Teresa Johnson [Thu, 17 Sep 2015 20:12:00 +0000 (20:12 +0000)]
Restore "Function bitcode index in Value Symbol Table and lazy reading support"
This reverts commit r247898 (which reverted r247894).
Patch fixed to address two issues exposed by buildbots:
- unused variable warning in NDEBUG mode
- std::initializer_list lifetime issue causing test failures
Original Summary:
Support for including the function bitcode indices in the Value Symbol
Table. This requires writing the VST after the function blocks, which in
turn requires a new VST forward declaration record encoding the offset of
the full VST (which is backpatched to contain the offset after the VST
is written).
This patch also enables the lazy function reader to use the new function
indices out of the VST. This support will be used by ThinLTO as well, which
will be in a follow on patch. Backwards compatibility with older bitcode
files is maintained.
A new test is also included.
The bitcode format (used for the lazy reader as well as the upcoming
ThinLTO patches) came out of discussions with Duncan and others and is
described here:
https://drive.google.com/file/d/0B036uwnWM6RWdnBLakxmeDdOeXc/view
Diego Novillo [Thu, 17 Sep 2015 19:05:48 +0000 (19:05 +0000)]
Temporarily fix gcov failures in big-endian hosts.
This test uses a gcov file generated in a little-endian host. The gcov
reader does not allow different endianness, so the test fails on big
endian hosts.
[WinEH] Add and use hasEHPadSuccessor instead of getLandingPadSuccessor
getLandingPadSuccessor assumes that each invoke can have at most one EH
pad successor, but WinEH invokes can have more than one. Two out of
three callers of getLandingPadSuccessor don't use the returned
landingpad, so we can make them use this simple predicate instead.
Eventually we'll have to circle back and fix SplitKit.cpp so that
register allocation works. Baby steps.
Teresa Johnson [Thu, 17 Sep 2015 16:19:10 +0000 (16:19 +0000)]
Revert "Function bitcode index in Value Symbol Table and lazy reading support"
Temporarily revert to fix some buildbot issues. One is a minor issue
with a variable unused in NDEBUG mode. More concerning are some test
failures on win7 that I need to dig into.
Daniel Sanders [Thu, 17 Sep 2015 16:08:39 +0000 (16:08 +0000)]
[mips] Add assembler support for the .cprestore directive.
Summary:
This assembler directive is used in O32 PIC to restore the current function's $gp after executing JAL's. The $gp is first stored on the stack at a user-specified offset.
It has the following format: ".cprestore 8" (where 8 is the offset).
Teresa Johnson [Thu, 17 Sep 2015 15:52:30 +0000 (15:52 +0000)]
Function bitcode index in Value Symbol Table and lazy reading support
Summary:
Support for including the function bitcode indices in the Value Symbol
Table. This requires writing the VST after the function blocks, which in
turn requires a new VST forward declaration record encoding the offset of
the full VST (which is backpatched to contain the offset after the VST
is written).
This patch also enables the lazy function reader to use the new function
indices out of the VST. This support will be used by ThinLTO as well, which
will be in a follow on patch. Backwards compatibility with older bitcode
files is maintained.
A new test is also included.
The bitcode format (used for the lazy reader as well as the upcoming
ThinLTO patches) came out of discussions with Duncan and others and is
described here:
https://drive.google.com/file/d/0B036uwnWM6RWdnBLakxmeDdOeXc/view
AVX-512: shufflevector for i1 vectors <2 x i1> .. <64 x i1>
AVX-512 does not provide an instruction that shuffles mask register. So I do the following way:
Diego Novillo [Thu, 17 Sep 2015 00:17:24 +0000 (00:17 +0000)]
GCC AutoFDO profile reader - Initial support.
This adds enough machinery to support reading simple GCC AutoFDO
profiles. It now supports reading flat profiles (no function calls).
Subsequent patches will add support for:
- Inlined calls (in particular, the inline call stack is not traversed
to accumulate samples).
- Working sets and modules. These are used mostly for GCC's LIPO
optimizations, so they're not needed in LLVM atm. I'm not sure that
we will ever need them. For now, I've if0'd around the calls.
The patch also adds support in GCOV.h for gcov version V704 (generated
by GCC's profile conversion tool).
ScalarEvolution: added tmp to avoid use-after-dtor in for loop.
Summary:
For loop destroyed current instance before invoking next.
Temporary variable added to prevent use-after-dtor when invoke
destructor on current instance.
David Majnemer [Wed, 16 Sep 2015 20:42:16 +0000 (20:42 +0000)]
[WinEHPrepare] Turn terminatepad into a cleanuppad + call + cleanupret
The MSVC doesn't really support exception specifications so let's just
turn these into cleanuppads. Later, we might use terminatepad to more
efficiently encode the "noexcept"-ness of a function body.
Summary:
`signum(x)` is sometimes implemented as `(x >> 63) | (-x >>> 63)` (for
an `i64` `x`). This change adds a matcher for that pattern, and an
instcombine rule to optimize `signum(x) s< 1`.
[WinEH] Pull Adjectives and CatchObj out of the catchpad arg list
Clang now passes the adjectives as an argument to catchpad.
Getting the CatchObj working is simply a matter of threading another
static alloca through codegen, first as an alloca, then as a frame
index, and finally as a frame offset.
David Majnemer [Wed, 16 Sep 2015 18:40:24 +0000 (18:40 +0000)]
[WinEHPrepare] Refactor explicit EH preparation
Split the preparation machinery into several functions, we will want to
selectively enable/disable different parts of it for an alternative
mechanism for dealing with cross-funclet uses.
Teresa Johnson [Wed, 16 Sep 2015 18:06:45 +0000 (18:06 +0000)]
Disable the second verification run when performing LTO through
gold in NDEBUG mode.
Follow on patch for r247729 - LTO: Disable extra verify runs in release
builds.
[WinEH] Skip state numbering when no EH pads are present
Otherwise we'd try to emit the thunk that passes the LSDA to
__CxxFrameHandler3. We don't emit the LSDA if there were no landingpads,
so we'd end up with an assembler error when trying to write the COFF
object.
Dan Gohman [Wed, 16 Sep 2015 16:51:30 +0000 (16:51 +0000)]
[WebAssembly] Check in an initial CFG Stackifier pass
This pass implements a simple algorithm for conversion from CFG to
wasm's structured control flow. It doesn't yet handle multiple-entry
loops; that will be added in a future patch.
It also adds initial support for switch statements.
After D10403, we had FMF in the DAG but disabled by default. Nick reported no crashing errors after some stress testing,
so I enabled them at r243687. However, Escha soon notified us of a bug not covered by any in-tree regression tests:
if we don't propagate the flags, we may fail to CSE DAG nodes because differing FMF causes them to not match. There is
one test case in this patch to prove that point.
This patch hopes to fix or leave a 'TODO' for all of the in-tree places where we create nodes that are FMF-capable. I
did this by putting an assert in SelectionDAG.getNode() to find any FMF-capable node that was being created without FMF
( D11807 ). I then ran all regression tests and test-suite and confirmed that everything passes.
This patch exposes remaining work to get DAG FMF to be fully functional: (1) add the flags to non-binary nodes such as
FCMP, FMA and FNEG; (2) add the flags to intrinsics; (3) use the flags as conditions for transforms rather than the
current global settings.
This patch adds support for msan on aarch64-linux for both 39 and
42-bit VMA. The support is enabled by defining the
SANITIZER_AARCH64_VMA compiler flag to either 39 or 42 at build time
for both clang/llvm and compiler-rt. The default VMA is 39 bits.
Daniel Sanders [Wed, 16 Sep 2015 11:49:49 +0000 (11:49 +0000)]
llvm-mc-fuzzer: A fuzzing tool for the MC layer.
Summary:
Only the disassembler is supported in this patch but it has already found a few
issues in the Mips disassembler (mostly invalid instructions being successfully
disassembled).
When trying emit a stack adjustments using pops, frame lowering selects an
arbitrary free GPR. It should always select one from an appropriate class...
This fixes PR24649.
[X86] Fix emitEpilogue() to make less assumptions about pops
This is the mirror image of r242395.
When X86FrameLowering::emitEpilogue() looks for where to insert the %esp addition that
deallocates stack space used for local allocations, it assumes that any sequence of pop
instructions from function exit backwards consists purely of restoring callee-save registers.
This may be false, since from some point backward, the pops may be clean-up of stack space
allocated for arguments to a call.
Mehdi Amini [Wed, 16 Sep 2015 05:34:32 +0000 (05:34 +0000)]
Make the default triple optional by allowing an empty string
When building LLVM as a (potentially dynamic) library that can be linked against
by multiple compilers, the default triple is not really meaningful.
We allow to explicitely set it to an empty string when configuring LLVM.
In this case, said "target independent" tests in the test suite that are using
the default triple are disabled by matching the newly available feature
"default_triple".
[Unroll] Fix a bug in UnrolledInstAnalyzer::visitLoad.
We only checked that a global is initialized with constants, which is
incorrect. We should be checking that GlobalVariable *is* a constant,
not just initialized with it.
Sanjoy Das [Tue, 15 Sep 2015 23:45:39 +0000 (23:45 +0000)]
[IndVars] Fix PR24783.
In `IndVarSimplify::ExpandSCEVIfNeeded`,
`SCEVExpander::findExistingExpansion` may return an `llvm::Value` that
differs in type from the SCEV it was asked to find an expansion for (but
computes the same value). In such cases, we fall back on
`expandCodeFor`; and rely on LLVM to CSE the two equivalent
expressions (different only by a no-op cast) into a single computation.
I tried a few other approaches to fixing PR24783, all of which turned
out to be more complex than this current version:
1. Move the `ExpandSCEVIfNeeded` logic into `expandCodeFor`. This got
problematic because currently we do not pass in the `Loop *` into
`expandCodeFor`. Changing the interface to do this is a more
invasive change, and really does not make much semantic sense unless
the SCEV being passed in is an add recurrence.
There is also the problem of `expandCodeFor` being used in places
other than `indvars` -- there may be performance / correctness
issues elsewhere if `expandCodeFor` is moved from always generating
IR from scratch to cache-like model.
2. Have `findExistingExpansion` only return expression with the correct
type. This would make `isHighCostExpansionHelper` and thus
`isHighCostExpansion` more conservative than necessary.
3. Insert casts on the value returned by `findExistingExpansion` if
needed using `InsertNoopCastOfTo`. This is complicated because
`InsertNoopCastOfTo` depends on internal state of its
`SCEVExpander` (specifically `Builder.GetInserPoint()`), and this
may not be set up when `ExpandSCEVIfNeeded` is called.
4. Manually insert casts on the value returned by
`findExistingExpansion` if needed using `InsertNoopCastOfTo` via
`CastInst::Create`. This is probably workable, but figuring out the
location where the cast instruction needs to be inserted has enough
edge cases (arguments, constants, invokes, LCSSA must be preserved)
makes me feel what I have right now is simplest solution.
[ASan] Don't instrument globals in .preinit_array/.init_array/.fini_array
These sections contain pointers to function that should be invoked
during startup/shutdown by __libc_csu_init and __libc_csu_fini.
Instrumenting these globals will append redzone to them, which will be
filled with zeroes. This will cause null pointer dereference at runtime.
Merge ASan regression tests for globals that should be ignored by
instrumentation pass.
The verifier currently runs three times in LTO: (1) after parsing, (2)
at the beginning of the optimization pipeline, and (3) at the end of it.
The first run is important, since we're not sure where the bitcode comes
from and it's nice to validate it, but in release builds the extra runs
aren't appropriate.
This commit:
- Allows these runs to be disabled in LTOCodeGenerator.
- Adds command-line options to llvm-lto.
- Adds command-line options to libLTO.dylib, and disables the verifier
by default in release builds (based on NDEBUG).
This shaves about 3.5% off the runtime of ld64 when linking
verify-uselistorder with -flto -g.
Revert "Clean up: Refactoring the hardcoded value of 6 for FindAvailableLoadedValue()'s parameter MaxInstsToScan." for preliminary community discussion (See. D12886)
[ShrinkWrapping] Fix an infinite loop while looking for restore point.
This may happen when the input program itself contains an infinite loop with no
exit block. In that case, we would fail to find a block post-dominating the loop
such that this block is outside of the loop.
This fixes PR24823.
Working on reducing the test case.
Broaden optimization of fcmp ([us]itofp x, constant) by instcombine.
The patch extends the optimization to cases where the constant's
magnitude is so small or large that the rounding of the conversion
is irrelevant. The "so small" case includes negative zero.
LazuValueInfo can prove that value is nonnull based on the context information.
Make use of this ability to infer nonnull attributes for the call arguments.
[mips][microMIPS] Fix an issue with disassembling lwm32 instruction
Fixed microMIPS disassembler crash on test case generated by llvm-mc-fuzzer.
Differential Revision: http://reviews.llvm.org/D12881
[SystemZ] Fix assertion failure in tryBuildVectorShuffle
Under certain circumstances, tryBuildVectorShuffle would attempt to
create a BUILD_VECTOR node with an invalid combination of types.
This happened when one of the components of the original BUILD_VECTOR
was itself a TRUNCATE node. That TRUNCATE was stripped off during
intermediate processing to simplify code, but when adding the node
back to the result vector, we still need it to get the type right.
Daniel Sanders [Tue, 15 Sep 2015 14:08:28 +0000 (14:08 +0000)]
Re-commit r247683: Replace Triple with a new TargetTuple in MCTargetDesc/* and related. NFC.
Summary:
This is the first patch in the series to migrate Triple's (which are ambiguous)
to TargetTuple's (which aren't).
For the moment, TargetTuple simply passes all requests to the Triple object it
holds. Once it has replaced Triple, it will start to implement the interface in
a more suitable way.
This change makes some changes to the public C++ API. In particular,
InitMCSubtargetInfo(), createMCRelocationInfo(), and createMCSymbolizer()
now take TargetTuples instead of Triples. The other public C++ API's have
been left as-is for the moment to reduce patch size.
This commit also contains a trivial patch to clang to account for the C++ API
change. Thanks go to Pavel Labath for fixing LLDB for me.
Daniel Sanders [Tue, 15 Sep 2015 13:17:40 +0000 (13:17 +0000)]
Replace Triple with a new TargetTuple in MCTargetDesc/* and related. NFC.
Summary:
This is the first patch in the series to migrate Triple's (which are ambiguous)
to TargetTuple's (which aren't).
For the moment, TargetTuple simply passes all requests to the Triple object it
holds. Once it has replaced Triple, it will start to implement the interface in
a more suitable way.
This change makes some changes to the public C++ API. In particular,
InitMCSubtargetInfo(), createMCRelocationInfo(), and createMCSymbolizer()
now take TargetTuples instead of Triples. The other public C++ API's have
been left as-is for the moment to reduce patch size.
This commit also contains a trivial patch to clang to account for the C++ API
change.
Sanjoy Das [Tue, 15 Sep 2015 01:42:48 +0000 (01:42 +0000)]
[PlaceSafepoints] Make the width of a counted loop settable.
Summary:
This change lets a `PlaceSafepoints` client change how wide the trip
count of a loop has to be for the loop to be considerd "counted", via
`CountedLoopTripWidth`. It also removes the boolean `SkipCounted` flag
and the `upperTripBound` constant -- we can get the old behavior of
`SkipCounted` == `false` by setting `CountedLoopTripWidth` to `13` (2 ^
13 == 8192).
Chris Bieneman [Mon, 14 Sep 2015 23:09:06 +0000 (23:09 +0000)]
[CMake] Refactor and cleanup generating and installing symlinks to tools.
Summary: This change generalizes symlink generation and makes symlinks to tools obey LLVM_TOOLCHAIN_TOOLS. It makes it so that if you exclude llvm-ar from LLVM_TOOLCHAIN_TOOLS you don't end up with broken symlinks to llvm-lib and llvm-ranlib in your install.
Adrian Prantl [Mon, 14 Sep 2015 22:10:22 +0000 (22:10 +0000)]
DwarfDebug: Emit dwo_id+dwo_name for DICompileUnits that provide a dwoId.
For module debugging clang emits prefabricated skeleton compile units
that can be recognized by a nonzero dwoId.
David Blaikie [Mon, 14 Sep 2015 21:47:27 +0000 (21:47 +0000)]
[opaque pointer types] Add an explicit value type to GlobalObject
This is needed by all GlobalObjects (GlobalAlias, Function,
GlobalVariable), see the GlobalObject::getValueType which is used in
many places. If at some point that can be removed, then we can remove
this member.
Dan Gohman [Mon, 14 Sep 2015 19:44:29 +0000 (19:44 +0000)]
Make AsmPrinter::EmitJumpTableInfo virtual
Targets that have non-traditional jump table mechanisms may need to do
something substantially different for jump tables than what
AsmPrinter::EmitJumpTableInfo does. This patch makes that function
virtual so that targets can override it.
- There are no duplicate registers in LiveRegs list we are copying from
and so we do not need to sort the registers.
- Simply use SmallVector::apend instead of a loop between begin() and end()
with push_back().
Chen Li [Mon, 14 Sep 2015 18:10:43 +0000 (18:10 +0000)]
[InstCombineCalls] Use isKnownNonNullAt() to check nullness of passing arguments at callsite
Summary: This patch replaces isKnownNonNull() with isKnownNonNullAt() when checking nullness of passing arguments at callsite. In this way it can handle cases where the argument does not have nonnull attribute but has a dominating null check from the CFG. It also adds assertions in isKnownNonNull() and isKnownNonNullFromDominatingCondition() to make sure the value checked is pointer type (as defined in LLVM document). These assertions might trip failures in things which are not covered under llvm/test, but fixes should be pretty obvious.
David Blaikie [Mon, 14 Sep 2015 18:01:59 +0000 (18:01 +0000)]
Revert "[opaque pointer type] Pass GlobalAlias the actual pointer type rather than decomposing it into pointee type + address space"
This was a flawed change - it just caused the getElementType call to be
deferred until later, when we really need to remove it. Now that the IR
for GlobalAliases has been updated, the root cause is addressed that way
instead and this change is no longer needed (and in fact gets in the way
- because we want to pass the pointee type directly down further).
Follow up patches to push this through GlobalValue, bitcode format, etc,
will come along soon.
Davide Italiano [Mon, 14 Sep 2015 17:10:01 +0000 (17:10 +0000)]
[llvm-mc] Better error handling in ENOENT case + test.
This is a follow up to r247518.
As a general note, I think we could do a much better job testing for
error conditions in tools. I already anticipated in a previous mail,
but while implementing this I noticed that the code coverage we have
for error checking is pretty low. I can arbitrarily remove checks from
several tools and the suite still passes.
projects / llvm / log