OSDN Git Service

android-x86/external-llvm.git
6 years ago[Dominators] Teach LoopUnswitch to use the incremental API
Jakub Kuderski [Thu, 17 Aug 2017 16:45:35 +0000 (16:45 +0000)]
[Dominators] Teach LoopUnswitch to use the incremental API

Summary:
This patch makes LoopUnswitch use new incremental API for updating dominators.
It also updates SplitCriticalEdge, as it is called in LoopUnswitch.

There doesn't seem to be any noticeable performance difference when bootstrapping clang with this patch.

Reviewers: dberlin, davide, sanjoy, grosser, chandlerc

Reviewed By: davide, grosser

Subscribers: mzolotukhin, llvm-commits

Differential Revision: https://reviews.llvm.org/D35528

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311093 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AVX512] Don't switch unmasked subvector insert/extract instructions when AVX512DQI...
Craig Topper [Thu, 17 Aug 2017 15:40:25 +0000 (15:40 +0000)]
[AVX512] Don't switch unmasked subvector insert/extract instructions when AVX512DQI is enabled.

There's no reason to switch instructions with and without DQI. It just creates extra isel patterns and test divergences.

There is however value in enabling the masked version of the instructions with DQI.

This required introducing some new multiclasses to enabling this splitting.

Differential Revision: https://reviews.llvm.org/D36661

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311091 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Remove memopmmx pattern fragment
Craig Topper [Thu, 17 Aug 2017 15:25:05 +0000 (15:25 +0000)]
[X86] Remove memopmmx pattern fragment

Summary: Just like the FIXME says, there is no alignment requirement for MMX.

Reviewers: RKSimon, zvi, igorb

Reviewed By: RKSimon

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D36815

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311090 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoMark Verifier/invalid-eh.ll as unsupported on windows
Victor Leschuk [Thu, 17 Aug 2017 15:07:03 +0000 (15:07 +0000)]
Mark Verifier/invalid-eh.ll as unsupported on windows

Mark this unsupported for now as it causes tests hangs on buildbot.
Will place it back when the problem is debugged.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311089 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[dfsan] Add explicit zero extensions for shadow parameters in function wrappers.
Simon Dardis [Thu, 17 Aug 2017 14:14:25 +0000 (14:14 +0000)]
[dfsan] Add explicit zero extensions for shadow parameters in function wrappers.

In the case where dfsan provides a custom wrapper for a function,
shadow parameters are added for each parameter of the function.
These parameters are i16s. For targets which do not consider this
a legal type, the lack of sign extension information would cause
LLVM to generate anyexts around their usage with phi variables
and calling convention logic.

Address this by introducing zero exts for each shadow parameter.

Reviewers: pcc, slthakur

Differential Revision: https://reviews.llvm.org/D33349

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311087 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[globalisel][tablegen] Generate TypeObject table. NFC
Daniel Sanders [Thu, 17 Aug 2017 13:18:35 +0000 (13:18 +0000)]
[globalisel][tablegen] Generate TypeObject table. NFC

Summary:
Generate the type table from the types used by a target rather than hard-coding
the union of types used by all targets.

Depends on D36084

Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar

Reviewed By: rovka

Subscribers: kristof.beyls, igorb, llvm-commits

Differential Revision: https://reviews.llvm.org/D36085

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311084 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[DAGCombiner] Add support for non-uniform constant vectors to (mul x, (1 << c)) ...
Simon Pilgrim [Thu, 17 Aug 2017 13:03:34 +0000 (13:03 +0000)]
[DAGCombiner] Add support for non-uniform constant vectors to (mul x, (1 << c)) -> x << c

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311083 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Refactoring of X86TargetLowering::EmitLoweredSelect. NFC.
Amjad Aboud [Thu, 17 Aug 2017 12:12:30 +0000 (12:12 +0000)]
[X86] Refactoring of X86TargetLowering::EmitLoweredSelect. NFC.

Authored by aivchenk
Differential Revision: https://reviews.llvm.org/D35685

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311082 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Verifier] Avoid visiting DIGlobalVariables twice.
Davide Italiano [Thu, 17 Aug 2017 11:32:21 +0000 (11:32 +0000)]
[Verifier] Avoid visiting DIGlobalVariables twice.

We currently visit them twice.
Once, through `visitMDNode()` -> (the code generated by)
  `../include/llvm/IR/Metadata.def:109` -> `visitDIGlobalVariable()`
Then, through `visitMDNode()` -> `visitDIGlobalVariableExpression()`
  -> `visitDIGlobalVariable()`

This results in verification failures printed twice, e.g.:

  $ ./opt -verify ../../test/DebugInfo/pr34186.ll
  missing global variable type
  !4 = distinct !DIGlobalVariable(name: "pat", scope: !0,
    file: !1, line: 27, isLocal: true, isDefinition: true)
  missing global variable type
  !4 = distinct !DIGlobalVariable(name: "pat", scope: !0,
    file: !1, line: 27, isLocal: true, isDefinition: true)
  ./opt: ../../test/DebugInfo/pr34186.ll: error: input module is broken!

The patch removes one call so we ensure each GV is visited exactly once.

Differential Revision:  https://reviews.llvm.org/D36797

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311081 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LV] Using VPlan to model the vectorized code and drive its transformation
Ayal Zaks [Thu, 17 Aug 2017 09:29:59 +0000 (09:29 +0000)]
[LV] Using VPlan to model the vectorized code and drive its transformation

VPlan is an ongoing effort to refactor and extend the Loop Vectorizer. This
patch introduces the VPlan model into LV and uses it to represent the vectorized
code and drive the generation of vectorized IR.

In this patch VPlan models the vectorized loop body: the vectorized control-flow
is represented using VPlan's Hierarchical CFG, with predication refactored from
being a post-vectorization-step into a vectorization planning step modeling
if-then VPRegionBlocks, and generating code inline with non-predicated code. The
vectorized code within each VPBasicBlock is represented as a sequence of
Recipes, each responsible for modelling and generating a sequence of IR
instructions. To keep the size of this commit manageable the Recipes in this
patch are coarse-grained and capture large chunks of LV's code-generation logic.
The constructed VPlans are dumped in dot format under -debug.

This commit retains current vectorizer output, except for minor instruction
reorderings; see associated modifications to lit tests.

For further details on the VPlan model see docs/Proposals/VectorizationPlan.rst
and its references.

Authors: Gil Rapaport and Ayal Zaks

Differential Revision: https://reviews.llvm.org/D32871

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311077 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRe-commit: [globalisel][tablegen] Support zero-instruction emission.
Daniel Sanders [Thu, 17 Aug 2017 09:26:14 +0000 (09:26 +0000)]
Re-commit: [globalisel][tablegen] Support zero-instruction emission.

Summary:
Support the case where an operand of a pattern is also the whole of the
result pattern. In this case the original result and all its uses must be
replaced by the operand. However, register class restrictions can require
a COPY. This patch handles both cases by always emitting the copy and
leaving it for the register allocator to optimize.

The previous commit failed on Windows machines due to a flaw in the sort
predicate which allowed both A < B < C and B == C to be satisfied
simultaneously. The cause of this was some sloppiness in the priority order of
G_CONSTANT instructions compared to other instructions. These had equal priority
because it makes no difference, however there were operands had higher priority
than G_CONSTANT but lower priority than any other instruction. As a result, a
priority order between G_CONSTANT and other instructions must be enforced to
ensure the predicate defines a strict weak order.

Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar

Subscribers: javed.absar, kristof.beyls, igorb, llvm-commits

Differential Revision: https://reviews.llvm.org/D36084

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311076 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SystemZ] Also wrap TII with #ifndef NDEBUG in constructor initilizer list.
Jonas Paulsson [Thu, 17 Aug 2017 09:18:02 +0000 (09:18 +0000)]
[SystemZ]  Also wrap TII with #ifndef NDEBUG in constructor initilizer list.

TII needs to be wrapped with #ifndef NDEBUG to silece compiler warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311075 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SystemZ] Add a wrapping with #ifndef NDEBUG to silence warning.
Jonas Paulsson [Thu, 17 Aug 2017 08:56:09 +0000 (08:56 +0000)]
[SystemZ]  Add a wrapping with #ifndef NDEBUG to silence warning.

SystemZHazardRecognizer::TII is only used for debug output, so it needs
also to be wrapped with #ifndef NDEBUG.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311074 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SystemZ, MachineScheduler] Improve post-RA scheduling.
Jonas Paulsson [Thu, 17 Aug 2017 08:33:44 +0000 (08:33 +0000)]
[SystemZ, MachineScheduler]  Improve post-RA scheduling.

The idea of this patch is to continue the scheduler state over an MBB boundary
in the case where the successor block has only one predecessor. This means
that the scheduler will continue in the successor block (after emitting any
branch instructions) with e.g. maintained processor resource counters.
Benchmarks have been confirmed to benefit from this.

The algorithm in MachineScheduler.cpp that extracts scheduling regions of an
MBB has been extended so that the strategy may optionally reverse the order
of processing the regions themselves. This is controlled by a new method
doMBBSchedRegionsTopDown(), which defaults to false.

Handling the top-most region of an MBB first also means that a top-down
scheduler can continue the scheduler state across any scheduling boundary
between to regions inside MBB.

Review: Ulrich Weigand, Matthias Braun, Andy Trick.
https://reviews.llvm.org/D35053

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311072 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SelectionDAG] Teach the vector-types operand scalarizer about SETCC
Elad Cohen [Thu, 17 Aug 2017 08:06:36 +0000 (08:06 +0000)]
[SelectionDAG] Teach the vector-types operand scalarizer about SETCC

When v1i1 is legal (e.g. AVX512) the legalizer can reach
a case where a v1i1 SETCC with an illgeal vector type operand
wasn't scalarized (since v1i1 is legal) but its operands does
have to be scalarized. This used to assert because SETCC was
missing from the vector operand scalarizer.

This patch attemps to teach the legalizer to handle these cases
by scalazring the operands, converting the node into a scalar
SETCC node.

Differential revision: https://reviews.llvm.org/D36651

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311071 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-dlltool] Improve an error message when unable to open files. NFC.
Martin Storsjo [Thu, 17 Aug 2017 06:26:42 +0000 (06:26 +0000)]
[llvm-dlltool] Improve an error message when unable to open files. NFC.

Differential Revision: https://reviews.llvm.org/D36818

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311069 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-dlltool] Don't crash if no def file is provided or it can't be opened
Martin Storsjo [Thu, 17 Aug 2017 05:58:27 +0000 (05:58 +0000)]
[llvm-dlltool] Don't crash if no def file is provided or it can't be opened

Differential Revision: https://reviews.llvm.org/D36780

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311068 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CGP] Fix the rematerialization of gc.relocates
Serguei Katkov [Thu, 17 Aug 2017 05:48:30 +0000 (05:48 +0000)]
[CGP] Fix the rematerialization of gc.relocates

If we want to substitute the relocation of derived pointer with gep of base then
we must ensure that relocation of base dominates the relocation of derived pointer.

Currently only check for basic block is present. However it is possible that both
relocation are in the same basic block but relocation of derived pointer is defined
earlier.

The patch moves the relocation of base pointer right before relocation of derived
pointer in this case.

Reviewers: sanjoy,artagnon,igor-laevsky,reames
Reviewed By: reames
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D36462

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311067 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert "[MachineCopyPropagation] Extend pass to do COPY source forwarding"
Geoff Berry [Thu, 17 Aug 2017 04:04:11 +0000 (04:04 +0000)]
Revert "[MachineCopyPropagation] Extend pass to do COPY source forwarding"

This reverts commit r311038.

Several buildbots are breaking, and at least one appears to be due to
the forwarding of physical regs enabled by this change.  Reverting while
I investigate further.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311062 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoARM: mark CPSR as clobbered for Windows VLAs
Saleem Abdulrasool [Thu, 17 Aug 2017 02:42:24 +0000 (02:42 +0000)]
ARM: mark CPSR as clobbered for Windows VLAs

When lowering a VLA, we emit a __chstk call.  However, this call can
internally clobber CPSR.  We did not mark this register as an ImpDef,
which could potentially allow a comparison to be hoisted above the call
to `__chkstk`.  In such a case, the CPSR could be clobbered, and the
check invalidated.  When the support was initially added, it seemed that
the call would take care of preventing CPSR from being clobbered, but
this is not the case.  Mark the register as clobbered to fix a possible
state corruption.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311061 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Exchange the memory op predicate for PALIGNR/VPALIGNR. I accidentally swapped...
Craig Topper [Thu, 17 Aug 2017 02:34:35 +0000 (02:34 +0000)]
[X86] Exchange the memory op predicate for PALIGNR/VPALIGNR. I accidentally swapped them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311060 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Cleanup multiclasses for SSE/AVX2 PALIGNR. Add missing load patterns.
Craig Topper [Thu, 17 Aug 2017 01:48:03 +0000 (01:48 +0000)]
[X86] Cleanup multiclasses for SSE/AVX2 PALIGNR. Add missing load patterns.

We used to have a separate multiclass for AVX2 and SSE/AVX. Now we have one multiclass and pass the relevant differences.

We were also missing load patterns, though we had them for the AVX-512 version.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311059 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Remove patterns for PALIGNR with non-vXi8 types.
Craig Topper [Thu, 17 Aug 2017 01:48:00 +0000 (01:48 +0000)]
[X86] Remove patterns for PALIGNR with non-vXi8 types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311058 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoReapply: [ADCE][Dominators] Teach ADCE to preserve dominators
Jakub Kuderski [Thu, 17 Aug 2017 01:41:49 +0000 (01:41 +0000)]
Reapply: [ADCE][Dominators] Teach ADCE to preserve dominators

Summary:
This patch teaches ADCE to preserve both DominatorTrees and PostDominatorTrees.

I didn't notice any performance impact when bootstrapping clang with this patch.

The patch was originally committed in r311039 and reverted in r311049.
This revision fixes the problem with not adding a dependency on the
DominatorTreeWrapperPass for the LegacyPassManager.

Reviewers: dberlin, chandlerc, sanjoy, davide, grosser, brzycki

Reviewed By: davide

Subscribers: grandinj, zhendongsu, llvm-commits, david2050

Differential Revision: https://reviews.llvm.org/D35869

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311057 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Put multiclass closer to its use and simplify slightly. NFC
Craig Topper [Wed, 16 Aug 2017 23:38:25 +0000 (23:38 +0000)]
[X86] Put multiclass closer to its use and simplify slightly. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311055 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Use a static array instead of a SmallVector for a small fixed size array. NFC
Craig Topper [Wed, 16 Aug 2017 23:16:43 +0000 (23:16 +0000)]
[X86] Use a static array instead of a SmallVector for a small fixed size array. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311054 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[x86] add cmov promotion tests for D36711; NFC
Sanjay Patel [Wed, 16 Aug 2017 22:50:11 +0000 (22:50 +0000)]
[x86] add cmov promotion tests for D36711; NFC

This way we can see what the current codegen looks like.
I've also explicitly added/removed the cmov attribute from the RUN lines,
so we know exactly what we're checking in the runs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311052 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstCombine] Teach canEvaluateTruncated to handle arithmetic shift (including those...
Amjad Aboud [Wed, 16 Aug 2017 22:42:38 +0000 (22:42 +0000)]
[InstCombine] Teach canEvaluateTruncated to handle arithmetic shift (including those with vector splat shift amount)

Differential Revision: https://reviews.llvm.org/D36784

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311050 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert "[ADCE][Dominators] Teach ADCE to preserve dominators"
Jakub Kuderski [Wed, 16 Aug 2017 22:10:53 +0000 (22:10 +0000)]
Revert "[ADCE][Dominators] Teach ADCE to preserve dominators"

This reverts commit r311039. The patch caused the
`test/Bindings/OCaml/Output/scalar_opts.ml` to fail.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311049 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Analysis] Fix some Clang-tidy modernize and Include What You Use warnings; other...
Eugene Zelenko [Wed, 16 Aug 2017 22:07:40 +0000 (22:07 +0000)]
[Analysis] Fix some Clang-tidy modernize and  Include What You Use warnings; other minor fixes (NFC).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311048 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstCombine] Make folding (X >s -1) ? C1 : C2 --> ((X >>s 31) & (C2 - C1)) + C1...
Craig Topper [Wed, 16 Aug 2017 21:52:07 +0000 (21:52 +0000)]
[InstCombine] Make folding (X >s -1) ? C1 : C2 --> ((X >>s 31) & (C2 - C1)) + C1 support splat vectors

This also uses decomposeBitTestICmp to decode the compare.

Differential Revision: https://reviews.llvm.org/D36781

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311044 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ADCE][Dominators] Teach ADCE to preserve dominators
Jakub Kuderski [Wed, 16 Aug 2017 20:50:23 +0000 (20:50 +0000)]
[ADCE][Dominators] Teach ADCE to preserve dominators

Summary:
This patch teaches ADCE to preserve both DominatorTrees and PostDominatorTrees.

I didn't notice any performance impact when bootstrapping clang with this patch.

Reviewers: dberlin, chandlerc, sanjoy, davide, grosser, brzycki

Reviewed By: davide

Subscribers: grandinj, zhendongsu, llvm-commits, david2050

Differential Revision: https://reviews.llvm.org/D35869

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311039 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MachineCopyPropagation] Extend pass to do COPY source forwarding
Geoff Berry [Wed, 16 Aug 2017 20:50:01 +0000 (20:50 +0000)]
[MachineCopyPropagation] Extend pass to do COPY source forwarding

This change extends MachineCopyPropagation to do COPY source forwarding.

This change also extends the MachineCopyPropagation pass to be able to
be run during register allocation, after physical registers have been
assigned, but before the virtual registers have been re-written, which
allows it to remove virtual register COPY LiveIntervals that become dead
through the forwarding of all of their uses.

Reviewers: qcolombet, javed.absar, MatzeB, jonpa

Subscribers: jyknight, nemanjai, llvm-commits, nhaehnle, mcrosier, mgorny

Differential Revision: https://reviews.llvm.org/D30751

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311038 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CMake][runtimes] Support for building target variants
Petr Hosek [Wed, 16 Aug 2017 19:13:45 +0000 (19:13 +0000)]
[CMake][runtimes] Support for building target variants

This can be used to build non-sanitized and sanitized versions of
runtimes, where sanitized versions use the just built sanitizer
which in turn may use the non-sanitized version.

Differential Revision: https://reviews.llvm.org/D36348

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311036 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LoopDataPrefetch][AArch64FalkorHWPFFix] Preserve ScalarEvolution
Geoff Berry [Wed, 16 Aug 2017 19:03:16 +0000 (19:03 +0000)]
[LoopDataPrefetch][AArch64FalkorHWPFFix] Preserve ScalarEvolution

Summary:
Mark LoopDataPrefetch and AArch64FalkorHWPFFix passes as preserving
ScalarEvolution since they do not alter loop structure and should not
alter any SCEV values (though LoopDataPrefetch may introduce new
instructions that won't have cached SCEV values yet).

This can result in slight code differences, mainly w.r.t. nsw/nuw flags
on SCEVs, since these are computed somewhat lazily when a zext/sext
instruction is encountered.  As a result, passes after the modified
passes may see SCEVs with more nsw/nuw flags present.

Reviewers: sanjoy, anemet

Subscribers: aemerson, rengolin, mzolotukhin, javed.absar, kristof.beyls, mcrosier, llvm-commits

Differential Revision: https://reviews.llvm.org/D36716

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311032 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips] Handle R_MIPS_TLS_DTPREL32/64 relocations in the RelocVisitor
Simon Atanasyan [Wed, 16 Aug 2017 19:01:22 +0000 (19:01 +0000)]
[mips] Handle R_MIPS_TLS_DTPREL32/64 relocations in the RelocVisitor

Debug information for TLS variables on MIPS might have R_MIPS_TLS_DTPREL32
or R_MIPS_TLS_DTPREL64 relocations. This patch adds a support for such
relocations in the `RelocVisitor`.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311031 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd a convenience overload of DWARFDie::dump() for debugging purposes.
Adrian Prantl [Wed, 16 Aug 2017 17:43:01 +0000 (17:43 +0000)]
Add a convenience overload of DWARFDie::dump() for debugging purposes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311026 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd more comment
Xinliang David Li [Wed, 16 Aug 2017 17:33:43 +0000 (17:33 +0000)]
Add more comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311025 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[PGO] Fix ThinLTO crash
Xinliang David Li [Wed, 16 Aug 2017 17:18:01 +0000 (17:18 +0000)]
[PGO] Fix ThinLTO crash

Differential Revsion: http://reviews.llvm.org/D36640

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311023 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AMDGPU] NFC: test commit
Evgeny Mankov [Wed, 16 Aug 2017 16:47:29 +0000 (16:47 +0000)]
[AMDGPU] NFC: test commit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311019 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAMDGPU/NFC: Sort files in CMakeLists.txt alphabetically
Konstantin Zhuravlyov [Wed, 16 Aug 2017 16:23:32 +0000 (16:23 +0000)]
AMDGPU/NFC: Sort files in CMakeLists.txt alphabetically

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311017 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Regenerate immediate store merging tests
Simon Pilgrim [Wed, 16 Aug 2017 16:22:19 +0000 (16:22 +0000)]
[X86] Regenerate immediate store merging tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311016 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Dominators] Introduce batch updates
Jakub Kuderski [Wed, 16 Aug 2017 16:12:52 +0000 (16:12 +0000)]
[Dominators] Introduce batch updates

Summary:
This patch introduces a way of informing the (Post)DominatorTree about multiple CFG updates that happened since the last tree update. This makes performing tree updates much easier, as it internally takes care of applying the updates in lockstep with the (virtual) updates to the CFG, which is done by reverse-applying future CFG updates.

The batch updater is able to remove redundant updates that cancel each other out. In the future, it should be also possible to reorder updates to reduce the amount of work needed to perform the updates.

Reviewers: dberlin, sanjoy, grosser, davide, brzycki

Reviewed By: brzycki

Subscribers: mgorny, llvm-commits

Differential Revision: https://reviews.llvm.org/D36167

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311015 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[BDCE] Don't check demanded bits on unsized types
Hal Finkel [Wed, 16 Aug 2017 16:09:22 +0000 (16:09 +0000)]
[BDCE] Don't check demanded bits on unsized types

To clear assumptions that are potentially invalid after trivialization, we need
to walk the use/def chain. Normally, the only way to reach an instruction with
an unsized type is via an instruction that has side effects (or otherwise will
demand its input bits). That would stop the walk. However, if we have a
readnone function that returns an unsized type (e.g., void), we must avoid
asking for the demanded bits of the function call's return value. A
void-returning readnone function is always dead (and so we can stop walking the
use/def chain here), but the check is necessary to avoid asserting.

Fixes PR34211.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311014 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Verifier] Reject globals without a type associated.
Davide Italiano [Wed, 16 Aug 2017 15:16:33 +0000 (15:16 +0000)]
[Verifier] Reject globals without a type associated.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311012 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AMDGPU][MC][GFX9] Added op_sel support for v_mad_*16, v_fma_f16, v_div_fixup_f16
Dmitry Preobrazhensky [Wed, 16 Aug 2017 15:16:32 +0000 (15:16 +0000)]
[AMDGPU][MC][GFX9] Added op_sel support for v_mad_*16, v_fma_f16, v_div_fixup_f16

This change implements features postponed in https://reviews.llvm.org/D35424 because of a dependency on https://reviews.llvm.org/D36322

Reviewers: SamWot, artem.tamazov, arsenm

Differential Revision: https://reviews.llvm.org/D36694

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311011 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[DemandedBits] simplify call; NFC
Sanjay Patel [Wed, 16 Aug 2017 14:28:23 +0000 (14:28 +0000)]
[DemandedBits] simplify call; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311009 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert "MachineInstr: Reason locally about some memory objects before going to AA."
Balaram Makam [Wed, 16 Aug 2017 14:17:43 +0000 (14:17 +0000)]
Revert "MachineInstr: Reason locally about some memory objects before going to AA."

r310825 caused the clang-ppc64le-linux-lnt bot to go red
(http://lab.llvm.org:8011/builders/clang-ppc64le-linux-lnt/builds/5712)
because of a test-suite failure of
SingleSource/UnitTests/2003-07-09-SignedArgs

This reverts commit 0028f6a87224fb595a1c19c544cde9b003035996.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311008 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AMDGPU][MC][GFX9] Added integer clamping support for VOP3 opcodes
Dmitry Preobrazhensky [Wed, 16 Aug 2017 13:51:56 +0000 (13:51 +0000)]
[AMDGPU][MC][GFX9] Added integer clamping support for VOP3 opcodes

See Bug 34152: https://bugs.llvm.org//show_bug.cgi?id=34152

Reviewers: SamWot, artem.tamazov, arsenm

Differential Revision: https://reviews.llvm.org/D36674

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311006 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CostModel][X86][XOP] Improve costs for XOP shuffles
Simon Pilgrim [Wed, 16 Aug 2017 13:50:20 +0000 (13:50 +0000)]
[CostModel][X86][XOP] Improve costs for XOP shuffles

VPPERM/VPERMIL2PD/VPERMIL2PS all provide more effective 2-input shuffles than regular AVX instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311005 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[DI] Every DIGlobalVariable should have a type.
Davide Italiano [Wed, 16 Aug 2017 13:39:07 +0000 (13:39 +0000)]
[DI] Every DIGlobalVariable should have a type.

I'll make this a verifier check to catch other violations. This
commit fixes the tests already in tree.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311004 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips] Handle variables with an explicit section and interactions with .sdata, .sbss
Simon Dardis [Wed, 16 Aug 2017 12:18:04 +0000 (12:18 +0000)]
[mips] Handle variables with an explicit section and interactions with .sdata, .sbss

If a variable has an explicit section such as .sdata or .sbss, it is placed
in that section and accessed in a gp relative manner. This overrides the global
-G setting.

Otherwise if a variable has a explicit section attached to it, such as '.rodata'
or '.mysection', it is not placed in the small data section. This also overrides
the global -G setting.

Reviewers: atanasyan, nitesh.jain

Differential Revision: https://reviews.llvm.org/D36616

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@311001 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM] Improve loop unrolling for Cortex-M
Sam Parker [Wed, 16 Aug 2017 07:42:44 +0000 (07:42 +0000)]
[ARM] Improve loop unrolling for Cortex-M

- Set the default runtime unroll count to 4 and use the newly added
  UnrollRemainder option.
- Create loop cost and force unroll for a cost less than 12.
- Disable unrolling on Thumb1 only targets.

Differential Revision: https://reviews.llvm.org/D36134

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310997 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[GlobalISel][X86] Fix mir tests, use correct physical register.NFC.
Igor Breger [Wed, 16 Aug 2017 07:25:51 +0000 (07:25 +0000)]
[GlobalISel][X86] Fix mir tests, use correct physical register.NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310996 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[COFF] Make the weak aliases optional
Martin Storsjo [Wed, 16 Aug 2017 05:22:49 +0000 (05:22 +0000)]
[COFF] Make the weak aliases optional

When creating an import library from lld, the cases with
Name != ExtName shouldn't end up as a weak alias, but as a real
export of the new name, which is what actually is exported from
the DLL.

This restores the behaviour of renamed exports to what it was in
4.0.

The other half of this commit, including test, goes into lld.

Differential Revision: https://reviews.llvm.org/D36633

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310991 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-dlltool] Fix creating stdcall/fastcall import libraries for i386
Martin Storsjo [Wed, 16 Aug 2017 05:18:36 +0000 (05:18 +0000)]
[llvm-dlltool] Fix creating stdcall/fastcall import libraries for i386

Hook up the -k option (that in the original GNU dlltool removes the
@n suffix from the symbol that the final executable ends up linked to).

In llvm-dlltool, make sure that functions end up with the undecorate
name type if this option is set and they are decorated. In mingw, when
creating import libraries from def files instead of creating an import
library as a side effect of linking a DLL, the symbol names in the def
contain the stdcall/fastcall decoration (but no leading underscore).

By setting the undecorate name type, a linker linking to the import
library will omit the decoration from the DLL import entry.

With this in place, mingw-w64 for i386 built with llvm-dlltool/clang
produces import libraries that actually work.

Differential Revision: https://reviews.llvm.org/D36548

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310990 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[COFF] Add SymbolName as a distinct field in COFFImportFile
Martin Storsjo [Wed, 16 Aug 2017 05:13:16 +0000 (05:13 +0000)]
[COFF] Add SymbolName as a distinct field in COFFImportFile

The previous Name and ExtName aren't enough to convey all the nuances
between weak aliases and stdcall decorated function names.

A test for this will be added in LLD.

Differential Revision: https://reviews.llvm.org/D36544

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310988 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[AMDGPU] Eliminate no effect instructions before s_endpgm
Stanislav Mekhanoshin [Wed, 16 Aug 2017 04:43:49 +0000 (04:43 +0000)]
[AMDGPU] Eliminate no effect instructions before s_endpgm

Differential Revision: https://reviews.llvm.org/D36585

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310987 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoMerge debug info when hoist then-else code to if.
Dehao Chen [Wed, 16 Aug 2017 01:55:26 +0000 (01:55 +0000)]
Merge debug info when hoist then-else code to if.

Summary: When we move then-else code to if, we need to merge its debug info, otherwise the hoisted instruction may have inaccurate debug info attached.

Reviewers: aprantl, probinson, dblaikie, echristo, loladiro

Reviewed By: aprantl

Subscribers: sanjoy, llvm-commits

Differential Revision: https://reviews.llvm.org/D36778

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310985 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[WebAssembly] Remove infinite loop from reg-stackify test
Derek Schuff [Wed, 16 Aug 2017 00:49:44 +0000 (00:49 +0000)]
[WebAssembly] Remove infinite loop from reg-stackify test

r310940 exposed reverse-unreachable code to some optimizers,
which caused some of the code in this test to be sunk, changing
the input to the pass and breaking the exptectations.

Since that change is irrelevant to this particular test, this change
just adds an exit node to work around the problem; the
test should really be more robust (or be an MIR test?) but this preserves
the existing test intent.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310981 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[VirtRegRewriter] Properly model the register liveness on undef subreg definition
Quentin Colombet [Wed, 16 Aug 2017 00:17:05 +0000 (00:17 +0000)]
[VirtRegRewriter] Properly model the register liveness on undef subreg definition

Undef subreg definition means that the content of the super register
doesn't matter at this point. While that's true for virtual registers,
this may not hold when replacing them with actual physical registers.
Indeed, some part of the physical register may be coalesced with the
related virtual register and thus, the values for those parts matter and
must be live.

The fix consists in checking whether or not subregs of the physical register
being assigned to an undef subreg definition are live through that def and
insert an implicit use if they are. Doing so, will keep them alive until
that point like they should be.

E.g., let vreg14 being assigned to R0_R1 then
%vreg14:gsub_0<def,read-undef> = COPY %R0 ; <-- R1 is still live here
%vreg14:gsub_1<def> = COPY %R1

Before this changes, the rewriter would change the code into:
%R0<def> = KILL %R0, %R0_R1<imp-def> ; <-- this tells R1 is redefined
%R1<def> = KILL %R1, %R0_R1<imp-def>, %R0_R1<imp-use> ; this value of this R1
                                                      ; is believed to come
                                                      ; from the previous
                                                      ; instruction

Because of this invalid liveness, later pass could make wrong choices and in
particular clobber live register as it happened with the register scavenger in
llvm.org/PR34107

Now we would generate:
%R0<def> = KILL %R0, %R0_R1<imp-def>, %R0_R1<imp-use> ; This tells R1 needs to
                                                      ; reach this point
%R1<def> = KILL %R1, %R0_R1<imp-def>, %R0_R1<imp-use>

The bug has been here forever, it got exposed recently because the register
scavenger got smarter.

Fixes llvm.org/PR34107

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310979 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert archive-* tests from r310953, there were test failures.
Kuba Mracek [Tue, 15 Aug 2017 23:41:34 +0000 (23:41 +0000)]
Revert archive-* tests from r310953, there were test failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310974 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstCombine] Teach canEvaluateZExtd and canEvaluateTruncated to handle vector shifts...
Craig Topper [Tue, 15 Aug 2017 22:48:41 +0000 (22:48 +0000)]
[InstCombine] Teach canEvaluateZExtd and canEvaluateTruncated to handle vector shifts with splat shift amount

We were only allowing ConstantInt before. This patch allows splat of ConstantInt too.

Differential Revision: https://reviews.llvm.org/D36763

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310970 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoReapply "[GlobalISel] Remove the GISelAccessor API."
Quentin Colombet [Tue, 15 Aug 2017 22:31:51 +0000 (22:31 +0000)]
Reapply "[GlobalISel] Remove the GISelAccessor API."

This reverts commit r310425, thus reapplying r310335 with a fix for link
issue of the AArch64 unittests on Linux bots when BUILD_SHARED_LIBS is ON.

Original commit message:
[GlobalISel] Remove the GISelAccessor API.

Its sole purpose was to avoid spreading around ifdefs related to
building global-isel. Since r309990, GlobalISel is not optional anymore,
thus, we can get rid of this mechanism all together.

NFC.

----
The fix for the link issue consists in adding the GlobalISel library in
the list of dependencies for the AArch64 unittests. This dependency
comes from the use of AArch64Subtarget that needs to know how
to destruct the GISel related APIs when being detroyed.

Thanks to Bill Seurer and Ahmed Bougacha for helping me reproducing and
understand the problem.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310969 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ThinLTO] Fix ThinLTO crash while destroying context
Charles Saternos [Tue, 15 Aug 2017 22:23:44 +0000 (22:23 +0000)]
[ThinLTO] Fix ThinLTO crash while destroying context

Fix for PR32763

An assert that checks if a Ref was untracked fails during ThinLTO context cleanup. The issue is because lazy loading temporary nodes didn't properly track ValueAsMetadata nodes. This patch ensures that the temporary nodes are properly tracked when they're replaced with the value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310967 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert changes in r310953 for llvm-symbolizer.test. The change causes a test failure.
Kuba Mracek [Tue, 15 Aug 2017 21:02:17 +0000 (21:02 +0000)]
Revert changes in r310953 for llvm-symbolizer.test. The change causes a test failure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310956 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoUpdate AMDGPUUsage.rst documentation:
Tony Tye [Tue, 15 Aug 2017 20:47:41 +0000 (20:47 +0000)]
Update AMDGPUUsage.rst documentation:

    1. Correct description of the kernel initial state for FLAT_SCRATCH_INIT.
    2. Add link to GFX9 architecture documentation.
    3. Update product names.
    4. Rename note record from NT_AMD_AMDGPU_METADATA to NT_AMD_AMDGPU_HSA_METADATA and move description to the AMDHSA coding convention section.
    5. Minor typo corrections.

Differential Revision: https://reviews.llvm.org/D36549

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310954 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm] Get rid of "%T" expansions
Kuba Mracek [Tue, 15 Aug 2017 20:29:24 +0000 (20:29 +0000)]
[llvm] Get rid of "%T" expansions

The %T lit expansion expands to a common directory shared between all the tests in the same directory, which is unexpected and unintuitive, and more importantly, it's been a source of subtle race conditions and flaky tests. In https://reviews.llvm.org/D35396, it was agreed that it would be best to simply ban %T and only keep %t, which is unique to each test. When a test needs a temporary directory, it can just create one using mkdir %t.

This patch removes %T in llvm.

Differential Revision: https://reviews.llvm.org/D36495

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310953 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstCombine] Added support for (X >>s C) << C --> X & (-1 << C)
Amjad Aboud [Tue, 15 Aug 2017 19:33:14 +0000 (19:33 +0000)]
[InstCombine] Added support for (X >>s C) << C --> X & (-1 << C)

Differential Revision: https://reviews.llvm.org/D36743

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310949 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ORC][Kaleidoscope] Update Chapter 1 of BuildingAJIT to incorporate recent ORC
Lang Hames [Tue, 15 Aug 2017 19:20:10 +0000 (19:20 +0000)]
[ORC][Kaleidoscope] Update Chapter 1 of BuildingAJIT to incorporate recent ORC
API changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310947 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstCombine] sink sext after ashr
Sanjay Patel [Tue, 15 Aug 2017 18:25:52 +0000 (18:25 +0000)]
[InstCombine] sink sext after ashr

Narrow ops are better for bit-tracking, and in the case of vectors,
may enable better codegen.

As the trunc test shows, this can allow follow-on simplifications.

There's a block of code in visitTrunc that deals with shifted ops
with FIXME comments. It may be possible to remove some of that now,
but I want to make sure there are no problems with this step first.

http://rise4fun.com/Alive/Y3a

Name: hoist_ashr_ahead_of_sext_1
  %s = sext i8 %x to i32
  %r = ashr i32 %s, 3  ; shift value is < than source bit width
  =>
  %a = ashr i8 %x, 3
  %r = sext i8 %a to i32

Name: hoist_ashr_ahead_of_sext_2
  %s = sext i8 %x to i32
  %r = ashr i32 %s, 8  ; shift value is >= than source bit width
  =>
  %a = ashr i8 %x, 7   ; so clamp this shift value
  %r = sext i8 %a to i32

Name: junc_the_trunc
  %a = sext i16 %v to i32
  %s = ashr i32 %a, 18
  %t = trunc i32 %s to i16
  =>
  %t = ashr i16 %v, 15

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310942 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Dominators] Include infinite loops in PostDominatorTree
Jakub Kuderski [Tue, 15 Aug 2017 18:14:57 +0000 (18:14 +0000)]
[Dominators] Include infinite loops in PostDominatorTree

Summary:
This patch teaches PostDominatorTree about infinite loops. It is built on top of D29705 by @dberlin which includes a very detailed motivation for this change.

What's new is that the patch also teaches the incremental updater how to deal with reverse-unreachable regions and how to properly maintain and verify tree roots. Before that, the incremental algorithm sometimes ended up preserving reverse-unreachable regions after updates that wouldn't appear in the tree if it was constructed from scratch on the same CFG.

This patch makes the following assumptions:
- A sequence of updates should produce the same tree as a recalculating it.
- Any sequence of the same updates should lead to the same tree.
- Siblings and roots are unordered.

The last two properties are essential to efficiently perform batch updates in the future.
When it comes to the first one, we can decide later that the consistency between freshly built tree and an updated one doesn't matter match, as there are many correct ways to pick roots in infinite loops, and to relax this assumption. That should enable us to recalculate postdominators less frequently.

This patch is pretty conservative when it comes to incremental updates on reverse-unreachable regions and ends up recalculating the whole tree in many cases. It should be possible to improve the performance in many cases, if we decide that it's important enough.
That being said, my experiments showed that reverse-unreachable are very rare in the IR emitted by clang when bootstrapping  clang. Here are the statistics I collected by analyzing IR between passes and after each removePredecessor call:

```
# functions:  52283
# samples:  337609
# reverse unreachable BBs:  216022
# BBs:  247840796
Percent reverse-unreachable:  0.08716159869015269 %
Max(PercRevUnreachable) in a function:  87.58620689655172 %
# > 25 % samples:  471 ( 0.1395104988314885 % samples )
... in 145 ( 0.27733680163724345 % functions )
```

Most of the reverse-unreachable regions come from invalid IR where it wouldn't be possible to construct a PostDomTree anyway.

I would like to commit this patch in the next week in order to be able to complete the work that depends on it before the end of my internship, so please don't wait long to voice your concerns :).

Reviewers: dberlin, sanjoy, grosser, brzycki, davide, chandlerc, hfinkel

Reviewed By: dberlin

Subscribers: nhaehnle, javed.absar, kparzysz, uabelho, jlebar, hiraditya, llvm-commits, dberlin, david2050

Differential Revision: https://reviews.llvm.org/D35851

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310940 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agotest-release.sh: Move test-suite setup to beginning of the script
Tom Stellard [Tue, 15 Aug 2017 18:11:56 +0000 (18:11 +0000)]
test-release.sh: Move test-suite setup to beginning of the script

Summary:
We want to catch failures early before do the full 3 stage build.

The goal here is to avoid running through the whole build process and have
it fail at the end (and not create the binary packages), just because
some prerequisites failed to install.

Reviewers: rovka, hans

Reviewed By: hans

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D36422

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310939 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ORC] Add case statements for AArch64 to the local stub and callback manager
Lang Hames [Tue, 15 Aug 2017 18:10:19 +0000 (18:10 +0000)]
[ORC] Add case statements for AArch64 to the local stub and callback manager
creation functions.

This should allow lli to lazily execute code using OrcLazyJIT on AArch64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310938 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstCombine] add tests for sext+ashr; NFC
Sanjay Patel [Tue, 15 Aug 2017 17:41:31 +0000 (17:41 +0000)]
[InstCombine] add tests for sext+ashr; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310935 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix -Wunused-lambda-capture for Release build.
Rui Ueyama [Tue, 15 Aug 2017 17:39:35 +0000 (17:39 +0000)]
Fix -Wunused-lambda-capture for Release build.

`I` and `this` are used only in assert or DEBUG, so they are unused
in Release build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310934 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-dwarfdump] - Attemp to fix BB after r310915.
George Rimar [Tue, 15 Aug 2017 16:42:21 +0000 (16:42 +0000)]
[llvm-dwarfdump] - Attemp to fix BB after r310915.

Now MIPS one is unhappy:
http://lab.llvm.org:8011/builders/llvm-mips-linux/builds/2221

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310928 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Doc] Update LangRef for new Module Flag Behavior
Steven Wu [Tue, 15 Aug 2017 16:16:33 +0000 (16:16 +0000)]
[Doc] Update LangRef for new Module Flag Behavior

Summary:
Add the documentation for the new module flag behavior. The new
ModFlagBehavior is added in r303590.

Reviewers: tejohnson

Reviewed By: tejohnson

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D36557

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310926 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-dwarfdump] - Refactor section name/uniqueness gathering.
George Rimar [Tue, 15 Aug 2017 15:54:43 +0000 (15:54 +0000)]
[llvm-dwarfdump] - Refactor section name/uniqueness gathering.

As was requested in D36313 thread,

with this patch section names and uniqueness calculated once,
and not every time when a range is dumped.

Differential revision: https://reviews.llvm.org/D36740

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310923 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert r310919 - [globalisel][tablegen] Support zero-instruction emission.
Daniel Sanders [Tue, 15 Aug 2017 15:10:31 +0000 (15:10 +0000)]
Revert r310919 - [globalisel][tablegen] Support zero-instruction emission.

As expected, this failed on the windows bots but the instrumentation showed
something interesting. The ADD8ri and INC8r rules are never directly compared
on the windows machines. That implies that the issue lies in transitivity of
the Compare predicate. I believe I've already verified that but maybe I missed
something.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310922 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRe-commit with some instrumentation: [globalisel][tablegen] Support zero-instruction...
Daniel Sanders [Tue, 15 Aug 2017 13:50:09 +0000 (13:50 +0000)]
Re-commit with some instrumentation: [globalisel][tablegen] Support zero-instruction emission.

Summary:
Support the case where an operand of a pattern is also the whole of the
result pattern. In this case the original result and all its uses must be
replaced by the operand. However, register class restrictions can require
a COPY. This patch handles both cases by always emitting the copy and
leaving it for the register allocator to optimize.

The previous commit failed on the windows bots and this one is likely to fail
on those same bots. However, the added instrumentation should reveal a particular
isHigherPriorityThan() evaluation which I'm expecting to expose that
these machines are weighing priority of two rules differently from the
non-windows machines.

Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar

Subscribers: javed.absar, kristof.beyls, igorb, llvm-commits

Differential Revision: https://reviews.llvm.org/D36084

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310919 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[DebugInfo] - Attemp to fix BB after r310915.
George Rimar [Tue, 15 Aug 2017 13:26:12 +0000 (13:26 +0000)]
[DebugInfo] - Attemp to fix BB after r310915.

Not sure what BB does not like.

While building module 'LLVM_DebugInfo_DWARF' imported from /home/buildbot/modules-slave-2/clang-x86_64-linux-selfhost-modules-2/llvm.src/lib/DebugInfo/DWARF/DWARFAbbreviationDeclaration.cpp:10:
In file included from <module-includes>:7:
In file included from /home/buildbot/modules-slave-2/clang-x86_64-linux-selfhost-modules-2/llvm.src/include/llvm/DebugInfo/DWARF/DWARFContext.h:29:
/home/buildbot/modules-slave-2/clang-x86_64-linux-selfhost-modules-2/llvm.src/include/llvm/DebugInfo/DWARF/DWARFObject.h:30:17: error: declaration of 'object' must be imported from module 'LLVM_Object.Decompressor' before it is required
  virtual const object::ObjectFile *getFile() const { return nullptr; }
                ^
/home/buildbot/modules-slave-2/clang-x86_64-linux-selfhost-modules-2/llvm.src/include/llvm/Object/Decompressor.h:18:11: note: previous declaration is here
namespace object {

http://lab.llvm.org:8011/builders/clang-x86_64-linux-selfhost-modules-2/builds/10766

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310918 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[RISCV] Add RISCVInstPrinter and basic MC assembler tests
Alex Bradbury [Tue, 15 Aug 2017 13:08:29 +0000 (13:08 +0000)]
[RISCV] Add RISCVInstPrinter and basic MC assembler tests

With the addition of RISCVInstPrinter, it is now possible to test the basic
operation of the RISCV MC layer.

Differential Revision: https://reviews.llvm.org/D23564

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310917 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[llvm-dwarfdump] - Print section name and index when dumping .debug_info ranges
George Rimar [Tue, 15 Aug 2017 12:32:54 +0000 (12:32 +0000)]
[llvm-dwarfdump] - Print section name and index when dumping .debug_info ranges

Teaches llvm-dwarfdump to print section index and name of range
when it dumps .debug_info.

Differential revision: https://reviews.llvm.org/D36313

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310915 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[RISCV] Recognize new relocation types
Alex Bradbury [Tue, 15 Aug 2017 12:11:10 +0000 (12:11 +0000)]
[RISCV] Recognize new relocation types

This patch adds all RISC-V relocation types, as of binutils 2.29. Note that
R_RISCV32_PCREL is not currently documented in the RISC-V ELF PSABI.

Differential Revision: https://reviews.llvm.org/D36455

Patch by Chih-Mao Chen (@PkmX)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310914 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LV] Minor savings to Sink casts to unravel first order recurrence
Ayal Zaks [Tue, 15 Aug 2017 08:32:59 +0000 (08:32 +0000)]
[LV] Minor savings to Sink casts to unravel first order recurrence

Two minor savings: avoid copying the SinkAfter map and avoid moving a cast if it
is not needed.

Differential Revision: https://reviews.llvm.org/D36408

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310910 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoPropagate error in LazyEmittingLayer::removeModule.
Frederich Munch [Tue, 15 Aug 2017 02:25:36 +0000 (02:25 +0000)]
Propagate error in LazyEmittingLayer::removeModule.

Summary:
Besides being the better thing to do, not doing so will triggers an assert with LLVM_ENABLE_ABI_BREAKING_CHECKS.

Reviewers: lhames

Reviewed By: lhames

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D36700

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310906 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SLPVectorizer] Replace VL[0] to VL0 with assert, add propagateIRFlags extra paramete...
Dinar Temirbulatov [Tue, 15 Aug 2017 00:31:49 +0000 (00:31 +0000)]
[SLPVectorizer] Replace VL[0] to VL0 with assert, add propagateIRFlags extra parameter VL0,
                replace E->Scalars[0] to VL0, NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310904 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CMake] Add install target for LLVMFuzzer
Petr Hosek [Mon, 14 Aug 2017 23:37:31 +0000 (23:37 +0000)]
[CMake] Add install target for LLVMFuzzer

This allows including LLVMFuzzer as distribution component.

Differential Revision: https://reviews.llvm.org/D36540

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310897 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd missing dependency in ICP. (NFC)
Dehao Chen [Mon, 14 Aug 2017 23:25:21 +0000 (23:25 +0000)]
Add missing dependency in ICP. (NFC)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310896 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MachineOutliner] Only outline candidates of length >= 2
Jessica Paquette [Mon, 14 Aug 2017 22:57:41 +0000 (22:57 +0000)]
[MachineOutliner] Only outline candidates of length >= 2

Since we don't factor in instruction lengths into outlining calculations
right now, it's never the case that a candidate could have length < 2.

Thus, we should quit early when we see such candidates.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310894 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstSimplify] Teach decomposeBitTestICmp to handle non-canonical compares
Craig Topper [Mon, 14 Aug 2017 22:11:43 +0000 (22:11 +0000)]
[InstSimplify] Teach decomposeBitTestICmp to handle non-canonical compares

This adds support non-canonical compare predicates. InstSimplify can't rely on canonicalization to have occurred.

Differential Revision: https://reviews.llvm.org/D36646

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310893 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRemove checks for debug info intrinsics in use lists, NFC
Reid Kleckner [Mon, 14 Aug 2017 22:10:54 +0000 (22:10 +0000)]
Remove checks for debug info intrinsics in use lists, NFC

These haven't done anything since debug info intrinsics stopped
appearing in Value use lists in 2014.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310892 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MIPS] Implement support for -mstack-alignment.
John Baldwin [Mon, 14 Aug 2017 21:49:38 +0000 (21:49 +0000)]
[MIPS] Implement support for -mstack-alignment.

Summary:
This is modeled on the implementation for x86 which stores the command line
option in a 'StackAlignOverride' field in MipsSubtarget and then uses this
to compute a 'stackAlignment' value in
MipsSubtarget::initializeSubtargetDependencies.

The stackAlignment() method in MipsSubTarget is renamed to getStackAlignment()
and returns the computed 'stackAlignment'.

Reviewers: sdardis

Reviewed By: sdardis

Subscribers: llvm-commits, arichardson

Differential Revision: https://reviews.llvm.org/D35874

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310891 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRecommit r310869, "[InstSimplify][InstCombine] Modify the interface of decomposeBitTe...
Craig Topper [Mon, 14 Aug 2017 21:39:51 +0000 (21:39 +0000)]
Recommit r310869, "[InstSimplify][InstCombine] Modify the interface of decomposeBitTestICmp and use it in the InstSimplify"

This recommits r310869, with the moved files and no extra changes.

Original commit message:

This addresses a fixme in InstSimplify about using decomposeBitTest. This also fixes InstSimplify to handle ugt and ult compares too.

I've modified the interface a little to return only the APInt version of the mask that InstSimplify needs. InstCombine now has a small wrapper routine to create a Constant out of it. I've also dropped the returning of 0 since InstSimplify doesn't need that. So InstCombine creates a zero constant itself.

I also had to make decomposeBitTest support vectors since InstSimplify needs that.

As InstSimplify can't use something from the Transforms library, I've moved the CmpInstAnalysis code to the Analysis library.

Differential Revision: https://reviews.llvm.org/D36593

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310889 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InlineCost] Refactor the checks for different analyses to be a bit more
Chandler Carruth [Mon, 14 Aug 2017 21:25:00 +0000 (21:25 +0000)]
[InlineCost] Refactor the checks for different analyses to be a bit more
localized to the code that uses those analyses.

Technically, this can change behavior as we no longer require the
existence of the ProfileSummaryInfo analysis to use local profile
information via BFI. We didn't actually require the PSI to have an
interesting profile though, so this only really impacts the behavior in
non-default pass pipelines.

IMO, this makes it substantially less surprising how everything works --
before an analysis that wasn't actually used had to exist to trigger
*any* profile aware inlining. I think the new organization makes it more
obvious where various checks for profile signals happen.

Differential Revision: https://reviews.llvm.org/D36710

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310888 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd strictfp attribute to prevent unwanted optimizations of libm calls
Andrew Kaylor [Mon, 14 Aug 2017 21:15:13 +0000 (21:15 +0000)]
Add strictfp attribute to prevent unwanted optimizations of libm calls

Differential Revision: https://reviews.llvm.org/D34163

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310885 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[libFuzzer] try to use less RAM while processing the initial corpus
Kostya Serebryany [Mon, 14 Aug 2017 20:34:35 +0000 (20:34 +0000)]
[libFuzzer] try to use less RAM while processing the initial corpus

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310881 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[libFuzzer] explicitly use -fsanitize-coverage=trace-pc-guard in test/dump_coverage...
Kostya Serebryany [Mon, 14 Aug 2017 19:55:23 +0000 (19:55 +0000)]
[libFuzzer] explicitly use -fsanitize-coverage=trace-pc-guard in test/dump_coverage.test; mark print_coverage/dump_coverage as To-be-deprecated

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310877 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoIPRA: Allow target to enable IPRA by default
Matt Arsenault [Mon, 14 Aug 2017 19:54:47 +0000 (19:54 +0000)]
IPRA: Allow target to enable IPRA by default

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@310876 91177308-0d34-0410-b5e6-96231b3b80d8