platform/upstream/llvm.git
7 years agoProfileData: clean up some stale declarations (NFC)
Saleem Abdulrasool [Mon, 24 Apr 2017 21:05:05 +0000 (21:05 +0000)]
ProfileData: clean up some stale declarations (NFC)

These were removed in SVN r300381.  Remove the declarations.

llvm-svn: 301252

7 years ago[analyzer] Improve subscripting null arrays for catching null dereferences.
Artem Dergachev [Mon, 24 Apr 2017 20:55:07 +0000 (20:55 +0000)]
[analyzer] Improve subscripting null arrays for catching null dereferences.

Array-to-pointer cast now works correctly when the pointer to the array
is concrete, eg. null, which allows further symbolic calculations involving
such values.

Inlined defensive checks are now detected correctly when the resulting null
symbol is being array-subscripted before dereference.

Differential Revision: https://reviews.llvm.org/D32291

llvm-svn: 301251

7 years agoRemove a dead field. NFC.
Vedant Kumar [Mon, 24 Apr 2017 20:54:36 +0000 (20:54 +0000)]
Remove a dead field. NFC.

Suggested by Adam Folwarczny!

llvm-svn: 301250

7 years ago[Coverage] Avoid null deref in skipRegionMappingForDecl (fixes PR32761)
Vedant Kumar [Mon, 24 Apr 2017 20:52:04 +0000 (20:52 +0000)]
[Coverage] Avoid null deref in skipRegionMappingForDecl (fixes PR32761)

Patch by Adam Folwarczny!

Differential Revision: https://reviews.llvm.org/D32406

llvm-svn: 301249

7 years agoSmall addition on how to add a builder.
Galina Kistanova [Mon, 24 Apr 2017 20:48:40 +0000 (20:48 +0000)]
Small addition on how to add a builder.

llvm-svn: 301248

7 years ago[AMDGPU][mc][tests][NFC] Bulk ISA tests: update for Gfx7/Gfx8, add for Gfx9.
Artem Tamazov [Mon, 24 Apr 2017 20:42:27 +0000 (20:42 +0000)]
[AMDGPU][mc][tests][NFC] Bulk ISA tests: update for Gfx7/Gfx8, add for Gfx9.

llvm-svn: 301247

7 years agoSpecify a target explicitly in libfuzzer driver flag test
George Karpenkov [Mon, 24 Apr 2017 20:38:56 +0000 (20:38 +0000)]
Specify a target explicitly in libfuzzer driver flag test

llvm-svn: 301246

7 years ago[Bitcode] Refactor attribute group writing to avoid getSlotAttributes
Reid Kleckner [Mon, 24 Apr 2017 20:38:30 +0000 (20:38 +0000)]
[Bitcode] Refactor attribute group writing to avoid getSlotAttributes

Summary:
That API creates a temporary AttributeList to carry an index and a
single AttributeSet. We need to carry the index in addition to the set,
because that is how attribute groups are currently encoded.

NFC

Reviewers: pcc

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D32262

llvm-svn: 301245

7 years agoUpdate profile during memory instrinsic optimization
Teresa Johnson [Mon, 24 Apr 2017 20:30:42 +0000 (20:30 +0000)]
Update profile during memory instrinsic optimization

Summary:
Ensure that the new merge BB (which contains the rest of the original BB
after the mem op being optimized) gets a profile frequency, in case
there are additional mem ops later in the BB. Otherwise they get skipped
as the merge BB looks cold.

Reviewers: davidxl, xur

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D32447

llvm-svn: 301244

7 years ago[asan] Use posix strerror_r interceptor on android.
Evgeniy Stepanov [Mon, 24 Apr 2017 20:25:39 +0000 (20:25 +0000)]
[asan] Use posix strerror_r interceptor on android.

This fixes a regression in r297315.

llvm-svn: 301243

7 years agoRevert "StructurizeCFG: Directly invert cmp instructions"
Matt Arsenault [Mon, 24 Apr 2017 20:25:01 +0000 (20:25 +0000)]
Revert "StructurizeCFG: Directly invert cmp instructions"

This reverts commit r300732. This breaks a few tests.
I think the problem is related to adding more uses of
the condition that don't yet exist at this point.

llvm-svn: 301242

7 years ago[LoopUnroll] Remove spurious newline.
Davide Italiano [Mon, 24 Apr 2017 20:17:38 +0000 (20:17 +0000)]
[LoopUnroll] Remove spurious newline.

Eli pointed out in the review, but I didn't squash the two commits
correctly. Pointy-hat to me.

llvm-svn: 301241

7 years agoRevert "Refactor DynamicLibrary so searching for a symbol will have a defined order"
Frederich Munch [Mon, 24 Apr 2017 20:16:01 +0000 (20:16 +0000)]
Revert "Refactor DynamicLibrary so searching for a symbol will have a defined order"
The i686-mingw32-RA-on-linux bot is still having errors.

This reverts commit r301236.

llvm-svn: 301240

7 years ago[LoopUnroll] Don't try to unroll non canonical loops.
Davide Italiano [Mon, 24 Apr 2017 20:14:11 +0000 (20:14 +0000)]
[LoopUnroll] Don't try to unroll non canonical loops.

The current Loop Unroll implementation works with loops having a
single latch that contains a conditional branch to a block outside
the loop (the other successor is, by defition of latch, the header).
If this precondition doesn't hold, avoid unrolling the loop as
the code is not ready to handle such circumstances.

Differential Revision:  https://reviews.llvm.org/D32261

llvm-svn: 301239

7 years ago[LIR] Obey non-integral pointer semantics
Sanjoy Das [Mon, 24 Apr 2017 20:12:10 +0000 (20:12 +0000)]
[LIR] Obey non-integral pointer semantics

Summary: See http://llvm.org/docs/LangRef.html#non-integral-pointer-type

Reviewers: haicheng

Reviewed By: haicheng

Subscribers: mcrosier, mzolotukhin, llvm-commits

Differential Revision: https://reviews.llvm.org/D32196

llvm-svn: 301238

7 years agoAvoid unnecessary copies in some for loops
Saleem Abdulrasool [Mon, 24 Apr 2017 20:01:03 +0000 (20:01 +0000)]
Avoid unnecessary copies in some for loops

Use constant references rather than `const auto` which will cause the
copy constructor.  These particular cases cause issues for the swift
compiler.

llvm-svn: 301237

7 years agoRefactor DynamicLibrary so searching for a symbol will have a defined order and
Frederich Munch [Mon, 24 Apr 2017 19:55:16 +0000 (19:55 +0000)]
Refactor DynamicLibrary so searching for a symbol will have a defined order and
libraries are properly unloaded when llvm_shutdown is called.

Summary:
This was mostly affecting usage of the JIT, where storing the library handles in
a set made iteration unordered/undefined. This lead to disagreement between the
JIT and native code as to what the address and implementation of particularly on
Windows with stdlib functions:

JIT: putenv_s("TEST", "VALUE") // called msvcrt.dll, putenv_s
JIT: getenv("TEST") -> "VALUE" // called msvcrt.dll, getenv
Native: getenv("TEST") -> NULL // called ucrt.dll, getenv

Also fixed is the issue of DynamicLibrary::getPermanentLibrary(0,0) on Windows
not giving priority to the process' symbols as it did on Unix.

Reviewers: chapuni, v.g.vassilev, lhames

Reviewed By: lhames

Subscribers: danalbert, srhines, mgorny, vsk, llvm-commits

Differential Revision: https://reviews.llvm.org/D30107

llvm-svn: 301236

7 years ago[cfi] Disable ThinLTO + CFI tests on Windows.
Evgeniy Stepanov [Mon, 24 Apr 2017 19:52:51 +0000 (19:52 +0000)]
[cfi] Disable ThinLTO + CFI tests on Windows.

PR32770.

llvm-svn: 301235

7 years agoMove value type list from TargetRegisterClass to TargetRegisterInfo
Krzysztof Parzyszek [Mon, 24 Apr 2017 19:51:12 +0000 (19:51 +0000)]
Move value type list from TargetRegisterClass to TargetRegisterInfo

Differential Revision: https://reviews.llvm.org/D31937

llvm-svn: 301234

7 years agoRevert r301223
Carlo Bertolli [Mon, 24 Apr 2017 19:50:35 +0000 (19:50 +0000)]
Revert r301223

llvm-svn: 301233

7 years agoRevert r301231: Accidentally committed stale files
Krzysztof Parzyszek [Mon, 24 Apr 2017 19:48:51 +0000 (19:48 +0000)]
Revert r301231: Accidentally committed stale files

I forgot to commit local changes before commit.

llvm-svn: 301232

7 years agoMove value type list from TargetRegisterClass to TargetRegisterInfo
Krzysztof Parzyszek [Mon, 24 Apr 2017 19:43:45 +0000 (19:43 +0000)]
Move value type list from TargetRegisterClass to TargetRegisterInfo

Differential Revision: https://reviews.llvm.org/D31937

llvm-svn: 301231

7 years agoAMDGPU: Select scratch mubuf offsets when pointer is a constant
Matt Arsenault [Mon, 24 Apr 2017 19:40:59 +0000 (19:40 +0000)]
AMDGPU: Select scratch mubuf offsets when pointer is a constant

In call sequence setups, there may not be a frame index base
and the pointer is a constant offset from the frame
pointer / scratch wave offset register.

llvm-svn: 301230

7 years agoAMDGPU: Set StackGrowsUp in MCAsmInfo
Matt Arsenault [Mon, 24 Apr 2017 19:40:51 +0000 (19:40 +0000)]
AMDGPU: Set StackGrowsUp in MCAsmInfo

Not sure what this does though.

llvm-svn: 301229

7 years ago[AMDGPU] Merge M0 initializations
Stanislav Mekhanoshin [Mon, 24 Apr 2017 19:37:54 +0000 (19:37 +0000)]
[AMDGPU] Merge M0 initializations

Merges equivalent initializations of M0 and hoists them into a common
dominator block. Technically the same code can be used with any
register, physical or virtual.

Differential Revision: https://reviews.llvm.org/D32279

llvm-svn: 301228

7 years agoHandle invariant.group.barrier in BasicAA
Piotr Padlewski [Mon, 24 Apr 2017 19:37:17 +0000 (19:37 +0000)]
Handle invariant.group.barrier in BasicAA

Summary:
llvm.invariant.group.barrier returns pointer that mustalias
pointer it takes. It can't be marked with `returned` attribute,
because it would be remove easily. The other reason is that
only Alias Analysis can know about this, because if any other
pass would know it, then the result would be replaced with it's
argument, which would be invalid.

We can think about returned pointer as something that mustalias, but
it doesn't have to be bitwise the same as the argument.

Reviewers: dberlin, chandlerc, hfinkel, sanjoy

Subscribers: reames, nlewycky, rsmith, anna, amharc

Differential Revision: https://reviews.llvm.org/D31585

llvm-svn: 301227

7 years ago[asan] Let the frontend disable gc-sections optimization for asan globals.
Evgeniy Stepanov [Mon, 24 Apr 2017 19:34:13 +0000 (19:34 +0000)]
[asan] Let the frontend disable gc-sections optimization for asan globals.

Also extend -asan-globals-live-support flag to all binary formats.

llvm-svn: 301226

7 years ago[asan] Disable ASan global-GC depending on the target and compiler flags.
Evgeniy Stepanov [Mon, 24 Apr 2017 19:34:12 +0000 (19:34 +0000)]
[asan] Disable ASan global-GC depending on the target and compiler flags.

llvm-svn: 301225

7 years ago[analyzer] Improve suppression for inlined defensive checks before operator &.
Artem Dergachev [Mon, 24 Apr 2017 19:30:33 +0000 (19:30 +0000)]
[analyzer] Improve suppression for inlined defensive checks before operator &.

Null dereferences are suppressed if the lvalue was constrained to 0 for the
first time inside a sub-function that was inlined during analysis, because
such constraint is a valid defensive check that does not, by itself,
indicate that null pointer case is anyhow special for the caller.

If further operations on the lvalue are performed, the symbolic lvalue is
collapsed to concrete null pointer, and we need to track where does the null
pointer come from.

Improve such tracking for lvalue operations involving operator &.

rdar://problem/27876009

Differential Revision: https://reviews.llvm.org/D31982

llvm-svn: 301224

7 years ago[OpenMP] Initial implementation of code generation for pragma 'distribute parallel...
Carlo Bertolli [Mon, 24 Apr 2017 19:26:11 +0000 (19:26 +0000)]
[OpenMP] Initial implementation of code generation for pragma 'distribute parallel for' on host

https://reviews.llvm.org/D29508

This patch makes the following additions:

1. It abstracts away loop bound generation code from procedures associated with pragma 'for' and loops in general, in such a way that the same procedures can be used for 'distribute parallel for' without the need for a full re-implementation.
2. It implements code generation for 'distribute parallel for' and adds regression tests. It includes tests for clauses.

It is important to notice that most of the clauses are implemented as part of existing procedures. For instance, firstprivate is already implemented for 'distribute' and 'for' as separate pragmas. As the implementation of 'distribute parallel for' is based on the same procedures, then we automatically obtain implementation for such clauses without the need to add new code. However, this requires regression tests that verify correctness of produced code.

Looking forward to comments.

llvm-svn: 301223

7 years ago[SimplifyCFG] Fix for non-determinism in codegen
Mandeep Singh Grang [Mon, 24 Apr 2017 19:20:45 +0000 (19:20 +0000)]
[SimplifyCFG] Fix for non-determinism in codegen

Summary: This patch fixes issues in codegen uncovered due to https://reviews.llvm.org/D26718

Reviewers: majnemer, chenli, davide

Reviewed By: davide

Subscribers: davide, arsenm, llvm-commits

Differential Revision: https://reviews.llvm.org/D26726

llvm-svn: 301222

7 years agoMove size and alignment information of regclass to TargetRegisterInfo
Krzysztof Parzyszek [Mon, 24 Apr 2017 18:55:33 +0000 (18:55 +0000)]
Move size and alignment information of regclass to TargetRegisterInfo

1. RegisterClass::getSize() is split into two functions:
   - TargetRegisterInfo::getRegSizeInBits(const TargetRegisterClass &RC) const;
   - TargetRegisterInfo::getSpillSize(const TargetRegisterClass &RC) const;
2. RegisterClass::getAlignment() is replaced by:
   - TargetRegisterInfo::getSpillAlignment(const TargetRegisterClass &RC) const;

This will allow making those values depend on subtarget features in the
future.

Differential Revision: https://reviews.llvm.org/D31783

llvm-svn: 301221

7 years agoDon't test setting sticky bits on files for modern BSDs
Dimitry Andric [Mon, 24 Apr 2017 18:54:48 +0000 (18:54 +0000)]
Don't test setting sticky bits on files for modern BSDs

Summary: In rL297945, jhenderson added methods for setting permissions
to sys::fs, but some of the unittests that attempt to set sticky bits
(01000) on files fail on modern BSDs, such as FreeBSD, NetBSD and
OpenBSD.  This is because those systems do not allow regular users to
set sticky bits on files, only on directories.  Fix it by disabling
these particular tests on modern BSDs.

Reviewers: emaste, brad, jhenderson

Reviewed By: jhenderson

Subscribers: joerg, krytarowski, llvm-commits

Differential Revision: https://reviews.llvm.org/D32120

llvm-svn: 301220

7 years agoDon't emit CFI instructions at the end of a function
Adrian Prantl [Mon, 24 Apr 2017 18:45:59 +0000 (18:45 +0000)]
Don't emit CFI instructions at the end of a function

When functions are terminated by unreachable instructions, the last
instruction might trigger a CFI instruction to be generated. However,
emitting it would be be illegal since the function (and thus the FDE
the CFI is in) has already ended with the previous instruction.

Darwin's dwarfdump --verify --eh-frame complains about this and the
specification supports this.
Relevant bits from the DWARF 5 standard (6.4 Call Frame Information):

"[The] address_range [field in an FDE]: The number of bytes of
 program instructions described by this entry."

"Row creation instructions: [...]
 The new location value is always greater than the current one."
The first quotation implies that a CFI cannot describe a target
address outside of the enclosing FDE's range.

rdar://problem/26244988

Differential Revision: https://reviews.llvm.org/D32246

llvm-svn: 301219

7 years agoPragma: Fix DebugOverflowStack() resulting in endless loop.
Matthias Braun [Mon, 24 Apr 2017 18:41:00 +0000 (18:41 +0000)]
Pragma: Fix DebugOverflowStack() resulting in endless loop.

Drive-by fix (noticed while working on https://reviews.llvm.org/D32205):
DebugOverflowStack() is supposed to provoke a stack overflow, however
LLVM was smart enough to use the red-zone and fold the load into a tail
jump on x86_64 optimizing this to an endless loop instead of a stack
overflow.

llvm-svn: 301218

7 years agoUpdates documentation for a syntax sugar libfuzzer flag,
George Karpenkov [Mon, 24 Apr 2017 18:39:52 +0000 (18:39 +0000)]
Updates documentation for a syntax sugar libfuzzer flag,

as implemented in https://reviews.llvm.org/D32193

llvm-svn: 301217

7 years agoRemove erroneous driver test for -fsanitize=fuzzer flag
George Karpenkov [Mon, 24 Apr 2017 18:36:31 +0000 (18:36 +0000)]
Remove erroneous driver test for -fsanitize=fuzzer flag

libfuzzer is not available on all platforms, and hence we can not always
rely that it was compiled.

llvm-svn: 301216

7 years agoCodeGen: Add a hook for getFenceOperandTy
Yaxun Liu [Mon, 24 Apr 2017 18:26:27 +0000 (18:26 +0000)]
CodeGen: Add a hook for getFenceOperandTy

Currently the operand type for ATOMIC_FENCE assumes value type of a pointer in address space 0.
This is fine for most targets. However for amdgcn target, the size of pointer in address space 0
depends on triple environment. For amdgiz environment, it is 64 bit but for other environment it is
32 bit. On the other hand, amdgcn target expects 32 bit fence operands independent of the target
triple environment. Therefore a hook is need in target lowering for getting the fence operand type.

This patch has no effect on targets other than amdgcn.

Differential Revision: https://reviews.llvm.org/D32186

llvm-svn: 301215

7 years agoRevert "Compute safety information in a much finer granularity."
Evgeniy Stepanov [Mon, 24 Apr 2017 18:25:07 +0000 (18:25 +0000)]
Revert "Compute safety information in a much finer granularity."

Use-after-free in llvm::isGuaranteedToExecute.

llvm-svn: 301214

7 years ago[InstSimplify] move (A & ~B) | (A ^ B) -> (A ^ B) from InstCombine
Sanjay Patel [Mon, 24 Apr 2017 18:24:36 +0000 (18:24 +0000)]
[InstSimplify] move (A & ~B) | (A ^ B) -> (A ^ B) from InstCombine

This is a straight cut and paste, but there's a bigger problem: if this
fold exists for simplifyOr, there should be a DeMorganized version for
simplifyAnd. But more than that, we have a patchwork of ad hoc logic
optimizations in InstCombine. There should be some structure to ensure
that we're not missing sibling folds across and/or/xor.

llvm-svn: 301213

7 years agoFlag -fsanitize=fuzzer to enable libfuzzer
George Karpenkov [Mon, 24 Apr 2017 18:23:24 +0000 (18:23 +0000)]
Flag -fsanitize=fuzzer to enable libfuzzer

Previously, adding libfuzzer to a project was a multi-step procedure,
involving libfuzzer compilation, linking the library, and specifying
coverage flags.
With this change,libfuzzer can be enabled by adding a single
-fsanitize=fuzzer flag instead.

llvm-svn: 301212

7 years agoX86RegisterInfo: eliminateFrameIndex: Avoid code duplication; NFC
Matthias Braun [Mon, 24 Apr 2017 18:15:00 +0000 (18:15 +0000)]
X86RegisterInfo: eliminateFrameIndex: Avoid code duplication; NFC

Re-Commit of r300922 and r300923 with less aggressive assert (see
discussion at the end of https://reviews.llvm.org/D32205)

X86RegisterInfo::eliminateFrameIndex() and
X86FrameLowering::getFrameIndexReference() both had logic to compute the
base register. This consolidates the code.

Also use MachineInstr::isReturn instead of manually enumerating tail
call instructions (return instructions were not included in the previous
list because they never reference frame indexes).

Differential Revision: https://reviews.llvm.org/D32206

llvm-svn: 301211

7 years agoUse DW_OP_stack_value when reconstructing variable values with arithmetic.
Adrian Prantl [Mon, 24 Apr 2017 18:11:42 +0000 (18:11 +0000)]
Use DW_OP_stack_value when reconstructing variable values with arithmetic.

When the location description of a source variable involves arithmetic
on the value itself, it needs to be marked with DW_OP_stack_value since it
is not describing the variable's location, but rather its value.

This is a follow-up to r297971 and fixes the source testcase quoted in
the comment in debuginfo-dce.ll.

rdar://problem/30725338

This reapplies r301093 without modifications.

llvm-svn: 301210

7 years agoAdd a testcase for DIExpression(DW_OP_stack_value)
Adrian Prantl [Mon, 24 Apr 2017 18:11:38 +0000 (18:11 +0000)]
Add a testcase for DIExpression(DW_OP_stack_value)
and relax the assertion that prohibited its emission.

This fixes the assertion failure uncovered by r301093.

llvm-svn: 301209

7 years agoAMDGPU: Add StackPtr and FramePtr registers to MFI
Matt Arsenault [Mon, 24 Apr 2017 18:05:16 +0000 (18:05 +0000)]
AMDGPU: Add StackPtr and FramePtr registers to MFI

These will be necessary for setting up call sequences.

llvm-svn: 301208

7 years ago[ELF] Account for R_386_TLS_LDO_32 addend
Shoaib Meenai [Mon, 24 Apr 2017 18:02:11 +0000 (18:02 +0000)]
[ELF] Account for R_386_TLS_LDO_32 addend

This relocation type has an implicit addend. Account for it when
processing the relocation. Add an offset to an existing test to ensure
it gets processed correctly.

Fixes PR32634.

Differential Revision: https://reviews.llvm.org/D32336

llvm-svn: 301207

7 years agoAMDGPU: Move trap lowering to DAG
Matt Arsenault [Mon, 24 Apr 2017 17:49:13 +0000 (17:49 +0000)]
AMDGPU: Move trap lowering to DAG

Fixes traps in any block besides the entry block,
and fixes depending on a live-in physical register
by using a virtual register copy.

Also happens to stop emitting a nop in the case
debug trap is not supported.

llvm-svn: 301206

7 years ago[DomPrinter] Add a way to programmatically dump a dot representation.
Davide Italiano [Mon, 24 Apr 2017 17:48:44 +0000 (17:48 +0000)]
[DomPrinter] Add a way to programmatically dump a dot representation.

Differential Revision:  https://reviews.llvm.org/D32145

llvm-svn: 301205

7 years ago[llvm-pdbdump] Merge functionality of graphical and text dumpers.
Zachary Turner [Mon, 24 Apr 2017 17:47:52 +0000 (17:47 +0000)]
[llvm-pdbdump] Merge functionality of graphical and text dumpers.

The *real* difference between these two was that

a) The "graphical" dumper could recurse, while the text one could
   not.
b) The "text" dumper could display nested types and functions,
   while the graphical one could not.

Merge these two so that there is only one dumper that can recurse
arbitrarily deep and optionally display nested types or not.

llvm-svn: 301204

7 years ago[llvm-pdbdump] Re-write the record layout code to be more resilient.
Zachary Turner [Mon, 24 Apr 2017 17:47:24 +0000 (17:47 +0000)]
[llvm-pdbdump] Re-write the record layout code to be more resilient.

This reworks the way virtual bases are handled, and also the way
padding is detected across multiple levels of aggregates, producing
a much more accurate result.

llvm-svn: 301203

7 years ago[Docs] Correct the path to the clang-format-diff.py script to include the clang-forma...
Craig Topper [Mon, 24 Apr 2017 17:39:35 +0000 (17:39 +0000)]
[Docs] Correct the path to the clang-format-diff.py script to include the clang-format directory.

llvm-svn: 301202

7 years ago[APInt] Simplify the zext and sext methods
Craig Topper [Mon, 24 Apr 2017 17:37:10 +0000 (17:37 +0000)]
[APInt] Simplify the zext and sext methods

This replaces a hand written copy loop with a call to memcpy for both zext and sext.

For sext, it replaces multiple if/else blocks propagating sign information forward. Now we just do a copy, a sign extension on the last copied word, a memset, and clearUnusedBits.

Differential Revision: https://reviews.llvm.org/D32417

llvm-svn: 301201

7 years agoTesting commit credentials
George Karpenkov [Mon, 24 Apr 2017 17:28:32 +0000 (17:28 +0000)]
Testing commit credentials

llvm-svn: 301200

7 years agoInstCombine: Fix assert when reassociating fsub with undef
Matt Arsenault [Mon, 24 Apr 2017 17:24:37 +0000 (17:24 +0000)]
InstCombine: Fix assert when reassociating fsub with undef

There is logic to track the expected number of instructions
produced. It thought in this case an instruction would
be necessary to negate the result, but here it folded
into a ConstantExpr fneg when the non-undef value operand
was cancelled out by the second fsub.

I'm not sure why we don't fold constant FP ops with undef currently,
but I think that would also avoid this problem.

llvm-svn: 301199

7 years ago[APInt] Add ashrInPlace method and rewrite ashr to make a copy and then call ashrInPlace.
Craig Topper [Mon, 24 Apr 2017 17:18:47 +0000 (17:18 +0000)]
[APInt] Add ashrInPlace method and rewrite ashr to make a copy and then call ashrInPlace.

This patch adds an in place version of ashr to match lshr and shl which were recently added.

I've tried to make this similar to the lshr code with additions to handle the sign extension. I've also tried to do this with less if checks than the current ashr code by sign extending the original result to a word boundary before doing any of the shifting. This removes a lot of the complexity of determining where to fill in sign bits after the shifting.

Differential Revision: https://reviews.llvm.org/D32415

llvm-svn: 301198

7 years agoAMDGPU: Move v_readlane lane select from VGPR to SGPR
Nicolai Haehnle [Mon, 24 Apr 2017 17:17:36 +0000 (17:17 +0000)]
AMDGPU: Move v_readlane lane select from VGPR to SGPR

Summary:
Fix a compiler bug when the lane select happens to end up in a VGPR.

Clarify the semantic of the corresponding intrinsic to be that of
the corresponding GLSL: the lane select must be uniform across a
wave front, otherwise results are undefined.

Reviewers: arsenm

Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits

Differential Revision: https://reviews.llvm.org/D32343

llvm-svn: 301197

7 years agoCompute safety information in a much finer granularity.
Xin Tong [Mon, 24 Apr 2017 17:12:22 +0000 (17:12 +0000)]
Compute safety information in a much finer granularity.

Summary:
Instead of keeping a variable indicating whether there are early exits
in the loop.  We keep all the early exits. This improves LICM's ability to
move instructions out of the loop based on is-guaranteed-to-execute.

I am going to update compilation time as well soon.

Reviewers: hfinkel, sanjoy, efriedma, mkuper

Reviewed By: hfinkel

Subscribers: llvm-commits, mzolotukhin

Differential Revision: https://reviews.llvm.org/D32433

llvm-svn: 301196

7 years agoInstCombine/AMDGPU: Fix constant folding of llvm.amdgcn.{icmp,fcmp}
Nicolai Haehnle [Mon, 24 Apr 2017 17:08:43 +0000 (17:08 +0000)]
InstCombine/AMDGPU: Fix constant folding of llvm.amdgcn.{icmp,fcmp}

Summary:
The return value of these intrinsics should always have 0 bits for
inactive threads. This means that when all arguments are constant
and the comparison evaluates to true, the intrinsic should return
the current exec mask.

Fixes some GL_ARB_shader_ballot tests.

Reviewers: arsenm

Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye

Differential Revision: https://reviews.llvm.org/D32344

llvm-svn: 301195

7 years ago[GlobalISel][X86] Lower FormalArgument/Ret using G_MERGE_VALUES/G_UNMERGE_VALUES.
Igor Breger [Mon, 24 Apr 2017 17:05:52 +0000 (17:05 +0000)]
[GlobalISel][X86] Lower FormalArgument/Ret using G_MERGE_VALUES/G_UNMERGE_VALUES.

Summary: [GlobalISel][X86] Lower FormalArgument/Ret using G_MERGE_VALUES/G_UNMERGE_VALUES.

Reviewers: zvi, t.p.northover, guyblank

Reviewed By: t.p.northover

Subscribers: dberris, rovka, llvm-commits, kristof.beyls

Differential Revision: https://reviews.llvm.org/D32288

llvm-svn: 301194

7 years ago[DAGCombiner] Updated bswap byte offset variable names to be more descriptive. NFC
Simon Pilgrim [Mon, 24 Apr 2017 17:05:14 +0000 (17:05 +0000)]
[DAGCombiner] Updated bswap byte offset variable names to be more descriptive. NFC

As discussed on D32039, use MaskByteOffset to describe the variable and also pull out repeated getOpcode() calls.

llvm-svn: 301193

7 years ago[APInt] Fix repeated word in comments. NFC
Craig Topper [Mon, 24 Apr 2017 17:00:22 +0000 (17:00 +0000)]
[APInt] Fix repeated word in comments. NFC

llvm-svn: 301192

7 years agoAMDGPU: Fix crash when scheduling non-memory SMRD instructions
Nicolai Haehnle [Mon, 24 Apr 2017 16:53:52 +0000 (16:53 +0000)]
AMDGPU: Fix crash when scheduling non-memory SMRD instructions

Summary: Fixes piglit spec/arb_shader_clock/execution/*

Reviewers: arsenm

Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye

Differential Revision: https://reviews.llvm.org/D32345

llvm-svn: 301191

7 years ago[tsan] Include __tsan_external_* API from a header file instead of declaring them...
Kuba Mracek [Mon, 24 Apr 2017 16:48:30 +0000 (16:48 +0000)]
[tsan] Include __tsan_external_* API from a header file instead of declaring them manually. NFC.

Differential Revision: https://reviews.llvm.org/D32384

llvm-svn: 301190

7 years ago[tsan] Remove the extra word "object" from description of external races
Kuba Mracek [Mon, 24 Apr 2017 16:42:29 +0000 (16:42 +0000)]
[tsan] Remove the extra word "object" from description of external races

Differential Revision: https://reviews.llvm.org/D32383

llvm-svn: 301189

7 years ago[clang-tidy] Some Cleanups for performance-faster-string-find check.
Haojian Wu [Mon, 24 Apr 2017 16:41:00 +0000 (16:41 +0000)]
[clang-tidy] Some Cleanups for performance-faster-string-find check.

NFC

llvm-svn: 301188

7 years ago[SDAG] Teach Chain Analysis about BaseIndexOffset addressing.
Nirav Dave [Mon, 24 Apr 2017 15:37:20 +0000 (15:37 +0000)]
[SDAG] Teach Chain Analysis about BaseIndexOffset addressing.

While we use BaseIndexOffset in FindBetterNeighborChains to
appropriately realize they're almost the same address and should be
improved concurrently we do not use it in isAlias using the non-index
understanding FindBaseOffset instead. Adding a BaseIndexOffset check
in isAlias like should allow indexed stores to be merged.

FindBaseOffset to be excised in subsequent patch.

Reviewers: jyknight, aditya_nandakumar, bogner

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D31987

llvm-svn: 301187

7 years agoUpdate two android XFAILS
Pavel Labath [Mon, 24 Apr 2017 15:23:21 +0000 (15:23 +0000)]
Update two android XFAILS

- XFAIL on TestNoreturnUnwind on all architectures
- TestStaticVariables fails with clang-3.8 as well

llvm-svn: 301186

7 years agoExtend readability-container-size-empty to add comparisons to empty-state objects.
Aaron Ballman [Mon, 24 Apr 2017 14:57:09 +0000 (14:57 +0000)]
Extend readability-container-size-empty to add comparisons to empty-state objects.

Patch by Josh Zimmerman.

llvm-svn: 301185

7 years ago[sanitizer] Cache SizeClassForTransferBatch in the 32-bit local cache
Kostya Kortchinsky [Mon, 24 Apr 2017 14:53:38 +0000 (14:53 +0000)]
[sanitizer] Cache SizeClassForTransferBatch in the 32-bit local cache

Summary:
`SizeClassForTransferBatch` is expensive and is called for every `CreateBatch`
and `DestroyBatch`. Caching it means `kNumClasses` calls in `InitCache`
instead. This should be a performance gain if more than `kNumClasses / 2`
batches are created and destroyed during the lifetime of the local cache.

I have chosen to fully remove the function and putting the code in `InitCache`,
which is a debatable choice.

In single threaded benchmarks leveraging primary backed allocations, this turns
out to be a sizeable gain in performances (greater than 5%). In multithreaded
benchmarks leveraging everything, it is less significant but still an
improvement (about 1%).

Reviewers: kcc, dvyukov, alekseyshl

Reviewed By: dvyukov

Subscribers: kubamracek, llvm-commits

Differential Revision: https://reviews.llvm.org/D32365

llvm-svn: 301184

7 years ago[index] If the 'external_source_symbol' attribute indicates 'Swift' as the language...
Argyrios Kyrtzidis [Mon, 24 Apr 2017 14:52:00 +0000 (14:52 +0000)]
[index] If the 'external_source_symbol' attribute indicates 'Swift' as the language then report it accordingly

llvm-svn: 301183

7 years agoclang-format: Fix bad corner case in formatting of function types.
Daniel Jasper [Mon, 24 Apr 2017 14:28:49 +0000 (14:28 +0000)]
clang-format: Fix bad corner case in formatting of function types.

Before:
  std::function<
      LoooooooooooongTemplatedType<SomeType>*(
          LooooooooooooooooooooongType
              type)>
      function;

After:
  std::function<
      LoooooooooooongTemplatedType<
          SomeType>*(
          LooooooooooooooooongType type)>
      function;

clang-format generally avoids having lines like "SomeType>*(" as they
lead to parameter lists that don't belong together to be aligned. However, in
case it is better than the alternative, which can even be violating the column
limit.

llvm-svn: 301182

7 years ago[X86][AVX] Add scheduling latency/throughput tests for missing AVX1 instructions
Simon Pilgrim [Mon, 24 Apr 2017 14:26:30 +0000 (14:26 +0000)]
[X86][AVX] Add scheduling latency/throughput tests for missing AVX1 instructions

Had to split btver2/znver1 checks as only btver2 suppresses zeroupper

llvm-svn: 301181

7 years ago[index] The relation between the declarations in template specializations
Alex Lorenz [Mon, 24 Apr 2017 14:04:58 +0000 (14:04 +0000)]
[index] The relation between the declarations in template specializations
that 'override' declarations in the base template should be recorded

This can be used for improved "go to definition" feature in Xcode.

rdar://31604739

Differential Revision: https://reviews.llvm.org/D32020

llvm-svn: 301180

7 years agoFix the new SocketAddressTest on Windows
Pavel Labath [Mon, 24 Apr 2017 13:34:35 +0000 (13:34 +0000)]
Fix the new SocketAddressTest on Windows

we need to call WSAStartup before we can use getaddrinfo.

llvm-svn: 301179

7 years ago[Devirtualization] Emit invariant.group loads with empty group md
Piotr Padlewski [Mon, 24 Apr 2017 12:58:43 +0000 (12:58 +0000)]
[Devirtualization] Emit invariant.group loads with empty group md

Summary:
As discussed here
http://lists.llvm.org/pipermail/llvm-dev/2017-January/109332.html
having different groups doesn't solve the problem entirly.

Reviewers: rjmccall, rsmith

Subscribers: amharc, cfe-commits

Differential Revision: https://reviews.llvm.org/D32110

llvm-svn: 301178

7 years ago[SystemZ] Update kill-flag in splitMove().
Jonas Paulsson [Mon, 24 Apr 2017 12:40:28 +0000 (12:40 +0000)]
[SystemZ]  Update kill-flag in splitMove().

EarlierMI needs to clear the kill flag on the first operand in case of a store.

Review: Ulrich Weigand
llvm-svn: 301177

7 years ago[DWARF] Move test to x86 directory
Renato Golin [Mon, 24 Apr 2017 12:37:11 +0000 (12:37 +0000)]
[DWARF] Move test to x86 directory

llvm-svn: 301176

7 years ago[RegionInfo] Fix dangling references created by moving RegionInfo objects
Philip Pfaffe [Mon, 24 Apr 2017 11:54:37 +0000 (11:54 +0000)]
[RegionInfo] Fix dangling references created by moving RegionInfo objects

Summary: Region objects capture the address of the creating RegionInfo instance. Because the RegionInfo class is movable, moving a RegionInfo object creates dangling references. This patch fixes these references by walking the Regions post-move, and updating references to the new parent.

Reviewers: Meinersbur, grosser

Reviewed By: Meinersbur, grosser

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D31719

llvm-svn: 301175

7 years agoAdd SUSE vendor
Ismail Donmez [Mon, 24 Apr 2017 11:18:29 +0000 (11:18 +0000)]
Add SUSE vendor

Summary: SUSE's ARM triples end with -gnueabi even though they are hard-float. This requires special handling of SUSE ARM triples. Hence we need a way to differentiate the SUSE as vendor. This CL adds that.

Reviewers: chandlerc, compnerd, echristo, rengolin

Reviewed By: rengolin

Subscribers: aemerson, rengolin, llvm-commits

Differential Revision: https://reviews.llvm.org/D32426

llvm-svn: 301174

7 years agoRevert [scudo] Enabling MIPS support for Scudo
Sagar Thakur [Mon, 24 Apr 2017 11:02:36 +0000 (11:02 +0000)]
Revert [scudo] Enabling MIPS support for Scudo

This patch broke the buildbot clang-cmake-mips. Investigating the issue.

llvm-svn: 301173

7 years ago[LLDB][MIPS] Move it into HandleLLVMOptions.cmake.
Nitesh Jain [Mon, 24 Apr 2017 10:56:01 +0000 (10:56 +0000)]
[LLDB][MIPS] Move it into HandleLLVMOptions.cmake.

The revison https://reviews.llvm.org/D32125 will fixed the off_t for GNU specific 32 bit platform. This fixed the difference in definition of off_t in LLDB and LLVM

Subscribers: jaydeep, bhushan, lldb-commits, slthakur, llvm-commits, krytarowski, emaste, zturner
llvm-svn: 301172

7 years ago[LLVM][MIPS] Fix different definition of off_t in LLDB and LLVM.
Nitesh Jain [Mon, 24 Apr 2017 10:36:46 +0000 (10:36 +0000)]
[LLVM][MIPS] Fix different definition of off_t in LLDB and LLVM.

Reviewers: beanz

Subscribers: jaydeep, bhushan, lldb-commits, slthakur, llvm-commits, krytarowski, emaste

Differential Revision: https://reviews.llvm.org/D32125

llvm-svn: 301171

7 years ago[DWARF] - Take relocations in account when extracting ranges from .debug_ranges
George Rimar [Mon, 24 Apr 2017 10:19:45 +0000 (10:19 +0000)]
[DWARF] - Take relocations in account when extracting ranges from .debug_ranges

I found this when investigated "Bug 32319 - .gdb_index is broken/incomplete" for LLD.

When we have object file with .debug_ranges section it may be filled with zeroes.
Relocations are exist in file to relocate this zeroes into real values later, but until that
a pair of zeroes is treated as terminator. And DWARF parser thinks there is no ranges at all
when I am trying to collect address ranges for building .gdb_index.

Solution implemented in this patch is to take relocations in account when parsing ranges.

Differential revision: https://reviews.llvm.org/D32228

llvm-svn: 301170

7 years agoAdd more arguments to SocketAddress::GetAddressInfo
Pavel Labath [Mon, 24 Apr 2017 09:39:56 +0000 (09:39 +0000)]
Add more arguments to SocketAddress::GetAddressInfo

Summary:
the reason for this is two-fold:
- getaddrinfo without the extra arguments will return the same
(network-level) address multiple times, once for each supported
transport protocol, which is not what is usually intended (it certainly
wasn't in D31823)
- it enables us to rewrite the getaddrinfo member function in terms of
the static GetAddressInfo function.

Reviewers: beanz, tberghammer

Subscribers: lldb-commits

Differential Revision: https://reviews.llvm.org/D32357

llvm-svn: 301168

7 years ago[clang-tidy] New check: modernize-replace-random-shuffle.
Mads Ravn [Mon, 24 Apr 2017 09:27:20 +0000 (09:27 +0000)]
[clang-tidy] New check: modernize-replace-random-shuffle.

This check will find occurrences of ``std::random_shuffle`` and replace it with ``std::shuffle``. In C++17 ``std::random_shuffle`` will no longer be available and thus we need to replace it.

Example of case that it fixes

```
  std::vector<int> v;

  // First example
  std::random_shuffle(vec.begin(), vec.end());

```

Reviewers: hokein, aaron.ballman, alexfh, malcolm.parsons, mclow.lists

Subscribers: cfe-commits

Differential Revision: https://reviews.llvm.org/D30158

llvm-svn: 301167

7 years ago[ARM] GlobalISel: Legalize s8 and s16 G_(S|U)DIV
Diana Picus [Mon, 24 Apr 2017 09:12:19 +0000 (09:12 +0000)]
[ARM] GlobalISel: Legalize s8 and s16 G_(S|U)DIV

We have to widen the operands to 32 bits and then we can either use
hardware division if it is available or lower to a libcall otherwise.

At the moment it is not enough to set the Legalizer action to
WidenScalar, since for libcalls it won't know what to do (it won't be
able to find what size to widen to, because it will find Libcall and not
Legal for 32 bits). To hack around this limitation, we request Custom
lowering, and as part of that we widen first and then we run another
legalizeInstrStep on the widened DIV.

llvm-svn: 301166

7 years ago[Arch64AsmParser] better diagnostic for isb
Sjoerd Meijer [Mon, 24 Apr 2017 08:22:20 +0000 (08:22 +0000)]
[Arch64AsmParser] better diagnostic for isb

Instruction isb takes as an operand either 'sy' or an immediate value. This
improves the diagnostic when the string is not 'sy' and adds a test case for
this which was missing. This also adds tests to check invalid inputs for dsb
and dmb.

Differential Revision: https://reviews.llvm.org/D32227

llvm-svn: 301165

7 years ago[ARM] GlobalISel: Support G_(S|U)DIV for s32
Diana Picus [Mon, 24 Apr 2017 08:20:05 +0000 (08:20 +0000)]
[ARM] GlobalISel: Support G_(S|U)DIV for s32

Add support for both targets with hardware division and without. For
hardware division we have to add support throughout the pipeline
(legalizer, reg bank select, instruction select). For targets without
hardware division, we only need to mark it as a libcall.

llvm-svn: 301164

7 years ago[GlobalISel] Legalize G_(S|U)DIV libcalls
Diana Picus [Mon, 24 Apr 2017 07:22:31 +0000 (07:22 +0000)]
[GlobalISel] Legalize G_(S|U)DIV libcalls

Treat them the same as the other binary operations that we have so far,
but on integers rather than floating point types. Extract the common
code into a helper.

This will be used in the ARM backend.

llvm-svn: 301163

7 years ago[ARM] GlobalISel: Select G_CONSTANT with CImm operands
Diana Picus [Mon, 24 Apr 2017 06:30:56 +0000 (06:30 +0000)]
[ARM] GlobalISel: Select G_CONSTANT with CImm operands

When selecting a G_CONSTANT to a MOVi, we need the value to be an Imm
operand. We used to just leave the G_CONSTANT operand unchanged, which
works in some cases (such as the GEP offsets that we create when
referring to stack slots). However, in many other places the G_CONSTANTs
are created with CImm operands. This patch makes sure to handle those as
well, and to error out gracefully if in the end we don't end up with an
Imm operand.

Thanks to Oliver Stannard for reporting this issue.

llvm-svn: 301162

7 years ago[XRay][tools] Fixup for pedantic and permissive errors/warnings
Dean Michael Berris [Mon, 24 Apr 2017 06:15:53 +0000 (06:15 +0000)]
[XRay][tools] Fixup for pedantic and permissive errors/warnings

Remove extraneous semicolons and fully qualify the Trace type.

Follow-up to D29320.

llvm-svn: 301161

7 years ago[XRay] A tool for Comparing xray function call graphs
Dean Michael Berris [Mon, 24 Apr 2017 05:54:33 +0000 (05:54 +0000)]
[XRay] A tool for Comparing xray function call graphs

Summary:
This is a tool for comparing the function graphs produced by the
llvm-xray graph too. It takes the form of a new subcommand of the
llvm-xray tool 'graph-diff'.

This initial version of the patch is very rough, but it is close to
feature complete.

Depends on D29363

Reviewers: dblaikie, dberris

Reviewed By: dberris

Subscribers: mgorny, llvm-commits

Differential Revision: https://reviews.llvm.org/D29320

llvm-svn: 301160

7 years ago[APInt] Make behavior of ashr by BitWidth consistent between single and multi word.
Craig Topper [Mon, 24 Apr 2017 05:38:26 +0000 (05:38 +0000)]
[APInt] Make behavior of ashr by BitWidth consistent between single and multi word.

Previously single word would always return 0 regardless of the original sign. Multi word would return all 0s or all 1s based on the original sign. Now single word takes into account the sign as well.

llvm-svn: 301159

7 years ago[scudo] Enabling MIPS support for Scudo
Sagar Thakur [Mon, 24 Apr 2017 04:29:44 +0000 (04:29 +0000)]
[scudo] Enabling MIPS support for Scudo

Adding MIPS 32-bit and 64-bit support for Scudo.

Reviewed by cryptoad
Differential: D31803

llvm-svn: 301158

7 years agoRevert "Refactor DynamicLibrary so searching for a symbol will have a defined order.”
Frederich Munch [Mon, 24 Apr 2017 03:33:30 +0000 (03:33 +0000)]
Revert "Refactor DynamicLibrary so searching for a symbol will have a defined order.”
The changes are causing the i686-mingw32 build to fail.

This reverts commit r301153, and the changes for a separate warning on i686-mingw32 in r301155  and r301156.

llvm-svn: 301157

7 years agoFix warning converting from boolean to pointer introduced in r301153.
Frederich Munch [Mon, 24 Apr 2017 03:12:16 +0000 (03:12 +0000)]
Fix warning converting from boolean to pointer introduced in r301153.

This reverts commit r301155, which was incorrect.

llvm-svn: 301156

7 years agoFix warning converting from void* to boolean introduced in r301153.
Frederich Munch [Mon, 24 Apr 2017 02:51:40 +0000 (02:51 +0000)]
Fix warning converting from void* to boolean introduced in r301153.

llvm-svn: 301155

7 years agoRevert "[SCEV] Enable SCEV verification by default in EXPENSIVE_CHECKS builds"
Sanjoy Das [Mon, 24 Apr 2017 02:35:19 +0000 (02:35 +0000)]
Revert "[SCEV] Enable SCEV verification by default in EXPENSIVE_CHECKS builds"

This reverts commit r301150.  It breaks CodeGen/Hexagon/hwloop-wrap2.ll, reverting
while I investigate.

llvm-svn: 301154

7 years agoRefactor DynamicLibrary so searching for a symbol will have a defined order and
Frederich Munch [Mon, 24 Apr 2017 02:30:12 +0000 (02:30 +0000)]
Refactor DynamicLibrary so searching for a symbol will have a defined order and
libraries are properly unloaded when llvm_shutdown is called.

Summary:
This was mostly affecting usage of the JIT, where storing the library handles in
a set made iteration unordered/undefined. This lead to disagreement between the
JIT and native code as to what the address and implementation of particularly on
Windows with stdlib functions:

JIT: putenv_s("TEST", "VALUE") // called msvcrt.dll, putenv_s
JIT: getenv("TEST") -> "VALUE" // called msvcrt.dll, getenv
Native: getenv("TEST") -> NULL // called ucrt.dll, getenv

Also fixed is the issue of DynamicLibrary::getPermanentLibrary(0,0) on Windows
not giving priority to the process' symbols as it did on Unix.

Reviewers: chapuni, v.g.vassilev, lhames

Reviewed By: lhames

Subscribers: danalbert, srhines, mgorny, vsk, llvm-commits

Differential Revision: https://reviews.llvm.org/D30107

llvm-svn: 301153

7 years ago[Orc] Fix a warning by removing an unused lambda capture.
Lang Hames [Mon, 24 Apr 2017 01:21:23 +0000 (01:21 +0000)]
[Orc] Fix a warning by removing an unused lambda capture.

llvm-svn: 301152