Erich Keane [Tue, 10 Jul 2018 20:46:46 +0000 (20:46 +0000)]
[NFC] Switch CodeGenFunction to use value init instead of member init lists
The member init list for the sole constructor for CodeGenFunction
has gotten out of hand, so this patch moves the non-parameter-dependent
initializations into the member value inits.
Summary:
Make sure that loop metadata only is put on the backedge
when expanding a do-while loop.
Previously we added the loop metadata also on the branch
in the pre-header. That could confuse optimization passes
and result in the loop metadata being associated with the
wrong loop.
Fixes https://bugs.llvm.org/show_bug.cgi?id=38011
Committing on behalf of deepak2427 (Deepak Panickal)
Paul Robinson [Tue, 10 Jul 2018 15:15:24 +0000 (15:15 +0000)]
Support -fdebug-prefix-map for assembler source (pass to cc1as). This
is useful to omit the debug compilation dir when compiling assembly
files with -g. Part of PR38050.
[analyzer] Add option to set maximum symbol complexity threshold
Summary:
This adds an option, max-symbol-complexity, so an user can set the maximum symbol complexity threshold.
Note that the current behaviour is equivalent to max complexity = 0, when taint analysis is not enabled and tests show that in a number of tests, having complexity = 25 yields the same results as complexity = 10000.
This patch was extracted and modified from Dominic Chen's patch, D35450.
[modules] Fix 37878; Autoload subdirectory modulemaps with specific LangOpts
Summary:
Reproducer and errors:
https://bugs.llvm.org/show_bug.cgi?id=37878
lookupModule was falling back to loadSubdirectoryModuleMaps when it couldn't
find ModuleName in (proper) search paths. This was causing iteration over all
files in the search path subdirectories for example "/usr/include/foobar" in
bugzilla case.
Users don't expect Clang to load modulemaps in subdirectories implicitly, and
also the disk access is not cheap.
if (AllowExtraModuleMapSearch) true with ObjC with @import ModuleName.
Martin Storsjo [Tue, 10 Jul 2018 10:46:51 +0000 (10:46 +0000)]
[MinGW] Skip adding default win32 api libraries if -lwindowsapp is specified
In this setup, skip adding all the default windows import libraries,
if linking to windowsapp (which replaces them, when targeting the
windows store/UWP api subset).
With GCC, the same is achieved by using a custom spec file, but
since clang doesn't use spec files, we have to allow other means of
overriding what default libraries to use (without going all the
way to using -nostdlib, which would exclude everything). The same
approach, in detecting certain user specified libraries and omitting
others from the defaults, was already used in SVN r314138.
Martin Storsjo [Tue, 10 Jul 2018 10:46:45 +0000 (10:46 +0000)]
[MinGW] Treat any -lucrt* as replacing -lmsvcrt
Since SVN r314138, we check if the user has specified any particular
alternative msvcrt/ucrt version, and skip the default -lmsvcrt
in those cases.
In addition to the existing names checked, we should also treat
a plain -lucrt in the same way, mingw-w64 has now added a separate
import library named libucrt.a, in addition to libucrtbase.a.
[X86] Lowering integer truncation intrinsics to native IR
This patch lowers the _mm[256|512]_cvtepi{64|32|16}_epi{32|16|8} intrinsics to
native IR in cases where the result's length is less than 128 bits.
The resulting IR for 256-bit inputs is folded into VPMOV instructions, while for
128-bit inputs the vpshufb (or, in the 64-to-32-bit case, vinsertps)
instructions are generated instead
[X86] Use masked the masked scalar fma builtins to implement the default rounding version of the fma intrinsics.
The rounding mode is checked in CGBuiltin.cpp to generate the correct intrinsic call.
Making this switch switchs the masking to use the i8 bitcast to <8 x i1> and extract i1 version of the IR for the mask. Previously we ended up with a scalar 'and' plus an icmp.
Richard Trieu [Tue, 10 Jul 2018 01:40:50 +0000 (01:40 +0000)]
[ODRHash] Merge the two function hashes into one.
Functions that are a sub-Decl of a record were hashed differently than other
functions. This change keeps the AddFunctionDecl function and the hash of
records now calls this function. In addition, AddFunctionDecl has an option
to perform a hash as if the body was absent, which is required for some
checks after loading modules. Additional logic prevents multiple error
message from being printed.
Fix parsing of privacy annotations in os_log format strings.
Privacy annotations shouldn't have to appear in the first
comma-delimited string in order to be recognized. Also, they should be
ignored if they are preceded or followed by non-whitespace characters.
[X86] Add __builtin_ia32_selectss_128 and __builtin_ia32_selectsd_128 that is suitable for use in scalar mask intrinsics.
This will convert the i8 mask argument to <8 x i1> and extract an i1 and then emit a select instruction. This replaces the '(__U & 1)" and ternary operator used in some of intrinsics. The old sequence was lowered to a scalar and and compare. The new sequence uses an i1 vector that will interoperate better with other mask intrinsics.
This removes the need to handle div_ss/sd specially in CGBuiltin.cpp. A follow up patch will add the GCCBuiltin name back in llvm and remove the custom handling.
I made some adjustments to legacy move_ss/sd intrinsics which we reused here to do a simpler extract and insert instead of 2 extracts and two inserts or a shuffle.
[OPENMP] Do not mark local variables as declare target.
When the parsing of the functions happens inside of the declare target
region, we may erroneously mark local variables as declare target
thought they are not. This attribute can be applied only to global
variables.
Alex Lorenz [Mon, 9 Jul 2018 19:41:28 +0000 (19:41 +0000)]
[libclang] evalute compound statement cursors before trying to evaluate
the cursor like a declaration
This change fixes a bug in libclang in which it tries to evaluate a statement
cursor as a declaration cursor, because that statement still has a pointer to
the declaration parent.
[Builtins][Attributes][X86] Tag all X86 builtins with their required vector width. Add a min_vector_width function attribute and tag all x86 instrinsics with it
This is part of an ongoing attempt at making 512 bit vectors illegal in the X86 backend type legalizer due to CPU frequency penalties associated with wide vectors on Skylake Server CPUs. We want the loop vectorizer to be able to emit IR containing wide vectors as intermediate operations in vectorized code and allow these wide vectors to be legalized to 256 bits by the X86 backend even though we are targetting a CPU that supports 512 bit vectors. This is similar to what happens with an AVX2 CPU, the vectorizer can emit wide vectors and the backend will split them. We want this splitting behavior, but still be able to use new Skylake instructions that work on 256-bit vectors and support things like masking and gather/scatter.
Of course if the user uses explicit vector code in their source code we need to not split those operations. Especially if they have used any of the 512-bit vector intrinsics from immintrin.h. And we need to make it so that merely using the intrinsics produces the expected code in order to be backwards compatible.
To support this goal, this patch adds a new IR function attribute "min-legal-vector-width" that can indicate the need for a minimum vector width to be legal in the backend. We need to ensure this attribute is set to the largest vector width needed by any intrinsics from immintrin.h that the function uses. The inliner will be reponsible for merging this attribute when a function is inlined. We may also need a way to limit inlining in the future as well, but we can discuss that in the future.
To make things more complicated, there are two different ways intrinsics are implemented in immintrin.h. Either as an always_inline function containing calls to builtins(can be target specific or target independent) or vector extension code. Or as a macro wrapper around a taget specific builtin. I believe I've removed all cases where the macro was around a target independent builtin.
To support the always_inline function case this patch adds attribute((min_vector_width(128))) that can be used to tag these functions with their vector width. All x86 intrinsic functions that operate on vectors have been tagged with this attribute.
To support the macro case, all x86 specific builtins have also been tagged with the vector width that they require. Use of any builtin with this property will implicitly increase the min_vector_width of the function that calls it. I've done this as a new property in the attribute string for the builtin rather than basing it on the type string so that we can opt into it on a per builtin basis and avoid any impact to target independent builtins.
There will be future work to support vectors passed as function arguments and supporting inline assembly. And whatever else we can find that isn't covered by this patch.
Special thanks to Chandler who suggested this direction and reviewed a preview version of this patch. And thanks to Eric Christopher who has had many conversations with me about this issue.
[OPENMP, NVPTX] Do not globalize local variables in parallel regions.
In generic data-sharing mode we are allowed to not globalize local
variables that escape their declaration context iff they are declared
inside of the parallel region. We can do this because L2 parallel
regions are executed sequentially and, thus, we do not need to put
shared local variables in the global memory.
[PCH] Add an option to not write comments into PCH
Summary:
Will be used in clangd, see the follow-up change.
Clangd does not use comments read from PCH to avoid crashes due to
changed contents of the file. However, reading them considerably slows
down code completion on files with large preambles.
[Preamble] Check system dependencies in preamble too
Summary:
PrecompiledPreamble hasn't checked if the system dependencies changed
before. This resulted in invalid preamble not being rebuilt if headers
that changed were found in -isystem include paths.
This pattern is sometimes used to avoid showing warnings in third
party code, so we want to correctly handle those cases.
Jacek Olesiak [Mon, 9 Jul 2018 07:08:45 +0000 (07:08 +0000)]
[clang-format/ObjC] Put ObjC method arguments into one line when they fit
Reapply D47195:
Currently BreakBeforeParameter is set to true everytime message receiver spans multiple lines, e.g.:
```
[[object block:^{
return 42;
}] aa:42 bb:42];
```
will be formatted:
```
[[object block:^{
return 42;
}] aa:42
bb:42];
```
even though arguments could fit into one line. This change fixes this behavior.
Jacek Olesiak [Mon, 9 Jul 2018 06:04:58 +0000 (06:04 +0000)]
[clang-format/ObjC] Prohibit breaking after a bracket opening ObjC method expression
Summary:
Don't break after a "[" opening an ObjC method expression.
Tests are added in D48719 where formatting is improved (to avoid adding and changing tests immediately).
Jacek Olesiak [Mon, 9 Jul 2018 05:58:51 +0000 (05:58 +0000)]
[clang-format/ObjC] Fix counting selector name parts for ObjC
Summary:
Counts selector parts also for method declarations and counts correctly for methods without arguments.
This is an internal change and doesn't influence formatting on its own (at the current state). Its lack would be visible after applying D48719.
Reka Kovacs [Sat, 7 Jul 2018 19:27:18 +0000 (19:27 +0000)]
[analyzer] Highlight c_str() call in DanglingInternalBufferChecker.
Add a bug visitor to DanglingInternalBufferChecker that places a note
at the point where the dangling pointer was obtained. The visitor is
handed over to MallocChecker and attached to the report there.
[MachineOutliner] Properly pass -moutline along to the toolchain
This moves the LTO-specific code for outlining from ToolChains/Clang.cpp to
ToolChains/Darwin.cpp. Passing -mllvm flags isn't sufficient for making sure
that the specified pass will actually run in LTO. This makes sure that when
-moutline is passed, the MachineOutliner will actually be added to the LTO
pass pipeline as expected.
[OPENMP] Make clauses closing loc point to right bracket.
For some of the clauses the closing location erroneously points to the
beginning of the next clause rather than on the location of the closing
bracket of the clause.
This patches adds support for passing -mcpu=native for AArch64. It will
get turned into the host CPU name, before we get the target features.
CPU = native is handled in a similar fashion in
getAArch64MicroArchFetauresFromMtune and getAArch64TargetCPU already.
Having a good test case for this is hard, as it depends on the host CPU
of the machine running the test. But we can check that native has been
replaced with something else.
When cross-compiling, we will get a CPU name from the host architecture
and get ` the clang compiler does not support '-mcpu=native'` as error
message, which seems reasonable to me.
Eric Liu [Fri, 6 Jul 2018 09:43:57 +0000 (09:43 +0000)]
[SemaCodeComplete] Expose a method to create CodeCompletionString for macros.
Summary:
The method only takes PPreprocessor and don't require structures that
might not be available (e.g. Sema and ASTContext) when CodeCompletionString
needs to be generated for macros.
Hans Wennborg [Fri, 6 Jul 2018 06:54:16 +0000 (06:54 +0000)]
[ms] Fix mangling of string literals used to initialize arrays larger or smaller than the literal
A Chromium developer reported a bug which turned out to be a mangling
collision between these two literals:
char s[] = "foo";
char t[32] = "foo";
They may look the same, but for the initialization of t we will (under
some circumstances) use a literal that's extended with zeros, and
both the length and those zeros should be accounted for by the mangling.
This actually makes the mangling code simpler: where it previously had
special logic for null terminators, which are not part of the
StringLiteral, that is now covered by the general algorithm.
(The problem was reported at https://crbug.com/857442)
Alex Lorenz [Thu, 5 Jul 2018 22:51:11 +0000 (22:51 +0000)]
[Sema] -Wformat-pedantic only for NSInteger/NSUInteger %tu/%td on Darwin
The '%tu'/'%td' as formatting specifiers have been used to print out the
NSInteger/NSUInteger values for a long time. Typically their ABI matches, but that's
not the case on watchOS. The ABI difference boils down to the following:
- Regular 32-bit darwin targets (like armv7) use 'ptrdiff_t' of type 'int',
which matches 'NSInteger'.
- WatchOS arm target (armv7k) uses 'ptrdiff_t' of type 'long', which doesn't
match 'NSInteger' of type 'int'.
Because of this ABI difference these specifiers trigger -Wformat warnings only
for watchOS builds, which is really inconvenient for cross-platform code.
This patch avoids this -Wformat warning for '%tu'/'%td' and NS[U]Integer only,
and instead uses the new -Wformat-pedantic warning that JF introduced in
https://reviews.llvm.org/D47290. This is acceptable because Darwin guarantees that,
despite the watchOS ABI differences, sizeof(ptrdiff_t) == sizeof(NS[U]Integer),
and alignof(ptrdiff_t) == alignof(NS[U]Integer) so the warning is therefore noisy
for pedantic reasons.
I'll update public documentation to ensure that this behaviour is properly
communicated.
Erich Keane [Thu, 5 Jul 2018 17:22:13 +0000 (17:22 +0000)]
[clang-cl, PCH] Implement support for MS-style PCH through headers
Implement support for MS-style PCH through headers.
This enables support for /Yc and /Yu where the through header is either
on the command line or included in the source. It replaces the current
support the requires the header also be specified with /FI.
This change adds a -cc1 option -pch-through-header that is used to either
start or stop compilation during PCH create or use.
When creating a PCH, the compilation ends after compilation of the through
header.
When using a PCH, tokens are skipped until after the through header is seen.
Erich Keane [Thu, 5 Jul 2018 15:52:58 +0000 (15:52 +0000)]
Fix __builtin_*_overflow when out-param isn't constexpr
As brought up on cfe-commits[1], r334650 causes the dependency of the
out parameter to the __builtin_*_overflow functions to be ignored. The result
was a usage that was otherwise constexpr (both operands to the operation were
constexpr) would be evaluated, but the out parameter wouldn't be modified, so
it would still be 'undef'.
This patch correctly handles the return value of handleAssignment to ensure that
this value is properly considered/evaluated.
[X86] Fix some vector cmp builtins - TRUE/FALSE predicates
This patch removes on optimization used with the TRUE/FALSE
predicates, as was suggested in https://reviews.llvm.org/D45616
for r335339.
The optimization was buggy, since r335339 used it also
for *_mask builtins, without actually applying the mask -- the
mask argument was just ignored.
[X86] NFC - add more test cases for vector cmp intrinsics
Add test cases with each predicate using the following
intrinsics:
_mm_cmp_pd
_mm_cmp_ps
_mm256_cmp_pd
_mm256_cmp_ps
_mm_cmp_pd_mask
_mm_cmp_ps_mask
_mm256_cmp_pd_mask
_mm256_cmp_ps_mask
_mm512_cmp_pd_mask
_mm512_cmp_ps_mask
_mm_mask_cmp_pd_mask
_mm_mask_cmp_ps_mask
_mm256_mask_cmp_pd_mask
_mm256_mask_cmp_ps_mask
_mm512_mask_cmp_pd_mask
_mm512_mask_cmp_ps_mask
Some of these are marked with FIXME, as there is bug in lowering
e.g. _mm512_mask_cmp_ps_mask.
[ASTImporter] Fix import of objects with anonymous types
Summary:
Currently, anonymous types are merged into the same redecl chain even if they
are structurally inequivalent. This results that global objects are not
imported, if there are at least two global objects with different anonymous
types. This patch provides a fix.
This patch is a preparation for another one containing meaningful
changes. This patch simply removes trailing whitespaces in few files
affected by the upcoming patch and reformats
Richard Smith [Wed, 4 Jul 2018 02:25:38 +0000 (02:25 +0000)]
PR33924: merge local declarations that have linkage of some kind within
merged function definitions; also merge functions with deduced return
types.
This seems like two independent fixes, but unfortunately they are hard
to separate because it's challenging to reliably test either one of them
without also testing the other.
A complication arises with deduced return type support: we need the type
of the function in order to know how to merge it, but we can't load the
actual type of the function because it might reference an entity
declared within the function (and we need to have already merged the
function to correctly merge that entity, which we would need to do to
determine if the function types match). So we instead compare the
declared function type when merging functions, and defer loading the
actual type of a function with a deduced type until we've finished
loading and merging the function.
This reverts r336175, reinstating r336021, with one change (for PR38015):
we look at the TypeSourceInfo of the first-so-far declaration of each
function when considering whether to merge two functions. This works
around a problem where the calling convention in the TypeSourceInfo for
subsequent redeclarations may not match if it was implicitly adjusted.
Michael Kruse [Wed, 4 Jul 2018 01:37:11 +0000 (01:37 +0000)]
[Sema] Consider all format_arg attributes.
If a function has multiple format_arg attributes, clang only considers
the first it finds (because AttributeLists are in reverse order, not
necessarily the textually first) and ignores all others.
Loop over all FormatArgAttr to print warnings for all declared
format_arg attributes.
For instance, libintl's ngettext (select plural or singular version of
format string) has two __format_arg__ attributes.
Erik Pilkington [Tue, 3 Jul 2018 22:15:36 +0000 (22:15 +0000)]
[Sema] Discarded statment should be an evaluatable context.
The constexpr evaluator was erroring out because these templates weren't
defined. Despite being used in a discarded statement, we still need to constexpr
evaluate them, which means that we need to instantiate them. Fixes PR37585.
Erich Keane [Tue, 3 Jul 2018 20:30:34 +0000 (20:30 +0000)]
Fix allocation of Nullability attribute.
Existing code always allocates for on the declarator's attribute pool,
but sometimes adds it to the declspec. This patch ensures that the
correct pool is used.
Discovered while testing: https://reviews.llvm.org/D48788
This happened during a recent refactor. toStringRefArray() returns
a vector<StringRef>, which was being implicitly converted to an
ArrayRef<StringRef>, and then the vector was immediately being
destroyed, so the ArrayRef<> was losing its backing storage.
Fix this by making sure the vector gets permanent storage.
Summary:
Scudo works on PPC64 as is, so mark the architecture as supported for it. This
will also require a change to config-ix.cmake on the compiler-rt side.
Hans Wennborg [Tue, 3 Jul 2018 07:51:41 +0000 (07:51 +0000)]
Revert r336021 "PR33924: merge local declarations that have linkage of some kind within"
This caused test failures in 32-bit builds (PR38015).
> merged function definitions; also merge functions with deduced return
> types.
>
> This seems like two independent fixes, but unfortunately they are hard
> to separate because it's challenging to reliably test either one of them
> without also testing the other.
>
> A complication arises with deduced return type support: we need the type
> of the function in order to know how to merge it, but we can't load the
> actual type of the function because it might reference an entity
> declared within the function (and we need to have already merged the
> function to correctly merge that entity, which we would need to do to
> determine if the function types match). So we instead compare the
> declared function type when merging functions, and defer loading the
> actual type of a function with a deduced type until we've finished
> loading and merging the function.
Steven Wu [Tue, 3 Jul 2018 04:15:49 +0000 (04:15 +0000)]
[Driver][Darwin] Use Host Triple to infer target os version
Summary:
When clang required to infer target os version from --target option and
the os version is not specified in targets, check the host triple. If the
host and target are both macOS, use host triple to infer target os
version.
projects / clang / log