OSDN Git Service

android-x86/external-llvm.git
5 years ago[MIPS GlobalISel] RegBankSelect for chains of ambiguous instructions
Petar Avramovic [Thu, 11 Jul 2019 09:22:49 +0000 (09:22 +0000)]
[MIPS GlobalISel] RegBankSelect for chains of ambiguous instructions

When one of the uses/defs of ambiguous instruction is also ambiguous
visit it recursively and search its uses/defs for instruction with
only one mapping available.
When all instruction in a chain are ambiguous arbitrary mapping can
be selected. For s64 operands in ambiguous chain fprb is selected since
it results in less instructions then having to narrow scalar s64 to s32.
For s32 both gprb and fprb result in same number of instructions and
gprb is selected like a general purpose option.

At the moment we always avoid cross register bank copies.
TODO: Implement a model for costs calculations of different mappings
on same instruction and cross bank copies. Allow cross bank copies
when appropriate according to cost model.

Differential Revision: https://reviews.llvm.org/D64485

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365743 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert Recommit "[CommandLine] Remove OptionCategory and SubCommand caches from the...
Haojian Wu [Thu, 11 Jul 2019 08:54:28 +0000 (08:54 +0000)]
Revert Recommit "[CommandLine] Remove OptionCategory and SubCommand caches from the Option class."

This reverts r365675 (git commit 43d75f977853c3ec891a440c362b2df183a211b5)

The patch causes a crash in SupportTests (CommandLineTest.AliasesWithArguments).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365742 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRemove some redundant code from r290372 and improve a comment.
Jay Foad [Thu, 11 Jul 2019 08:49:52 +0000 (08:49 +0000)]
Remove some redundant code from r290372 and improve a comment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365741 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ARM][ParallelDSP] Change the search for smlads
Sam Parker [Thu, 11 Jul 2019 07:47:50 +0000 (07:47 +0000)]
[ARM][ParallelDSP] Change the search for smlads

Two functional changes have been made here:
- Now search up from any add instruction to find the chains of
  operations that we may turn into a smlad. This allows the
  generation of a smlad which doesn't accumulate into a phi.
- The search function has been corrected to stop it falsely searching
  up through an invalid path.

The bulk of the changes have been making the Reduction struct a class
and making it more C++y with getters and setters.

Differential Revision: https://reviews.llvm.org/D61780

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365740 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[test] Silence gcc 7.4 warning [NFC]
Mikael Holmen [Thu, 11 Jul 2019 07:07:23 +0000 (07:07 +0000)]
[test] Silence gcc 7.4 warning [NFC]

Without this gcc 7.4.0 complains with
 ../unittests/Analysis/ValueTrackingTest.cpp:937:66: error: ISO C++11 requires at least one argument for the "..." in a variadic macro [-Werror]
                          ::testing::ValuesIn(IsBytewiseValueTests));
                                                                   ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365738 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WebAssembly] Print error message for llvm.clear_cache intrinsic
Heejin Ahn [Thu, 11 Jul 2019 05:55:47 +0000 (05:55 +0000)]
[WebAssembly] Print error message for llvm.clear_cache intrinsic

Summary:
Wasm does not currently support `llvm.clear_cache` intrinsic, and this
prints a proper error message instead of segfault.

Reviewers: dschuff, sbc100, sunfish

Subscribers: jgravelle-google, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64322

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365731 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine][NFCI] Add test coverage to onehot_merge.ll
Huihui Zhang [Thu, 11 Jul 2019 04:56:37 +0000 (04:56 +0000)]
[InstCombine][NFCI] Add test coverage to onehot_merge.ll

Prep work for upcoming patch D64275.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365729 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[SCEV] teach SCEV symbolical execution about overflow intrinsics folding.
Chen Zheng [Thu, 11 Jul 2019 02:18:22 +0000 (02:18 +0000)]
[SCEV] teach SCEV symbolical execution about overflow intrinsics folding.
Differential Revision: https://reviews.llvm.org/D64422

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365726 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoReplace three "strip & accumulate" implementations with a single one
Johannes Doerfert [Thu, 11 Jul 2019 01:14:48 +0000 (01:14 +0000)]
Replace three "strip & accumulate" implementations with a single one

This patch replaces the three almost identical "strip & accumulate"
implementations for constant pointer offsets with a single one,
combining the respective functionalities. The old interfaces are kept
for now.

Differential Revision: https://reviews.llvm.org/D64468

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365723 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[NFC] Adjust "invalid.ll.bc" tests to check for AttrKind #255 not #63
Johannes Doerfert [Thu, 11 Jul 2019 01:14:30 +0000 (01:14 +0000)]
[NFC] Adjust "invalid.ll.bc" tests to check for AttrKind #255 not #63

We are about to add enum attributes with AttrKind numbers >= 63. This
means we cannot use AttrKind #63 to test for an invalid attribute number
in the RAW format anymore. This patch changes the number of an invalid
attribute to #255. There is no change to the character of the tests.

Differential Revision: https://reviews.llvm.org/D64531

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365722 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Don't convert 8 or 16 bit ADDs to LEAs on Atom in FixupLEAPass.
Craig Topper [Thu, 11 Jul 2019 01:01:39 +0000 (01:01 +0000)]
[X86] Don't convert 8 or 16 bit ADDs to LEAs on Atom in FixupLEAPass.

We use the functions that convert to three address to do the
conversion, but changing an 8 or 16 bit will cause it to create
a virtual register. This can't be done after register allocation
where this pass runs.

I've switched the pass completely to a white list of instructions
that can be converted to LEA instead of a blacklist that was
incorrect. This will avoid surprises if we enhance the three
address conversion function to include additional instructions
in the future.

Fixes PR42565.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365720 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AMDGPU] gfx908 atomic fadd and atomic pk_fadd
Stanislav Mekhanoshin [Thu, 11 Jul 2019 00:10:17 +0000 (00:10 +0000)]
[AMDGPU] gfx908 atomic fadd and atomic pk_fadd

Differential Revision: https://reviews.llvm.org/D64435

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365717 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AMDGPU] gfx908 dot instruction support
Stanislav Mekhanoshin [Thu, 11 Jul 2019 00:00:27 +0000 (00:00 +0000)]
[AMDGPU] gfx908 dot instruction support

Differential Revision: https://reviews.llvm.org/D64431

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365715 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert [llvm-objcopy] Allow strip symtab from executables and DSOs
Jordan Rupprecht [Wed, 10 Jul 2019 23:32:44 +0000 (23:32 +0000)]
Revert [llvm-objcopy] Allow strip symtab from executables and DSOs

This reverts r365193 (git commit 194f16b3548bcb23a7f0fd638778ed72edd18d37)

This patch doesn't work with binaries built w/ `--emit-relocs`, e.g.

```
$ echo 'int main() { return 0; }' | clang -Wl,--emit-relocs -x c - -o foo && llvm-objcopy --strip-unneeded foo
llvm-objcopy: error: 'foo': not stripping symbol '__gmon_start__' because it is named in a relocation
```

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365712 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[SDAG] commute setcc operands to match a subtract
Sanjay Patel [Wed, 10 Jul 2019 23:23:54 +0000 (23:23 +0000)]
[SDAG] commute setcc operands to match a subtract

If we have:

R = sub X, Y
P = cmp Y, X

...then flipping the operands in the compare instruction can allow using a subtract that sets compare flags.

Motivated by diffs in D58875 - not sure if this changes anything there,
but this seems like a good thing independent of that.

There's a more involved version of this transform already in IR (in instcombine
although that seems misplaced to me) - see "swapMayExposeCSEOpportunities()".

Differential Revision: https://reviews.llvm.org/D63958

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365711 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAdd IsBytewiseValue unit test
Vitaly Buka [Wed, 10 Jul 2019 22:56:15 +0000 (22:56 +0000)]
Add IsBytewiseValue unit test

Subscribers: llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63966

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365710 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoNFC: Pass DataLayout into isBytewiseValue
Vitaly Buka [Wed, 10 Jul 2019 22:53:52 +0000 (22:53 +0000)]
NFC: Pass DataLayout into isBytewiseValue

Summary:
We will need to handle IntToPtr which I will submit in a separate patch as it's
not going to be NFC.

Reviewers: eugenis, pcc

Reviewed By: eugenis

Subscribers: hiraditya, cfe-commits, llvm-commits

Tags: #clang, #llvm

Differential Revision: https://reviews.llvm.org/D63940

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365709 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Add patterns with and_flag_nocf for BLSI and TBM instructions.
Craig Topper [Wed, 10 Jul 2019 22:44:32 +0000 (22:44 +0000)]
[X86] Add patterns with and_flag_nocf for BLSI and TBM instructions.

Fixes similar issues to r352306.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365705 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Add a few more TBM and BLSI test cases that show the same issue that r352306...
Craig Topper [Wed, 10 Jul 2019 22:44:24 +0000 (22:44 +0000)]
[X86] Add a few more TBM and BLSI test cases that show the same issue that r352306 fixed for BLSR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365704 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Add BLSR and BLSMSK to isUseDefConvertible.
Craig Topper [Wed, 10 Jul 2019 22:14:39 +0000 (22:14 +0000)]
[X86] Add BLSR and BLSMSK to isUseDefConvertible.

Unfortunately subo formation in CGP prevents obvious ways of
testing this.

But we already have BLSI in here and the flag behavior is
well understood.

Might become more useful if we improve PR42571.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365702 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[NFC]Fix IR/MC depency issue for function descriptor SDAG implementation
David Tenty [Wed, 10 Jul 2019 22:13:55 +0000 (22:13 +0000)]
[NFC]Fix IR/MC depency issue for function descriptor SDAG implementation

Summary: llvm/IR/GlobalValue.h can't be included in MC, that creates a circular dependency between MC and IR libraries. This circular dependency is causing an issue for build system that enforce layering.

Author: Xiangling_L

Reviewers: sfertile, jasonliu, hubert.reinterpretcast, gribozavr

Reviewed By: gribozavr

Subscribers: wuzish, nemanjai, hiraditya, kbarton, MaskRay, jsji, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64445

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365701 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Remove unused variable. NFC
Craig Topper [Wed, 10 Jul 2019 21:01:34 +0000 (21:01 +0000)]
[X86] Remove unused variable. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365697 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[NFC][InstCombine] Comb through just-added "omit mask before left-shift" tests once...
Roman Lebedev [Wed, 10 Jul 2019 19:58:13 +0000 (19:58 +0000)]
[NFC][InstCombine] Comb through just-added "omit mask before left-shift" tests once more

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365694 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAdd support for building with Python 3
Nico Weber [Wed, 10 Jul 2019 19:31:25 +0000 (19:31 +0000)]
Add support for building with Python 3

LLDB supports Python 3 now.

Differential Revision: https://reviews.llvm.org/D64443

Patch from Christian Biesinger <cbiesinger@google.com>!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365692 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AArch64][GlobalISel] Optimize compare and branch cases with G_INTTOPTR and unknown...
Amara Emerson [Wed, 10 Jul 2019 19:21:43 +0000 (19:21 +0000)]
[AArch64][GlobalISel] Optimize compare and branch cases with G_INTTOPTR and unknown values.

Since we have distinct types for pointers and scalars, G_INTTOPTRs can sometimes
obstruct attempts to find constant source values. These usually come about when
try to do some kind of null pointer check. Teaching getConstantVRegValWithLookThrough
about this operation allows the CBZ/CBNZ optimization to catch more cases.

This change also improves the case where we can't find a constant source at all.
Previously we would emit a cmp, cset and tbnz for that. Now we try to just emit
a cmp and conditional branch, saving an instruction.

The cumulative code size improvement of this change plus D64354 is 5.5% geomean
on arm64 CTMark -O0.

Differential Revision: https://reviews.llvm.org/D64377

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365690 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agogn build: Merge r365687
Nico Weber [Wed, 10 Jul 2019 19:20:17 +0000 (19:20 +0000)]
gn build: Merge r365687

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365689 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert "[ELF] Loose a condition for relocation with a symbol"
Nikola Prica [Wed, 10 Jul 2019 18:58:05 +0000 (18:58 +0000)]
Revert "[ELF] Loose a condition for relocation with a symbol"

This reverts commit 8507eca1647118e73435b0ce1de8a1952a021d01.

Reveting due to some suspicious failurse in santizer-x86_64-linux.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365685 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[GlobalISel][AArch64] Use getOpcodeDef instead of findMIFromReg
Jessica Paquette [Wed, 10 Jul 2019 18:46:56 +0000 (18:46 +0000)]
[GlobalISel][AArch64] Use getOpcodeDef instead of findMIFromReg

Some minor cleanup.

This function in Utils does the same thing as `findMIFromReg`. It also looks
through copies, which `findMIFromReg` didn't.

Delete `findMIFromReg` and use `getOpcodeDef` instead. This only happens in
`tryOptVectorDup` right now.

Update opt-shuffle-splat to show that we can look through the copies now, too.

Differential Revision: https://reviews.llvm.org/D64520

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365684 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[GlobalISel][AArch64][NFC] Use getDefIgnoringCopies from Utils where we can
Jessica Paquette [Wed, 10 Jul 2019 18:44:57 +0000 (18:44 +0000)]
[GlobalISel][AArch64][NFC] Use getDefIgnoringCopies from Utils where we can

There are a few places where we walk over copies throughout
AArch64InstructionSelector.cpp. In Utils, there's a function that does exactly
this which we can use instead.

Note that the utility function works with the case where we run into a COPY
from a physical register. We've run into bugs with this a couple times, so using
it should defend us from similar future bugs.

Also update opt-fold-compare.mir to show that we still handle physical registers
properly.

Differential Revision: https://reviews.llvm.org/D64513

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365683 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert "[System Model] [TTI] Update cache and prefetch TTI interfaces"
David Greene [Wed, 10 Jul 2019 18:25:58 +0000 (18:25 +0000)]
Revert "[System Model] [TTI] Update cache and prefetch TTI interfaces"

This broke some PPC prefetching tests.

This reverts commit 9fdfb045ae8bb643ab0d0455dcf9ecaea3b1eb3c.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365680 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoMove three folds for FADD, FSUB and FMUL in the DAG combiner away from Unsafe to...
Michael Berg [Wed, 10 Jul 2019 18:23:26 +0000 (18:23 +0000)]
Move three folds for FADD, FSUB and FMUL in the DAG combiner away from Unsafe to more aligned checks that reflect context

Summary: Unsafe does not map well alone for each of these three cases as it is missing NoNan context when accessed directly with clang.  I have migrated the fold guards to reflect the expectations of handing nan and zero contexts directly (NoNan, NSZ) and some tests with it.  Unsafe does include NSZ, however there is already precedent for using the target option directly to reflect that context.

Reviewers: spatel, wristow, hfinkel, craig.topper, arsenm

Reviewed By: arsenm

Subscribers: michele.scandale, wdng, javed.absar

Differential Revision: https://reviews.llvm.org/D64450

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365679 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[System Model] [TTI] Update cache and prefetch TTI interfaces
David Greene [Wed, 10 Jul 2019 18:07:01 +0000 (18:07 +0000)]
[System Model] [TTI] Update cache and prefetch TTI interfaces

Rework the TTI cache and software prefetching APIs to prepare for the
introduction of a general system model.  Changes include:

- Marking existing interfaces const and/or override as appropriate
- Adding comments
- Adding BasicTTIImpl interfaces that delegate to a subtarget
  implementation
- Adding a default "no information" subtarget implementation

Only a handful of targets use these interfaces currently: AArch64,
Hexagon, PPC and SystemZ.  AArch64 already has a custom subtarget
implementation, so its custom TTI implementation is migrated to use
the new facilities in BasicTTIImpl to invoke its custom subtarget
implementation.  The custom TTI implementations continue to exist for
the other targets with this change.  They are not moved over to
subtarget-based implementations.

The end goal is to have the default subtarget implementation defer to
the system model defined by the target.  With this change, the default
subtarget implementation essentially returns "no information" for
these interfaces.  None of the existing users of TTI will hit that
implementation because they define their own custom TTI
implementations and won't use the BasicTTIImpl implementations.

Once system models are in place for the targets that use these
interfaces, their custom TTI implementations can be removed.

Differential Revision: https://reviews.llvm.org/D63614

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365676 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRecommit "[CommandLine] Remove OptionCategory and SubCommand caches from the Option...
Don Hinton [Wed, 10 Jul 2019 17:57:05 +0000 (17:57 +0000)]
Recommit "[CommandLine] Remove OptionCategory and SubCommand caches from the Option class."

Previously reverted in 364141 due to buildbot breakage, and fixed here
by making GeneralCategory global a ManagedStatic.

Summary:
This change processes `OptionCategory`s and `SubCommand`s as they
are seen instead of caching them in the Option class and processing
them later.  Doing so simplifies the work needed to be done by the Global
parser and significantly reduces the size of the Option class to a mere 64
bytes.

Removing  the `OptionCategory` cache saved 24 bytes, and removing
the `SubCommand` cache saved an additional 48 bytes, for a total of a
72 byte reduction.

Reviewed By: serge-sans-paille

Tags: #llvm, #clang

Differential Revision: https://reviews.llvm.org/D62105

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365675 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] EltsFromConsecutiveLoads - clean up element size calcs. NFCI.
Simon Pilgrim [Wed, 10 Jul 2019 17:49:27 +0000 (17:49 +0000)]
[X86] EltsFromConsecutiveLoads - clean up element size calcs. NFCI.

Determine the element/load size calculations earlier and assert that they are whole bytes in size.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365674 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[LoopRotate + MemorySSA] Keep an <instruction-cloned instruction> map.
Alina Sbirlea [Wed, 10 Jul 2019 17:36:56 +0000 (17:36 +0000)]
[LoopRotate + MemorySSA] Keep an <instruction-cloned instruction> map.

Summary:
The map kept in loop rotate is used for instruction remapping, in order
to simplify the clones of instructions. Thus, if an instruction can be
simplified, its simplified value is placed in the map, even when the
clone is added to the IR. MemorySSA in contrast needs to know about that
clone, so it can add an access for it.
To resolve this: keep a different map for MemorySSA.

Reviewers: george.burgess.iv

Subscribers: jlebar, Prazek, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63680

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365672 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ORC] Add custom IR compiler configuration to LLJITBuilder to enable obj caches.
Lang Hames [Wed, 10 Jul 2019 17:24:24 +0000 (17:24 +0000)]
[ORC] Add custom IR compiler configuration to LLJITBuilder to enable obj caches.

LLJITBuilder now has a setCompileFunctionCreator method which can be used to
construct a CompileFunction for the LLJIT instance being created. The motivating
use-case for this is supporting ObjectCaches, which can now be set up at
compile-function construction time. To demonstrate this an example project,
LLJITWithObjectCache, is included.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365671 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Regenerate tests. NFCI.
Simon Pilgrim [Wed, 10 Jul 2019 17:22:31 +0000 (17:22 +0000)]
[X86] Regenerate tests. NFCI.

Hasn't been regenerated since the update script could merge 32/64-bit checks.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365670 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Add tests for an alternative sequence for _mm_storel_pi/_mm_storeh_pi intrinsic...
Craig Topper [Wed, 10 Jul 2019 17:11:18 +0000 (17:11 +0000)]
[X86] Add tests for an alternative sequence for _mm_storel_pi/_mm_storeh_pi intrinsics. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365667 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[TargetLowering] support BlockAddress as "i" inline asm constraint
Nick Desaulniers [Wed, 10 Jul 2019 17:08:25 +0000 (17:08 +0000)]
[TargetLowering] support BlockAddress as "i" inline asm constraint

Summary:
This allows passing address of labels to inline assembly "i" input
constraints.

Fixes pr/42502.

Reviewers: ostannard

Reviewed By: ostannard

Subscribers: void, echristo, nathanchance, ostannard, javed.absar, hiraditya, llvm-commits, srhines

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64167

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365664 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[NFC][InstCombine] Fixup some tests in just-added "omit mask before left-shift" tests
Roman Lebedev [Wed, 10 Jul 2019 16:54:13 +0000 (16:54 +0000)]
[NFC][InstCombine] Fixup some tests in just-added "omit mask before left-shift" tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365663 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoMC: AArch64: Add support for pg_hi21_nc relocation specifier.
Peter Collingbourne [Wed, 10 Jul 2019 16:36:46 +0000 (16:36 +0000)]
MC: AArch64: Add support for pg_hi21_nc relocation specifier.

Differential Revision: https://reviews.llvm.org/D64455

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365661 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[CodeExtractor] Fix sinking of allocas with multiple bitcast uses (PR42451)
Vedant Kumar [Wed, 10 Jul 2019 16:32:20 +0000 (16:32 +0000)]
[CodeExtractor] Fix sinking of allocas with multiple bitcast uses (PR42451)

An alloca which can be sunk into the extraction region may have more
than one bitcast use. Move these uses along with the alloca to prevent
use-before-def.

Testing: check-llvm, stage2 build of clang

Fixes llvm.org/PR42451.

Differential Revision: https://reviews.llvm.org/D64463

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365660 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[CodeExtractor] Simplify findAllocas, NFC
Vedant Kumar [Wed, 10 Jul 2019 16:32:16 +0000 (16:32 +0000)]
[CodeExtractor] Simplify findAllocas, NFC

Split getLifetimeMarkers out into its own method and have it return a
struct.

Differential Revision: https://reviews.llvm.org/D64467

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365659 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoGlobalISel: Legalization for G_FMINNUM/G_FMAXNUM
Matt Arsenault [Wed, 10 Jul 2019 16:31:19 +0000 (16:31 +0000)]
GlobalISel: Legalization for G_FMINNUM/G_FMAXNUM

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365658 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoGlobalISel: Define the full family of FP min/max instructions
Matt Arsenault [Wed, 10 Jul 2019 16:31:15 +0000 (16:31 +0000)]
GlobalISel: Define the full family of FP min/max instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365657 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] EltsFromConsecutiveLoads - remove duplicate check for element size. NFCI.
Simon Pilgrim [Wed, 10 Jul 2019 16:22:31 +0000 (16:22 +0000)]
[X86] EltsFromConsecutiveLoads - remove duplicate check for element size. NFCI.

We've already checked that each element is the correct contributory size for VT when we inspect the elements for Undef/Zero/Load.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365656 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] EltsFromConsecutiveLoads - ensure element reg/store sizes are the same size...
Simon Pilgrim [Wed, 10 Jul 2019 16:14:26 +0000 (16:14 +0000)]
[X86] EltsFromConsecutiveLoads - ensure element reg/store sizes are the same size. NFCI.

This renames the type so it doesn't sound like its based off the load size - as we're moving towards supporting combining loads of different sizes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365655 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAMDGPU: Serialize mode from MachineFunctionInfo
Matt Arsenault [Wed, 10 Jul 2019 16:09:26 +0000 (16:09 +0000)]
AMDGPU: Serialize mode from MachineFunctionInfo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365653 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[PatternMatch] Generalize m_SpecificInt_ULT() to take ICmpInst::Predicate
Roman Lebedev [Wed, 10 Jul 2019 16:07:35 +0000 (16:07 +0000)]
[PatternMatch] Generalize m_SpecificInt_ULT() to take ICmpInst::Predicate

As discussed in the original review, this may be useful,
so let's just do it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365652 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Remarks] Add cl::Hidden to -remarks-yaml-string-table
Francis Visoiu Mistrih [Wed, 10 Jul 2019 15:46:36 +0000 (15:46 +0000)]
[Remarks] Add cl::Hidden to -remarks-yaml-string-table

It was showing up in a lot of unrelated tools.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365647 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agodocs/GithubMove.rst: Remove obsolete information
Tom Stellard [Wed, 10 Jul 2019 15:39:37 +0000 (15:39 +0000)]
docs/GithubMove.rst: Remove obsolete information

Summary:
Remove references to the multirepo and update the document to
reflect the current state of the github repository.

Reviewers: mehdi_amini, jyknight

Subscribers: jdoerfert, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D58420

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365645 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[FileCheck] Use bool operator to test Expected
Thomas Preud'homme [Wed, 10 Jul 2019 15:31:19 +0000 (15:31 +0000)]
[FileCheck] Use bool operator to test Expected

Use bool() consistently to get boolean value of Error, Optional and
Expected types in EXPECT calls. While static_cast is used in all cases
but one, bool provides more clarity and makes more sense as a new
default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365644 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[NFC][InstCombine] Redundant masking before left-shift (PR42563)
Roman Lebedev [Wed, 10 Jul 2019 15:08:06 +0000 (15:08 +0000)]
[NFC][InstCombine] Redundant masking before left-shift (PR42563)

alive proofs:
a,b:     https://rise4fun.com/Alive/4zsf
c,d,e,f: https://rise4fun.com/Alive/RC49

Indeed, not all of these patterns are canonical.
But since this fold will only produce a single instruction
i'm really interested in handling even uncanonical patterns.

Other than these 6 patterns, i can't think of any other
reasonable variants right now, although i'm sure they exist.

For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.

https://bugs.llvm.org/show_bug.cgi?id=42563

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365641 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AMDGPU] Allow abs/neg source modifiers on v_cndmask_b32
Jay Foad [Wed, 10 Jul 2019 14:53:47 +0000 (14:53 +0000)]
[AMDGPU] Allow abs/neg source modifiers on v_cndmask_b32

Summary:
D59191 added support for these modifiers in the assembler and
disassembler. This patch just teaches instruction selection that it can
use them.

Reviewers: arsenm, tstellar

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

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64497

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365640 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine] pow(C,x) -> exp2(log2(C)*x)
David Bolvansky [Wed, 10 Jul 2019 14:43:27 +0000 (14:43 +0000)]
[InstCombine] pow(C,x) -> exp2(log2(C)*x)

Summary:
Transform
pow(C,x)

To
exp2(log2(C)*x)

if C > 0, C != inf, C != NaN (and C is not power of 2, since we have some fold for such case already).

log(C) is folded by the compiler and exp2 is much faster to compute than pow.

Reviewers: spatel, efriedma, evandro

Reviewed By: evandro

Subscribers: lebedev.ri, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64099

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365637 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InferFunctionAttrs] add/adjust tests for dereferenceable; NFC
Sanjay Patel [Wed, 10 Jul 2019 14:41:47 +0000 (14:41 +0000)]
[InferFunctionAttrs] add/adjust tests for dereferenceable; NFC

Based on review comments for D64258.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365636 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agogn build: Merge r365585
Nico Weber [Wed, 10 Jul 2019 14:05:13 +0000 (14:05 +0000)]
gn build: Merge r365585

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365633 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[docs][llvm-symbolizer] Fix grammar
James Henderson [Wed, 10 Jul 2019 13:40:45 +0000 (13:40 +0000)]
[docs][llvm-symbolizer] Fix grammar

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365630 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] EltsFromConsecutiveLoads - cleanup Zero/Undef/Load element collection. NFCI.
Simon Pilgrim [Wed, 10 Jul 2019 13:28:13 +0000 (13:28 +0000)]
[X86] EltsFromConsecutiveLoads - cleanup Zero/Undef/Load element collection. NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365628 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[MIPS GlobalISel] Select float and double phi
Petar Avramovic [Wed, 10 Jul 2019 13:18:13 +0000 (13:18 +0000)]
[MIPS GlobalISel] Select float and double phi

Select float and double phi for MIPS32.

Differential Revision: https://reviews.llvm.org/D64420

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365627 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[MIPS GlobalISel] Select float and double load and store
Petar Avramovic [Wed, 10 Jul 2019 12:55:21 +0000 (12:55 +0000)]
[MIPS GlobalISel] Select float and double load and store

Select float and double load and store for MIPS32.

Differential Revision: https://reviews.llvm.org/D64419

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365626 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[FileCheck] Simplify numeric variable interface
Thomas Preud'homme [Wed, 10 Jul 2019 12:49:28 +0000 (12:49 +0000)]
[FileCheck] Simplify numeric variable interface

Summary:
This patch simplifies 2 aspects in the FileCheckNumericVariable code.

First, setValue() method is turned into a void function since being
called only on undefined variable is an invariant and is now asserted
rather than returned. This remove the assert from the callers.

Second, clearValue() method is also turned into a void function since
the only caller does not check its return value since it may be trying
to clear the value of variable that is already cleared without this
being noteworthy.

Reviewers: jhenderson, chandlerc, jdenny, probinson, grimar, arichardson, rnk

Subscribers: JonChesterfield, rogfer01, hfinkel, kristina, rnk, tra, arichardson, grimar, dblaikie, probinson, llvm-commits, hiraditya

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64231

llvm-svn: 365249

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365625 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[FileCheck] Fix @LINE value after match failure
Thomas Preud'homme [Wed, 10 Jul 2019 12:49:17 +0000 (12:49 +0000)]
[FileCheck] Fix @LINE value after match failure

Summary:
The value of the FileCheckNumericVariable class instance representing
the @LINE numeric variable is set and cleared respectively before and
after substitutions are made, if any. However, when a substitution
fails, the value is not cleared. This causes the next substitution of
@LINE later on to give the wrong value since setValue is a nop if the
value is already set. This is what caused failures after commit r365249.

Reviewers: jhenderson, chandlerc, jdenny, probinson, grimar, arichardson, rnk

Subscribers: hiraditya, llvm-commits, probinson, dblaikie, grimar, arichardson, tra, rnk, kristina, hfinkel, rogfer01, JonChesterfield

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64449

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365624 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[NFC][ARM] Convert lambdas to static helpers
Sam Parker [Wed, 10 Jul 2019 12:29:43 +0000 (12:29 +0000)]
[NFC][ARM] Convert lambdas to static helpers

Break up and convert some of the lambdas in ARMLowOverheadLoops into
static functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365623 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] EltsFromConsecutiveLoads - LDBase is non-null. NFCI.
Simon Pilgrim [Wed, 10 Jul 2019 12:22:59 +0000 (12:22 +0000)]
[X86] EltsFromConsecutiveLoads - LDBase is non-null. NFCI.

Don't bother checking for LDBase != null - it should be (and we assert that it is).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365622 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DAGCombine] visitINSERT_SUBVECTOR - use uint64_t subvector index. NFCI.
Simon Pilgrim [Wed, 10 Jul 2019 12:21:35 +0000 (12:21 +0000)]
[DAGCombine] visitINSERT_SUBVECTOR - use uint64_t subvector index. NFCI.

Keep the uint64_t type from getZExtValue() to stop truncation/extension overflow warnings in MSVC in subvector index math.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365621 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] EltsFromConsecutiveLoads - store Loads on a per-element basis. NFCI.
Simon Pilgrim [Wed, 10 Jul 2019 11:26:57 +0000 (11:26 +0000)]
[X86] EltsFromConsecutiveLoads - store Loads on a per-element basis. NFCI.

Cache the LoadSDNode nodes so we can easily map to/from the element index instead of packing them together - this will be useful for future patches for PR16739 etc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365620 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ELF] Loose a condition for relocation with a symbol
Nikola Prica [Wed, 10 Jul 2019 11:17:48 +0000 (11:17 +0000)]
[ELF] Loose a condition for relocation with a symbol

Deleted code was introduced as a work around for a bug in the gold linker
(http://sourceware.org/PR16794). Test case that was given as a reason for
this part of code, the one on previous link, now works for the gold.
This condition is too strict and when a code is compiled with debug info
it forces generation of numerous relocations with symbol for architectures
that do not have relocation addend.

Reviewers: arsenm, espindola

Reviewed By: MaskRay

Differential Revision: https://reviews.llvm.org/D64327

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365618 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86][SSE] EltsFromConsecutiveLoads - add basic dereferenceable support
Simon Pilgrim [Wed, 10 Jul 2019 10:46:36 +0000 (10:46 +0000)]
[X86][SSE] EltsFromConsecutiveLoads - add basic dereferenceable support

This patch checks to see if the vector element loads are based off a dereferenceable pointer that covers the entire vector width, in which case we don't need to have element loads at both extremes of the vector width - just the start (base pointer) of it.

Another step towards partial vector loads......

Differential Revision: https://reviews.llvm.org/D64205

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365614 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix const/non-const lambda return type warning. NFCI.
Simon Pilgrim [Wed, 10 Jul 2019 10:45:09 +0000 (10:45 +0000)]
Fix const/non-const lambda return type warning. NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365613 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix "result of 32-bit shift implicitly converted to 64 bits" warning. NFCI.
Simon Pilgrim [Wed, 10 Jul 2019 10:34:44 +0000 (10:34 +0000)]
Fix "result of 32-bit shift implicitly converted to 64 bits" warning. NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365612 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[SimpleLoopUnswitch] Don't consider unswitching `switch` insructions with one unique...
Serguei Katkov [Wed, 10 Jul 2019 10:25:22 +0000 (10:25 +0000)]
[SimpleLoopUnswitch] Don't consider unswitching `switch` insructions with one unique successor

Only instructions with two or more unique successors should be considered for unswitching.

Patch Author: Daniil Suchkov.

Reviewers: reames, asbirlea, skatkov
Reviewed By: skatkov
Subscribers: hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D64404

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365611 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ARM] Enable VPUSH/VPOP aliases when either MVE or VFP is present
Mikhail Maltsev [Wed, 10 Jul 2019 08:59:17 +0000 (08:59 +0000)]
[ARM] Enable VPUSH/VPOP aliases when either MVE or VFP is present

Summary:
Use the same predicates as VSTMDB/VLDMIA since VPUSH/VPOP alias to
these.

Patch by Momchil Velikov.

Reviewers: ostannard, simon_tatham, SjoerdMeijer, samparker, t.p.northover, dmgreen

Reviewed By: dmgreen

Subscribers: javed.absar, kristof.beyls, hiraditya, dmgreen, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64413

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365604 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[docs][llvm-dwarfdump] Normalise some wording
James Henderson [Wed, 10 Jul 2019 08:56:13 +0000 (08:56 +0000)]
[docs][llvm-dwarfdump] Normalise some wording

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365603 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[test/Object/nm-trivial-object.test] - Remove 4 precompiled binaries.
George Rimar [Wed, 10 Jul 2019 08:40:11 +0000 (08:40 +0000)]
[test/Object/nm-trivial-object.test] - Remove 4 precompiled binaries.

This converts 5 precompiled binaries to YAMLs,
removes 4 from inputs and performs a cleanup.

Differential revision: https://reviews.llvm.org/D64246

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365601 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[SimpleLoopUnswitch] Add a test case exposing a bug
Serguei Katkov [Wed, 10 Jul 2019 08:25:48 +0000 (08:25 +0000)]
[SimpleLoopUnswitch] Add a test case exposing a bug

This test exposes a bug in SimpleLoopUnswitch that leads to a crash on
assert(SuccessorsCount > 1 && "Cannot unswitch a condition without multiple distinct successors!");
when SimpleLoopUnswitch considers unswitching of a loop by a switch with one successor.

Fix will be submitted soon.

Patch Author: Daniil Suchkov.

Reviewers: reames, asbirlea, skatkov
Reviewed By: skatkov
Subscribers: zzheng, llvm-commits
Differential Revision: https://reviews.llvm.org/D64403

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365600 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoSilence gcc warning by adding parentheses to condition [NFC]
Mikael Holmen [Wed, 10 Jul 2019 06:18:03 +0000 (06:18 +0000)]
Silence gcc warning by adding parentheses to condition [NFC]

Without this gcc 7.4.0 complains with

  ../include/llvm/CodeGen/GlobalISel/LegalizationArtifactCombiner.h:457:54: error: suggest parentheses around '&&' within '||' [-Werror=parentheses]
                    isArtifactCast(TmpDef->getOpcode()) &&
                    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~
                        "Expecting copy or artifact cast here");
                        ~

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365597 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix modular build issues caused by BitCodes.h
Kristina Brooks [Wed, 10 Jul 2019 03:52:31 +0000 (03:52 +0000)]
Fix modular build issues caused by BitCodes.h

Consolidate llvm::BWH_* statics into an enum to fix
module build issues. This fixes the LLVM_Bitcode module,
getting rid of -Wmodules-ambiguous-internal-linkage.

Differential Revision: https://reviews.llvm.org/D64469

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365594 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Limit getTargetConstantFromNode to only work on NormalLoads not extending loads.
Craig Topper [Wed, 10 Jul 2019 00:40:01 +0000 (00:40 +0000)]
[X86] Limit getTargetConstantFromNode to only work on NormalLoads not extending loads.

This seems to fix a failure reported by Jordan Rupprecht, but we
don't have a reduced test case yet.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365589 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Support] Move llvm::MemoryBuffer to sys::fs::file_t
Reid Kleckner [Wed, 10 Jul 2019 00:34:13 +0000 (00:34 +0000)]
[Support] Move llvm::MemoryBuffer to sys::fs::file_t

Summary:
On Windows, Posix integer file descriptors are a compatibility layer
over native file handles provided by the C runtime. There is a hard
limit on the maximum number of file descriptors that a process can open,
and the limit is 8192. LLD typically doesn't run into this limit because
it opens input files, maps them into memory, and then immediately closes
the file descriptor. This prevents it from running out of FDs.

For various reasons, I'd like to open handles to every input file and
keep them open during linking. That requires migrating MemoryBuffer over
to taking open native file handles instead of integer FDs.

Reviewers: aganea, Bigcheese

Reviewed By: aganea

Subscribers: smeenai, silvas, mehdi_amini, hiraditya, steven_wu, dexonsmith, dang, llvm-commits, zturner

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63453

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365588 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAMDGPU/GlobalISel: Add support for wide loads >= 256-bits
Tom Stellard [Wed, 10 Jul 2019 00:22:41 +0000 (00:22 +0000)]
AMDGPU/GlobalISel: Add support for wide loads >= 256-bits

Summary:
This adds support for the most commonly used wide load types:
<8xi32>, <16xi32>, <4xi64>, and <8xi64>

Reviewers: arsenm

Reviewed By: arsenm

Subscribers: hiraditya, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, t-tye, volkan, Petar.Avramovic, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D57399

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365586 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoGlobalISel: Implement lower for G_FCOPYSIGN
Matt Arsenault [Tue, 9 Jul 2019 23:34:29 +0000 (23:34 +0000)]
GlobalISel: Implement lower for G_FCOPYSIGN

In SelectionDAG AMDGPU treated these as legal, but this was mostly
because the bitcasts required for FP types were painful. Theoretically
the bitpattern should eventually match to bfi, so don't bother trying
to get the patterns to import.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365583 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Bitcode] Explicitly include Bitstream/BitCodes.h and BitstreamWriter.h
Francis Visoiu Mistrih [Tue, 9 Jul 2019 23:20:01 +0000 (23:20 +0000)]
[Bitcode] Explicitly include Bitstream/BitCodes.h and BitstreamWriter.h

This fixes a modules issue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365580 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[docs][Remarks] Add documentation for remarks in LLVM
Francis Visoiu Mistrih [Tue, 9 Jul 2019 23:16:59 +0000 (23:16 +0000)]
[docs][Remarks] Add documentation for remarks in LLVM

This adds documentation that describes remarks in LLVM.

It aims at explaining what remarks are, how to enable them, and what
users can do with the different modes.

It lists all the available flags in LLVM (excluding clang), and
describes the expected YAML structure as well as the tools that support
the YAML format today.

Differential Revision: https://reviews.llvm.org/D64355

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365578 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Don't form extloads in combineExtInVec unless the load extension is legal.
Craig Topper [Tue, 9 Jul 2019 23:05:54 +0000 (23:05 +0000)]
[X86] Don't form extloads in combineExtInVec unless the load extension is legal.

This should prevent doing this on pre-sse4.1 targets or for 256
bit vectors without avx2.

I don't know of a failure from this. Op legalization will probably
take care of, but seemed better to be safe.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365577 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAMDGPU/GlobalISel: Fix legality for G_BUILD_VECTOR
Matt Arsenault [Tue, 9 Jul 2019 22:48:04 +0000 (22:48 +0000)]
AMDGPU/GlobalISel: Fix legality for G_BUILD_VECTOR

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365575 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AMDGPU] gfx908 v_pk_fmac_f16 support
Stanislav Mekhanoshin [Tue, 9 Jul 2019 22:42:24 +0000 (22:42 +0000)]
[AMDGPU] gfx908 v_pk_fmac_f16 support

Differential Revision: https://reviews.llvm.org/D64433

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365573 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agogn build: Merge r365536.
Peter Collingbourne [Tue, 9 Jul 2019 22:37:51 +0000 (22:37 +0000)]
gn build: Merge r365536.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365572 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agogn build: Merge r365532.
Peter Collingbourne [Tue, 9 Jul 2019 22:37:41 +0000 (22:37 +0000)]
gn build: Merge r365532.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365571 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agogn build: Merge r365541.
Peter Collingbourne [Tue, 9 Jul 2019 22:37:25 +0000 (22:37 +0000)]
gn build: Merge r365541.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365570 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agogn build: Merge r365531.
Peter Collingbourne [Tue, 9 Jul 2019 22:37:09 +0000 (22:37 +0000)]
gn build: Merge r365531.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365569 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAdd lldb type unit support to the release notes
Pavel Labath [Tue, 9 Jul 2019 22:36:43 +0000 (22:36 +0000)]
Add lldb type unit support to the release notes

Reviewers: JDevlieghere, teemperor

Subscribers: llvm-commits, lldb-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64366

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365568 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoGlobalISel: Combine unmerge of merge with intermediate cast
Matt Arsenault [Tue, 9 Jul 2019 22:19:13 +0000 (22:19 +0000)]
GlobalISel: Combine unmerge of merge with intermediate cast

This eliminates some illegal intermediate vectors when operations are
scalarized.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365566 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Profile] Support raw/indexed profiles larger than 4GB
Vedant Kumar [Tue, 9 Jul 2019 22:01:04 +0000 (22:01 +0000)]
[Profile] Support raw/indexed profiles larger than 4GB

rdar://45955976

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365565 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-objdump] Keep warning for --disassemble-functions in correct order.
Yuanfang Chen [Tue, 9 Jul 2019 21:53:33 +0000 (21:53 +0000)]
[llvm-objdump] Keep warning for --disassemble-functions in correct order.

relative to normal output when dumping archive files.

prepare for PR35351.

Reviewers: jhenderson, grimar, MaskRay, rupprecht

Reviewed by: MaskRay, jhenderson

Differential Revision: https://reviews.llvm.org/D64165

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365564 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AMDGPU] gfx908 mAI instructions, MC part
Stanislav Mekhanoshin [Tue, 9 Jul 2019 21:43:09 +0000 (21:43 +0000)]
[AMDGPU] gfx908 mAI instructions, MC part

Differential Revision: https://reviews.llvm.org/D64446

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365563 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[SLP] Optimize getSpillCost(); NFCI
Nikita Popov [Tue, 9 Jul 2019 20:24:44 +0000 (20:24 +0000)]
[SLP] Optimize getSpillCost(); NFCI

For a given set of live values, the spill cost will always be the
same for each call. Compute the cost once and multiply it by the
number of calls.

(I'm not sure this spill cost modeling makes sense if there are
multiple calls, as the spill cost will likely be shared across
calls in that case. But that's how it currently works.)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365552 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agohwasan: Improve precision of checks using short granule tags.
Peter Collingbourne [Tue, 9 Jul 2019 20:22:36 +0000 (20:22 +0000)]
hwasan: Improve precision of checks using short granule tags.

A short granule is a granule of size between 1 and `TG-1` bytes. The size
of a short granule is stored at the location in shadow memory where the
granule's tag is normally stored, while the granule's actual tag is stored
in the last byte of the granule. This means that in order to verify that a
pointer tag matches a memory tag, HWASAN must check for two possibilities:

* the pointer tag is equal to the memory tag in shadow memory, or
* the shadow memory tag is actually a short granule size, the value being loaded
  is in bounds of the granule and the pointer tag is equal to the last byte of
  the granule.

Pointer tags between 1 to `TG-1` are possible and are as likely as any other
tag. This means that these tags in memory have two interpretations: the full
tag interpretation (where the pointer tag is between 1 and `TG-1` and the
last byte of the granule is ordinary data) and the short tag interpretation
(where the pointer tag is stored in the granule).

When HWASAN detects an error near a memory tag between 1 and `TG-1`, it
will show both the memory tag and the last byte of the granule. Currently,
it is up to the user to disambiguate the two possibilities.

Because this functionality obsoletes the right aligned heap feature of
the HWASAN memory allocator (and because we can no longer easily test
it), the feature is removed.

Also update the documentation to cover both short granule tags and
outlined checks.

Differential Revision: https://reviews.llvm.org/D63908

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365551 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[PoisonChecking] Flesh out complete todo list for full coverage
Philip Reames [Tue, 9 Jul 2019 19:59:39 +0000 (19:59 +0000)]
[PoisonChecking] Flesh out complete todo list for full coverage

Note: I don't actually plan to implement all of the cases at the moment, I'm just documenting them for completeness.  There's a couple of cases left which are practically useful for me in debugging loop transforms, and I'll probably stop there for the moment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365550 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86][AMDGPU][DAGCombiner] Move call to allowsMemoryAccess into isLoadBitCastBenefici...
Craig Topper [Tue, 9 Jul 2019 19:55:28 +0000 (19:55 +0000)]
[X86][AMDGPU][DAGCombiner] Move call to allowsMemoryAccess into isLoadBitCastBeneficial/isStoreBitCastBeneficial to allow X86 to bypass it

Basically the problem is that X86 doesn't set the Fast flag from
allowsMemoryAccess on certain CPUs due to slow unaligned memory
subtarget features. This prevents bitcasts from being folded into
loads and stores. But all vector loads and stores of the same width
are the same cost on X86.

This patch merges the allowsMemoryAccess call into isLoadBitCastBeneficial to allow X86 to skip it.

Differential Revision: https://reviews.llvm.org/D64295

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@365549 91177308-0d34-0410-b5e6-96231b3b80d8