Matt Arsenault [Sun, 13 Nov 2016 18:20:54 +0000 (18:20 +0000)]
AMDGPU: Implement SGPR spilling with scalar stores
nThis avoids the nasty problems caused by using
memory instructions that read the exec mask while
spilling / restoring registers used for control flow
masking, but only for VI when these were added.
This always uses the scalar stores when enabled currently,
but it may be better to still try to spill to a VGPR
and use this on the fallback memory path.
The cache also needs to be flushed before wave termination
if a scalar store is used.
Analysis: Simplify the ScalarEvolution::getGEPExpr() interface. NFCI.
All existing callers were manually extracting information out of an existing
GEP instruction and passing it to getGEPExpr(). Simplify the interface by
changing it to take a GEPOperator instead.
Lang Hames [Sat, 12 Nov 2016 23:12:41 +0000 (23:12 +0000)]
[ORC] Remove the 'const' qualifier from the member function wrapper, make the
lambda in wrapHandler mutable to allow it to pass the handler through as a
non-const value.
llvm-strings: trivialise logic until we support more options
Until we have handling for ignoring unloaded sections, simplify the logic to
the point of triviality. This fixes the scanning of archives, particularly when
embedded in archives.
Craig Topper [Sat, 12 Nov 2016 18:04:46 +0000 (18:04 +0000)]
[AVX-512] Remove the remaining masked shift by immediate or by single value. Autoupgrade them to recently introduced unmasked versions and a select.
After this I'll add the unmasked intrinsics to InstCombineCalls to finish making our handling of these types of shuffles consistent between AVX-512 and the legacy intrinsics.
Michal Gorny [Sat, 12 Nov 2016 14:58:30 +0000 (14:58 +0000)]
[OCaml] Clear cross-target test deps when building out-of-tree
Clear cross-target test dependencies when using LLVM_OCAML_OUT_OF_TREE,
in order to make it possible to run check-llvm-bindings-ocaml without
rebuilding the whole LLVM.
Craig Topper [Sat, 12 Nov 2016 05:28:24 +0000 (05:28 +0000)]
[AVX-512] Add unmasked version of shift by immediate and shift by single element in XMM.
Summary:
This is the first step towards being able to add the avx512 shift by immediate intrinsics to InstCombineCalls where we aleady support the sse2 and avx2 intrinsics. We need to the unmasked versions so we can avoid having to teach InstCombineCalls that it would need to insert selects sometimes. Instead we'll just add the selects around the new instrinsics in the frontend.
This change should also enable the shift by i32 intrinsics to take a non-constant shift value just like the avx2 and sse intrinsics. This will enable us to fix PR30691 once we update clang.
Next I'll switch clang to use the new builtins. Then we'll come back to the backend and remove/autoupgrade the old intrinsics. Then I'll work on the same series for variable shifts.
Craig Topper [Sat, 12 Nov 2016 05:05:27 +0000 (05:05 +0000)]
[AVX-512] Add support for lowering shuffles to VALIGND/VALIGNQ
Summary: VALIGND and VALIGNQ are similar to PALIGNR but instead of working on a 128-bit lane they work on the entire vector register. This change leverages the shuffle rotate detection code used for PALIGNR to detect these cases.
Lang Hames [Sat, 12 Nov 2016 02:19:31 +0000 (02:19 +0000)]
[ORC] Add a WrappedHandlerReturn type to map handler return types onto error
return types.
This class allows user provided handlers to return either error-wrapped types
or plain types. In the latter case, the plain type is wrapped with a success
value of Error or Expected<T> type to fit it into the rest of the serialization
machinery.
This patch allows us to remove the RPC unit-test workaround added in r286646.
Zachary Turner [Fri, 11 Nov 2016 23:57:40 +0000 (23:57 +0000)]
[Support] Introduce llvm::formatv() function.
This introduces a new type-safe general purpose formatting
library. It provides compile-time type safety, does not require
a format specifier (since the type is deduced), and provides
mechanisms for extending the format capability to user defined
types, and overriding the formatting behavior for existing types.
This patch additionally adds documentation for the API to the
LLVM programmer's manual.
Mailing List Thread:
http://lists.llvm.org/pipermail/llvm-dev/2016-October/105836.html
Rui Ueyama [Fri, 11 Nov 2016 23:41:13 +0000 (23:41 +0000)]
Define DbiStreamBuilder::addSectionContribs.
This patch defines a new function to add a SectionContribs stream
to a PDB file. Unlike SectionMap, SectionContribs contains a list
of input sections as opposed to output sections.
Note that this patch needs improving because currently we do not
set Module field in SectionContribs entries. In a follow-up patch,
I'll add Modules and then fix it after that.
Tom Stellard [Fri, 11 Nov 2016 23:35:42 +0000 (23:35 +0000)]
AMDGPU/SI: Fix visit order assumption in SIFixSGPRCopies
Summary:
This pass was assuming that when a PHI instruction defined a register
used by another PHI instruction that the defining insstruction would
be legalized before the using instruction.
This assumption was causing the pass to not legalize some PHI nodes
within divergent flow-control.
Anna Zaks [Fri, 11 Nov 2016 23:01:02 +0000 (23:01 +0000)]
[tsan][llvm] Implement the function attribute to disable TSan checking at run time
This implements a function annotation that disables TSan checking for the
function at run time. The benefit over attribute((no_sanitize("thread")))
is that the accesses within the callees will also be suppressed.
The motivation for this attribute is a guarantee given by the objective C
language that the calls to the reference count decrement and object
deallocation will be synchronized. To model this properly, we would need
to intercept all ref count decrement calls (which are very common in ObjC
due to use of ARC) and also every single message send. Instead, we propose
to just ignore all accesses made from within dealloc at run time. The main
downside is that this still does not introduce any synchronization, which
means we might still report false positives if the code that relies on this
synchronization is not executed from within dealloc. However, we have not seen
this in practice so far and think these cases will be very rare.
Unfortunately given the current structure of optimization diagnostics we
lack the capability to tell whether the user has
passed -Rpass-analysis=loop-vectorize since this is local to the
front-end (BackendConsumer::OptimizationRemarkHandler).
So rather than printing this even if the user has already
passed -Rpass-analysis, this patch just punts and stops recommending
this option. I don't think that getting this right is worth the
complexity.
Matthias Braun [Fri, 11 Nov 2016 22:37:34 +0000 (22:37 +0000)]
MachineScheduler/ScheduleDAG: Add support to skipping a node.
The DAG mutators in the scheduler cannot really remove DAG nodes as
additional anlysis information such as ScheduleDAGToplogicalSort are
already computed at this point and rely on a fixed number of DAG nodes.
Alleviate the missing removal with a new flag: Setting the new skip
flag on a node ignores it during scheduling.
Erik Eckstein [Fri, 11 Nov 2016 22:21:39 +0000 (22:21 +0000)]
FunctionComparator: don't rely on argument evaluation order.
This is a follow-up on the recent refactoring of the FunctionMerge pass.
It should fix a fail of the new FunctionComparator unittest whe compiling with MSVC.
Evgeniy Stepanov [Fri, 11 Nov 2016 21:39:26 +0000 (21:39 +0000)]
[cfi] Fix weak functions handling.
When a function pointer is replaced with a jumptable pointer, special
case is needed to preserve the semantics of extern_weak functions.
Since a jumptable entry can not be extern_weak, we emulate that
behaviour by replacing all references to F (the extern_weak function)
with the following expression: F != nullptr ? JumpTablePtr : nullptr.
Extra special care is needed for global initializers, since most (or
probably all) backends can not lower an initializer that includes
this kind of constant expression. Initializers like that are replaced
with a global constructor (i.e. a runtime initializer).
Erik Eckstein [Fri, 11 Nov 2016 21:15:13 +0000 (21:15 +0000)]
Make the FunctionComparator of the MergeFunctions pass a stand-alone utility.
This is pure refactoring. NFC.
This change moves the FunctionComparator (together with the GlobalNumberState
utility) in to a separate file so that it can be used by other passes.
For example, the SwiftMergeFunctions pass in the Swift compiler:
https://github.com/apple/swift/blob/master/lib/LLVMPasses/LLVMMergeFunctions.cpp
Details of the change:
*) The big part is just moving code out of MergeFunctions.cpp into FunctionComparator.h/cpp
*) Make FunctionComparator member functions protected (instead of private)
so that a derived comparator class can use them.
Following refactoring helps to share code between the base FunctionComparator
class and a derived class:
*) Add a beginCompare() function
*) Move some basic function property comparisons into a separate function compareSignature()
*) Do the GEP comparison inside cmpOperations() which now has a new
needToCmpOperands reference parameter
Fixed the lost FastMathFlags for FCmp operations in SLPVectorizer.
Reviewer: Michael Zolotukhin.
Differential Revision: https://reviews.llvm.org/D26543
Bitcode: Clean up error handling for certain bitcode query functions.
The functions getBitcodeTargetTriple(), isBitcodeContainingObjCCategory(),
getBitcodeProducerString() and hasGlobalValueSummary() now return errors
via their return value rather than via the diagnostic handler.
To make this work, re-implement these functions using non-member functions
so that they can be used without the LLVMContext required by BitcodeReader.
Lang Hames [Fri, 11 Nov 2016 19:42:44 +0000 (19:42 +0000)]
[ORC] Refactor the ORC RPC utilities to add some new features.
(1) Add support for function key negotiation.
The previous version of the RPC required both sides to maintain the same
enumeration for functions in the API. This means that any version skew between
the client and server would result in communication failure.
With this version of the patch functions (and serializable types) are defined
with string names, and the derived function signature strings are used to
negotiate the actual function keys (which are used for efficient call
serialization). This allows clients to connect to any server that supports a
superset of the API (based on the function signatures it supports).
(2) Add a callAsync primitive.
The callAsync primitive can be used to install a return value handler that will
run as soon as the RPC function's return value is sent back from the remote.
(3) Launch policies for RPC function handlers.
The new addHandler method, which installs handlers for RPC functions, takes two
arguments: (1) the handler itself, and (2) an optional "launch policy". When the
RPC function is called, the launch policy (if present) is invoked to actually
launch the handler. This allows the handler to be spawned on a background
thread, or added to a work list. If no launch policy is used, the handler is run
on the server thread itself. This should only be used for short-running
handlers, or entirely synchronous RPC APIs.
(4) Zero cost cross type serialization.
You can now define serialization from any type to a different "wire" type. For
example, this allows you to call an RPC function that's defined to take a
std::string while passing a StringRef argument. If a serializer from StringRef
to std::string has been defined for the channel type this will be used to
serialize the argument without having to construct a std::string instance.
This allows buffer reference types to be used as arguments to RPC calls without
requiring a copy of the buffer to be made.
Evgeniy Stepanov [Fri, 11 Nov 2016 18:49:09 +0000 (18:49 +0000)]
[cfi] Implement cfi-icall using inline assembly.
The current implementation is emitting a global constant that happens
to evaluate to the same bytes + relocation as a jump instruction on
X86. This does not work for PIE executables and shared libraries
though, because we end up with a wrong relocation type. And it has no
chance of working on ARM/AArch64 which use different relocation types
for jump instructions (R_ARM_JUMP24) that is never generated for
data.
This change replaces the constant with module-level inline assembly
followed by a hidden declaration of the jump table. Works fine for
ARM/AArch64, but has some drawbacks.
* Extra symbols are added to the static symbol table, which inflate
the size of the unstripped binary a little. Stripped binaries are not
affected. This happens because jump table declarations must be
external (because their body is in the inline asm).
* Original functions that were anonymous are now named
<original name>.cfi, and it affects symbolization sometimes. This is
necessary because the only user of these functions is the (inline
asm) jump table, so they had to be added to @llvm.used, which does
not allow unnamed functions.
Adrian Prantl [Fri, 11 Nov 2016 17:50:09 +0000 (17:50 +0000)]
Revert "Use private linkage for MergedGlobals variables" on Darwin.
This is a partial revert of r244615 (http://reviews.llvm.org/D11942),
which caused a major regression in debug info quality.
Turning the artificial __MergedGlobal symbols into private symbols
(l__MergedGlobal) means that the linker will not include them in the
symbol table of the final executable. Without a symbol table entry
dsymutil is not be able to process the debug info for any of the
merged globals and thus drops the debug info for all of them.
This patch is enabling the old behavior for all MachO targets while
leaving all other targets unaffected.
Greg Clayton [Fri, 11 Nov 2016 16:55:31 +0000 (16:55 +0000)]
Fix windows buildbot where warnings are errors. We had a switch statement where all enumerations were handled, but some compilers don't recognize this. Simplify the logic so that all compilers will know a return value is returned in all cases.
Greg Clayton [Fri, 11 Nov 2016 16:21:37 +0000 (16:21 +0000)]
Clean up DWARFFormValue by reducing duplicated code and removing DWARFFormValue::getFixedFormSizes()
In preparation for a follow on patch that improves DWARF parsing speed, clean up DWARFFormValue so that we have can get the fixed byte size of a form value given a DWARFUnit or given the version, address byte size and dwarf32/64.
This patch cleans up code so that everyone is using one of the new DWARFFormValue functions:
This patch changes DWARFFormValue::skipValue() to rely on the output of DWARFFormValue::getFixedByteSize(...) instead of duplicating the code in each function. This will reduce the number of changes we need to make to DWARF to fewer places in DWARFFormValue when we add support for new form.
This patch also starts to support DWARF64 so that we can get correct byte sizes for forms that vary according the DWARF 32/64.
To reduce the code duplication a new FormSizeHelper pure virtual class was created that can be created as a FormSizeHelperDWARFUnit when you have a DWARFUnit, or FormSizeHelperManual where you manually specify the DWARF version, address byte size and DWARF32/DWARF64. There is now a single implementation of a function that gets the fixed byte size (instead of two where one took a DWARFUnit and one took the DWARF version, address byte size and DWARFFormat enum) and one function to skip the form values.
Nemanja Ivanovic [Fri, 11 Nov 2016 14:41:19 +0000 (14:41 +0000)]
[PowerPC] Add vector conversion builtins to altivec.h - LLVM portion
This patch corresponds to review:
https://reviews.llvm.org/D26307
Adds all the intrinsics used for various conversion builtins that will
be added to altivec.h. These are type conversions between various types of
vectors.
projects / llvm / log