Simon Pilgrim [Sun, 2 Aug 2015 15:28:10 +0000 (15:28 +0000)]
Fix invalid shufflevector operands
This patch fixes bug 23800 ( https://llvm.org/bugs/show_bug.cgi?id=23800#c2 ). There existed a case where the index operand from extractelement was directly used to create a shufflevector mask. Since the index can be of any integral type but the mask must only contain 32 bit integers a 64 bit index operand led to an assertion error later on.
Committed on behalf of mpflanzer (Moritz Pflanzer)
Most of these had no need to check `tag:` field, but did so as a way of
getting to the `name:` field. In a few cases I've converted the `tag:`
checks to `arg:` or `CHECK-NOT: arg:`.
David Majnemer [Fri, 31 Jul 2015 17:58:45 +0000 (17:58 +0000)]
[MS ABI] Hook clang up to the new EH instructions
The new EH instructions make it possible for LLVM to generate .xdata
tables that the MSVC personality routines will be happy about. Because
this is experimental, hide it behind a -cc1 flag (-fnew-ms-eh).
DI: Update for LLVM API change for local variables
Adjust to LLVM DIBuilder API changes in r243764, using
`createAutoVariable()` and `createParameterVariable()` in place of
`createLocalVariable()`. No real functionality change here.
David Blaikie [Thu, 30 Jul 2015 21:42:22 +0000 (21:42 +0000)]
Split DWARF: Allow -gmlt/-gsplit-dwarf to override rather than complement each other
It doesn't make any sense to enable -gmlt with -gsplit-dwarf, since
-gmlt is designed for on-line symbolication (and -gsplit-dwarf normally
emits all the -gmlt data into the .o anyway - so there's nothing to
split out except redundant/duplicate info).
With this change they override each other, -gmlt -gsplit-dwarf is the
same as -gsplit-dwarf and -gsplit-dwarf -gmlt is the same as -gmlt.
Summary:
Currently, if the argument to _Pragma is not a parenthesised string
literal, the bad token will be consumed, as well as the ')', if present.
If additional bad tokens are passed to the _Pragma, this results in
extra error messages which may distract from the true problem.
The proposed patch causes all tokens to be consumed until the closing
')' or a new line, whichever is reached first.
[ARM] Allow setting the generic features for native cpus
Copying the already existing code for x86 to ARM to set the correct CPU
features when using -mcpu=native. We can already detect the CPU name
but we were not setting the correct feature bits.
Moving fpu/hwdiv down to make sure they override whatever we set the
default to be.
No tests because this is native detection, and not all ARM-enabled builds
will hapen at a specific CPU, or even ARM. I have tested locally and it
works as expected.
This is because the "Ident_pixel" variable was uninitialized
in the getLangOpts().ZVector case, but we'd still call into
clang::Parser::TryAltiVecTokenOutOfLine, which uses the variable.
The simplest fix for this without sprinkling !getLangOpts().ZVector
checks all over the code seems to be to just initialize the variable
to nullptr; this will then do the right thing on ZVector.
[SystemZ] Add support for vecintrin.h vector built-in functions
This patch adds support for the System Z vector built-in functions.
The API-defined header file has the name vecintrin.h.
The user-level functions are defined in the same style as the clang
version of altivec.h, making heavy use of the __overloadable__ and
__always_inline__ attributes. Where possible the functions expand to
generic operations rather than specific built-in functions, in the hope
that that form can be optimised better.
Where a built-in routine is specified to require an immediate integer
argument, the __enable_if__ attribute is used to verify the argument is
in fact constant and in the appropriate range.
Add support for System z vector language extensions
The z13 vector facility has an associated language extension,
closely modeled on AltiVec/VSX. The main differences are:
- vector long, vector float and vector pixel are not supported
- vector long long and vector double are supported (like VSX)
- comparison operators return a vector rather than a scalar integer
- shift operators behave like the OpenCL shift operators
- vector bool is only supported as argument to certain operators;
some operators allow mixing a bool with a non-bool vector
This patch adds clang support for the extension. It is closely modelled
on the AltiVec support. Similarly to the -faltivec option, there's a
new -fzvector option to enable the extensions (as well as an -mzvector
alias for compatibility with GCC). There's also a separate LangOpt.
The extension as implemented here is intended to be compatible with
the -mzvector extension recently implemented by GCC.
[OPENMP 4.1] Initial support for extended 'ordered' clause.
OpenMP 4.1 introduces optional argument '(n)' for 'ordered' clause, where 'n' is a number of loops that immediately follow the directive.
'n' must be constant positive integer expressions and it must be less or equal than the number of the loops in the resulting loop nest.
Patch adds parsing and semantic analysis for this optional argument.
Sean Silva [Thu, 30 Jul 2015 01:21:56 +0000 (01:21 +0000)]
Remove bad test.
We currently don't canonicalize paths in the preprocessed files.
But we do when writing to PCH.
This causes a discrepancy on Windows with the test below.
This test fails even on unix if you change the test to use
`%S//preprocess.h`.
I am led to conclude that the invariant that this test was intending to
test has not been upheld for a while (and may never have been).
Sean Silva [Thu, 30 Jul 2015 00:26:34 +0000 (00:26 +0000)]
Avoid failure to canonicalize '..'.
Also fix completely broken and untested code which was hiding the
primary bug. The !LLVM_ON_UNIX branch of the ifdef was actually a no-op.
I ran into this in the wild. It was causing failures in our SDK build.
Ideally we'd have a perfect llvm::sys::fs::canonical, but at least this
is a step in the right direction, and fixes an obviously broken case.
In some sense the test case I've added here is an integration test. We
should have these routines thoroughly unit tested in llvm::sys::fs.
Richard Trieu [Wed, 29 Jul 2015 23:47:19 +0000 (23:47 +0000)]
Fix -Wredundant-move warning.
Without DR1579 implemented, the only case for -Wredundant-move is for a
parameter being returned with the same type as the function return type. Also
include a check to verify that the move constructor will be used by matching
nodes in the AST dump.
Richard Smith [Wed, 29 Jul 2015 23:38:25 +0000 (23:38 +0000)]
[modules] When performing redeclaration lookup for a using declaration, prefer
UsingShadowDecls over other declarations of the same entity in the lookup
results. This ensures that we build correct redeclaration chains for the
UsingShadowDecls (otherwise we could see assertions and other misbehavior in
modules builds, when merging combines multiple redeclaration chains for the
same entity from the same module into one chain).
Richard Trieu [Wed, 29 Jul 2015 17:03:34 +0000 (17:03 +0000)]
Disable -Wpessimizing-move and -Wredundant-move in template instantiations.
Dependent types can throw off the analysis for these warnings, possibly giving
conflicting warnings and fix-its. Disabling the warning in template
instantiations will prevent this problem, and will still catch the
non-dependent cases in templates.
In case of an existing GlobalVariable, the comdat is created using the name of the
new GV (usually NAME.1) instead of the correct NAME of the old GV. Moving comdat
creation after GV replacement solves this. Patch + testcase.
This commit changes the driver to save subtarget feature "+strict-align"
to the IR instead of using backend option "aarch64-strict-align". This is
needed for LTO.
This commit changes the driver to save subtarget feature "+strict-align" to the
IR instead of using backend option "arm-strict-align". This is needed for LTO.
Also, move the logic in ARM backend that was deciding whether strict alignment
should be forced to the front-end.
--cuda-host-only should not disable linking phase.
Host-only cuda compilation does produce valid host object
file and in some cases users do want to proceed on to the linking phase.
The change removes special case that stopped compilation pipeline at
the Assembly phase. Device-side compilation is still stopped early
by the types::getCompilationPhases().
Richard Trieu [Tue, 28 Jul 2015 20:53:46 +0000 (20:53 +0000)]
Use an iterative method instead of recursion for printing macro backtraces.
Store the locations for a macro expansion in a vector, then iterate over them
instead of using recursion. This simplifies the logic around the backtrace
limit and gives easier access to the source locations. No functionality change.
Richard Trieu [Tue, 28 Jul 2015 19:06:16 +0000 (19:06 +0000)]
Do not give a -Wredundant-move warning when removing the move will result in an
error.
If the object being moved has a move constructor and a deleted copy constructor,
std::move is required, otherwise Clang will give a deleted constructor error.
Also move "pragma comment" warning from -Wmicrosoft to -Wignored-pragmas.
-Wmicrosoft currently covers many different areas, some more useful than
others. Split it into many targeted flags, so that projects can choose to
enable only a subset of these warnings. This is also useful for incrementally
fixing and turning on these warnings.
-Wno-microsoft still disables all these warnings, and -Wmicrosoft still enables
them all. After this change, it's possible to pass `-Wno-microsoft
-Wmicrosoft-unqualified-friend` to only enable -Wmicrosoft-unqualified-friend,
and `-Wmicrosoft -Wno-microsoft-unqualified-friend` to enable all other
Microsoft warnings.
I put all the template-related warnings behind -Wmicrosoft-template; if that
turns out to be too coarse we can make that finer later on. (In practice, I
haven't seen the template-related warnings fire frequently.)
InstrProf: Fix a misuse of the FunctionDecl API when generating coverage
This was calling FD->hasBody(), meaning "Does the function that this
decl refers to have a body?", rather than
FD->doesThisDeclarationHaveABody(), meaning "Is this decl a
non-deleted definition?".
[ARM] Implement -Wa,-mfpu and friends for assemblers
This patch allows Clang to pass on -Wa,-mfpu, -Wa,-mhwdiv and
-Wa,-mcpu to the integrated assembler (via target-features), but
-march is still not being passed, but validated.
In case the command line has both -mxxx and -Wa,-mxxx, we warn
that the naked one will not be used in assembler mode.
[ARM] Changes to Arch/CPU handling to help with -Wa,-mfpu etc. (nfc)
To be able to handle -Wa, options in the assembler (ClangAs), we need to
make the handling of options based on the value of the options, not direct
Arguments from the list, since the list is immutable.
No functional change in this patch, but this allows validating of -Wa,-mfpu
and friends in the same way we validate -mfpu and friends, *just* for the
assembler.
Use CGLoopInfo to emit metadata for loop hint pragmas.
When ‘#pragma clang loop vectorize(assume_safety)’ was specified on a loop other loop hints were lost. The problem is that CGLoopInfo attaches metadata differently than EmitCondBrHints in CGStmt. For do-loops CGLoopInfo attaches metadata to the br in the body block and for while and for loops, the inc block. EmitCondBrHints on the other hand always attaches data to the br in the cond block. When specifying assume_safety CGLoopInfo emits an empty llvm.loop metadata shadowing the metadata in the cond block. Loop transformations like rotate and unswitch would then eliminate the cond block and its non-empty metadata.
This patch unifies both approaches for adding metadata and modifies the existing safety tests to include non-assume_safety loop hints.
Simon Pilgrim [Mon, 27 Jul 2015 19:01:52 +0000 (19:01 +0000)]
[X86] Add missing _m_prefetch intrinsic
The 3DNOW/PRFCHW cpu targets define both the PREFETCHW (set cache line modified) and PREFETCH (set cache line exclusive) instructions but only the _m_prefetchw (PREFETCHW) intrinsic is included in the header. This patch adds the missing _m_prefetch intrinsic.
I'm basing this off AMD documentation - the intel docs on the support for PREFETCHW isn't clear whether Silvermont/Broadwell properly support PREFETCH but given that the intrinsic implementation is a default __builtin_prefetch call, it is safe whatever.
Currently trigger to select hard-float linker is only based of -gnueabihf
appearing in target triplet, but we should also select it when hardfloat
is requested via cmdline.
David Majnemer [Mon, 27 Jul 2015 07:32:11 +0000 (07:32 +0000)]
[clang-cl] Handle -O correctly
We had multiple bugs here:
- We didn't support multiple optimization options in one argument.
e.g. -O2y-
- We didn't correctly expand -O[12dx] to their respective options.
- We treated -O1 as clang -O1 instead of clang -Os.
- We treated -Ox as clang -O3 instead of clang -O2. In fact, cl's -Ox
option is *less* powerful than cl's -O2 option despite -Ox described
as "Full Optimization".
Rename highly-confusing isWhitespace that returns false on \n to
isWhitespaceExceptNL. If that's wasn't bad enough, we have an identically
named function in Basic/CharInfo.h.
While at it, remove unnecessary includes, .str() calls and #ifdef.
projects / clang / log