both an abstract base and
a particular region. This region class is now an abstract base and no longer
occupies GenericMemSpaceRegionKind. Instead, a new class, CodeSpaceRegion,
is introduced for handling the unique use case for MemSpaceRegion as
"the generic memory space" (when it represents a memory space that holds all
executable code).
3. BEG_ prefixes in memory region kind ranges are renamed to BEGIN_ for
consisitency with symbol kind ranges.
4. FunctionTextRegion and BlockTextRegion are renamed to FunctionCodeRegion and
BlockCodeRegion, respectively. The term 'code' is less jargony than 'text' and
we already refer to BlockTextRegion as a 'code region' in BlockDataRegion.
Alexey Bataev [Wed, 13 Jan 2016 11:18:54 +0000 (11:18 +0000)]
[OPENMP] Fix for declarative/standalone directives use.
Fixes processing of declarative directives and standalone executable directives. Declarative directives should not be allowed as an immediate statements and standalone executable directives are allowed to be used in case-stmt constructs.
In {CG,}ExprConstant.cpp, we weren't treating vector splats properly.
This patch makes us treat splats more properly.
Additionally, this patch adds a new cast kind which allows a bool->int
cast to result in -1 or 0, instead of 1 or 0 (for true and false,
respectively), so we can sanely model OpenCL bool->int casts in the AST.
Justin Lebar [Wed, 13 Jan 2016 01:24:35 +0000 (01:24 +0000)]
[CUDA] Rename check-prefixes in cuda-options.cu and cuda-unused-arg-warning.cu.
Summary:
Rename the args to be more human-readable. Among other things, this
lets us get rid of a bunch of comments (e.g. "ensure we don't run the
linker"), greatly shortening these tests.
Also apply consistent formatting and fix some English nits while we're
at it.
Bob Wilson [Wed, 13 Jan 2016 01:19:02 +0000 (01:19 +0000)]
Generalize r256026 to apply to all MachO targets, not just Darwin targets.
The PIC default is set for the MachO toolchain, not just the Darwin toolchain,
so this treats those the same. The behavior with -static should be the same
for all MachO targets. rdar://24152327
Frederic Riss [Tue, 12 Jan 2016 23:47:59 +0000 (23:47 +0000)]
[Darwin] Fix deployment target detection
There was a thinko in the deployment target detection code that
made the -isysroot parsing have precedence over the environment
variable for tvOS. This patch makes this logic symetric for all
platforms (the env variable must have precedence).
Richard Smith [Tue, 12 Jan 2016 21:59:26 +0000 (21:59 +0000)]
Improve AST dumping:
1) When dumping a declaration that declares a name for a type, also dump the named type.
2) Add a #pragma clang __debug dump X, that dumps the lookup results for X in
the current context.
Adrian Prantl [Tue, 12 Jan 2016 21:01:56 +0000 (21:01 +0000)]
Module debugging: Make the module format part of the module hash instead
of the file name. This is consistent with how other HeaderSearchOptions
are handled.
Due to the other inputs of the module hash (revision number) this is not
really testable in a meaningful way.
Ismail Donmez [Tue, 12 Jan 2016 10:41:20 +0000 (10:41 +0000)]
Revert r253898 and r253899, this breaks mingw compilation on openSUSE.
Will re-apply after llvm 3.8 is branched.
Original commit message:
Driver: fallback to the location of clang if no sysroot,
hard coding /usr makes little sense for mingw-w64.
If we have portable toolchains having /usr breaks that.
If the clang we use is in /usr/bin or /usr/sbin etc this will
still detect as though it was hard coded to /usr
This makes the most sense going forward for mingw-w64 toolchains
on both linux and mac
Pierre Gousseau [Tue, 12 Jan 2016 10:07:56 +0000 (10:07 +0000)]
[analyzer] Evaluate integral casts as cast symbols if truncations are detected.
The current workaround for truncations not being modelled is that the evaluation of integer to integer casts are simply bypassed and so the original symbol is used as the new casted symbol (cf SimpleSValBuilder::evalCastFromNonLoc).
This lead to the issue described in PR25078, as the RangeConstraintManager associates ranges with symbols.
The new evalIntegralCast method added by this patch wont bypass the cast if it finds the range of the symbol to be greater than the maximum value of the target type.
The fix to RangeSet::pin mentioned in the initial review will be committed separately.
Alexey Bataev [Tue, 12 Jan 2016 09:12:20 +0000 (09:12 +0000)]
PR18513: make gcc compatible layout for bit-fields with explicit aligned attribute, by Dmitry Polukhin
Fix binary compatibility issue with GCC.
Differential Revision: http://reviews.llvm.org/D14980
Alexey Bataev [Tue, 12 Jan 2016 09:01:25 +0000 (09:01 +0000)]
Fix infinite recursion for invalid declaration, by Dmitry Polukhin
Fix for a case found by fuzzing PR23057 (comment #25 https://llvm.org/bugs/show_bug.cgi?id=23057#c25).
Differential Revision: http://reviews.llvm.org/D16065
Richard Smith [Mon, 11 Jan 2016 22:41:53 +0000 (22:41 +0000)]
When a tag is declared in prototype scope in C, if we've decided that it
redeclares an existing tag but are creating a new declaration anyway (because
it has attributes or changes the visibility of the name), don't warn that it
won't be visible outside the current scope. That's not true.
Also narrow down the set of cases where we create these extra declarations when
building modules; previously, all tag declarations but the first in a module
header would get this treatment if -fmodules-local-submodule-visibility. (This
isn't a functional change, but we try to avoid creating these extra
declarations whenever we can.)
James Y Knight [Mon, 11 Jan 2016 22:00:22 +0000 (22:00 +0000)]
PR26087: Use nonstandard MSVC extension for VS2015 as well.
In r256564, I had conditioned the workaround in has_getDecl to only be
used for MSVC before the 2015 release, believing that 2015 could handle
the standard code. But, that was incorrect.
Benjamin Kramer [Mon, 11 Jan 2016 16:27:16 +0000 (16:27 +0000)]
[clang-format] Fix comment aligning when there are changes within the comment
As soon as a comment had whitespace changes inside of the token, we
couldn't identify the whole comment as a trailing comment anymore and
alignment stopped working. Add a new boolean to Change for this special
case and fix trailing comment identification to use it.
This also changes WhitespaceManager to sum the length of all Changes
inside of a token into the first Change.
Daniel Jasper [Mon, 11 Jan 2016 11:01:05 +0000 (11:01 +0000)]
clang-format: Slightly row back on r257257.
r257257 change the way clang-format enforces line breaks after a
templated type has been line-wrapped. This was to fix an incorrect line
break if BinPackParameters is set to false. However, it also leads to
an unwanted line break in a different case. Thus, for now, only do this
when BinPackParameters is false. This isn't ideal yet, but helps us
until we have a better solution.
Yury Gribov [Mon, 11 Jan 2016 09:38:48 +0000 (09:38 +0000)]
AnalysisConsumer: use canonical decl for both lookup and store of
visited decls.
Due to redeclarations, the function may have different declarations used
in CallExpr and in the definition. However, we need to use a unique
declaration for both store and lookup in VisitedCallees. This patch
fixes issues with analysis in topological order. A simple test is
included.
Daniel Jasper [Sat, 9 Jan 2016 21:12:45 +0000 (21:12 +0000)]
clang-format: Fix the counting of leading whitespace in tok::unknown tokens
Previously, all whitespace characters would increase the starting
column, which doesn't make sense. This fixes a problem, e.g. with the
length calculation in JS template strings.
Benjamin Kramer [Sat, 9 Jan 2016 16:33:16 +0000 (16:33 +0000)]
[vfs] Normalize working directory if requested.
FixedCompilationDatabase sets the working dir to "." by default. For
chdir(".") this is a noop but this lead to InMemoryFileSystem to create
bogus paths. Fixes PR25327.
Richard Smith [Sat, 9 Jan 2016 06:58:48 +0000 (06:58 +0000)]
[modules] If we're treating an elaborated-type-specifier as if it introduces a
tag (because the previous declaration was found in a different module), inject
the tag into the appropriate scope (that is, the enclosing scope if we're in a
function prototype scope in C++).