platform/upstream/llvm.git
6 years ago[MemorySSAUpdater] Remove deleted trivial Phis from active workset
Alexandros Lamprineas [Mon, 16 Jul 2018 07:51:27 +0000 (07:51 +0000)]
[MemorySSAUpdater] Remove deleted trivial Phis from active workset

Bug fix for PR37808. The regression test is a reduced version of the
original reproducer attached to the bug report. As stated in the report,
the problem was that InsertedPHIs was keeping dangling pointers to
deleted Memory-Phis. MemoryPhis are created eagerly and sometimes get
zapped shortly afterwards. I've used WeakVH instead of an expensive
removal operation from the active workset.

Differential Revision: https://reviews.llvm.org/D48372

llvm-svn: 337149

6 years ago[Sema] Add fixit for unused lambda captures
Alexander Shaposhnikov [Mon, 16 Jul 2018 07:23:47 +0000 (07:23 +0000)]
[Sema] Add fixit for unused lambda captures

This diff adds a fixit to suggest removing unused lambda captures
in the appropriate diagnostic.

Patch by Andrew Comminos!

Test plan: make check-all

Differential revision: https://reviews.llvm.org/D48845

llvm-svn: 337148

6 years ago[X86] Merge the FR128 and VR128 regclass since they have identical spill and alignmen...
Craig Topper [Mon, 16 Jul 2018 06:56:09 +0000 (06:56 +0000)]
[X86] Merge the FR128 and VR128 regclass since they have identical spill and alignment characteristics.

This unfortunately requires a bunch of bitcasts to be added added to SUBREG_TO_REG, COPY_TO_REGCLASS, and instructions in output patterns. Otherwise tablegen seems to default to picking f128 and then we fail when something tries to get the register class for f128 which isn't always valid.

The test changes are because we were previously mixing fr128 and vr128 due to contrainRegClass finding FR128 first and passes like live range shrinking weren't handling that well.

llvm-svn: 337147

6 years ago[MinGW] Automatically mangle Windows-specific entry points as C
Martin Storsjo [Mon, 16 Jul 2018 05:42:25 +0000 (05:42 +0000)]
[MinGW] Automatically mangle Windows-specific entry points as C

This mangles entry points wmain, WinMain, wWinMain or DllMain as C
functions, to match the ABI for these functions.

We already did the same for these functions in MSVC mode, but we also
should do the same in the Itanium ABI.

This fixes PR38124.

Differential Revision: https://reviews.llvm.org/D49354

llvm-svn: 337146

6 years ago[x86/SLH] Fix an unused variable warning in release builds after
Chandler Carruth [Mon, 16 Jul 2018 04:42:27 +0000 (04:42 +0000)]
[x86/SLH] Fix an unused variable warning in release builds after
r337144.

llvm-svn: 337145

6 years ago[x86/SLH] Teach speculative load hardening to correctly harden the
Chandler Carruth [Mon, 16 Jul 2018 04:17:51 +0000 (04:17 +0000)]
[x86/SLH] Teach speculative load hardening to correctly harden the
indices used by AVX2 and AVX-512 gather instructions.

The index vector is hardened by broadcasting the predicate state
into a vector register and then or-ing. We don't even have to worry
about EFLAGS here.

I've added a test for all of the gather intrinsics to make sure that we
don't miss one. A particularly interesting creation is the gather
prefetch, which needs to be marked as potentially "loading" to get the
correct behavior. It's a memory access in many ways, and is actually
relevant for SLH. Based on discussion with Craig in review, I've moved
it to be `mayLoad` and `mayStore` rather than generic side effects. This
matches how we model other prefetch instructions.

Many thanks to Craig for the review here.

Differential Revision: https://reviews.llvm.org/D49336

llvm-svn: 337144

6 years ago[InstCombine] add more SPFofSPF folding
Chen Zheng [Mon, 16 Jul 2018 02:23:00 +0000 (02:23 +0000)]
[InstCombine] add more SPFofSPF folding

Differential Revision: https://reviews.llvm.org/D49238

llvm-svn: 337143

6 years ago[XRay][compiler-rt] Use `SANITIZER_CXX_ABI_LIBRARY` for XRay unit tests
Dean Michael Berris [Mon, 16 Jul 2018 01:08:56 +0000 (01:08 +0000)]
[XRay][compiler-rt] Use `SANITIZER_CXX_ABI_LIBRARY` for XRay unit tests

Summary:
Fix a TODO in CMake config for XRay tests to use the detected C++ ABI
library in the tests.

Also make the tests depend on the llvm-xray target when built in-tree.

Reviewers: kpw, eizan

Reviewed By: eizan

Subscribers: mgorny, llvm-commits

Differential Revision: https://reviews.llvm.org/D49358

llvm-svn: 337142

6 years ago[InstCombine] fold icmp pred (sub 0, X) C for vector type
Chen Zheng [Mon, 16 Jul 2018 00:51:40 +0000 (00:51 +0000)]
[InstCombine] fold icmp pred (sub 0, X) C for vector type

Differential Revision: https://reviews.llvm.org/D49283

llvm-svn: 337141

6 years agoRecommit r335794 "Add support for generating a call graph profile from Branch Frequen...
Michael J. Spencer [Mon, 16 Jul 2018 00:28:24 +0000 (00:28 +0000)]
Recommit r335794 "Add support for generating a call graph profile from Branch Frequency Info." with fix for removed functions.

llvm-svn: 337140

6 years agoPrune empty directory.
Joerg Sonnenberger [Sun, 15 Jul 2018 23:52:15 +0000 (23:52 +0000)]
Prune empty directory.

llvm-svn: 337139

6 years ago[x86/SLH] Extract one of the bits of logic to its own function. NFC.
Chandler Carruth [Sun, 15 Jul 2018 23:46:36 +0000 (23:46 +0000)]
[x86/SLH] Extract one of the bits of logic to its own function. NFC.

This is just a refactoring to start cleaning up the code here and make
it more readable and approachable.

llvm-svn: 337138

6 years ago[X86] Add custom execution domain fixing for 128/256-bit integer logic operations...
Craig Topper [Sun, 15 Jul 2018 23:32:36 +0000 (23:32 +0000)]
[X86] Add custom execution domain fixing for 128/256-bit integer logic operations with AVX512F, but not AVX512DQ.

AVX512F only has integer domain logic instructions. AVX512DQ added FP domain logic instructions.

Execution domain fixing runs before EVEX->VEX. So if we have AVX512F and not AVX512DQ we fail to do execution domain switching of the logic operations. This leads to mismatches in execution domain and more test differences.

This patch adds custom domain fixing that switches EVEX integer logic operations to VEX fp logic operations if XMM16-31 are not used.

llvm-svn: 337137

6 years ago[X86] Add load patterns for cases where we select X86Movss/X86Movsd to blend instruct...
Craig Topper [Sun, 15 Jul 2018 21:49:01 +0000 (21:49 +0000)]
[X86] Add load patterns for cases where we select X86Movss/X86Movsd to blend instructions.

This allows us to fold the load during isel without waiting for the peephole pass to do it.

llvm-svn: 337136

6 years ago[X86] Use 128-bit blends instead vmovss/vmovsd for 512-bit vzmovl patterns to match...
Craig Topper [Sun, 15 Jul 2018 18:51:08 +0000 (18:51 +0000)]
[X86] Use 128-bit blends instead vmovss/vmovsd for 512-bit vzmovl patterns to match AVX.

llvm-svn: 337135

6 years ago[X86] Use 128-bit ops for 256-bit vzmovl patterns.
Craig Topper [Sun, 15 Jul 2018 18:51:07 +0000 (18:51 +0000)]
[X86] Use 128-bit ops for 256-bit vzmovl patterns.

128-bit ops implicitly zero the upper bits. This should address the comment about domain crossing for the integer version without AVX2 since we can use a 128-bit VBLENDW without AVX2.

The only bad thing I see here is that we failed to reuse an vxorps in some of the tests, but I think that's already known issue.

llvm-svn: 337134

6 years ago[cmake] Fix libomptarget/test/CMakeLists.txt
Azharuddin Mohammed [Sun, 15 Jul 2018 17:29:43 +0000 (17:29 +0000)]
[cmake] Fix libomptarget/test/CMakeLists.txt

Summary:
Should be variable name instead of variable reference. If the variable is
somehow unset, it messes up the if condition expression and causes a CMake
error.

Reviewers: jlpeyton, AndreyChurbanov, Hahnfeld

Reviewed By: Hahnfeld

Subscribers: mgorny, llvm-commits, openmp-commits

Differential Revision: https://reviews.llvm.org/D47221

llvm-svn: 337133

6 years ago[DAGCombiner] fix typo in comment; NFC
Sanjay Patel [Sun, 15 Jul 2018 17:09:35 +0000 (17:09 +0000)]
[DAGCombiner] fix typo in comment; NFC

llvm-svn: 337132

6 years ago[InstCombine] Corrections in comments for division transformation (NFC)
Sanjay Patel [Sun, 15 Jul 2018 17:06:59 +0000 (17:06 +0000)]
[InstCombine] Corrections in comments for division transformation (NFC)

The actual code seems to be correct, but the comments were misleading.

Patch by Aaron Puchert!

Differential Revision: https://reviews.llvm.org/D49276

llvm-svn: 337131

6 years ago[DAGCombiner] extend(ifpositive(X)) -> shift-right (not X)
Sanjay Patel [Sun, 15 Jul 2018 16:27:07 +0000 (16:27 +0000)]
[DAGCombiner] extend(ifpositive(X)) -> shift-right (not X)

This is almost the same as an existing IR canonicalization in instcombine,
so I'm assuming this is a good early generic DAG combine too.

The motivation comes from reduced bit-hacking for select-of-constants in IR
after rL331486. We want to restore that functionality in the DAG as noted in
the commit comments for that change and the llvm-dev discussion here:
http://lists.llvm.org/pipermail/llvm-dev/2018-July/124433.html

The PPC and AArch tests show that those targets are already doing something
similar. x86 will be neutral in the minimal case and generally better when
this pattern is extended with other ops as shown in the signbit-shift.ll tests.

Note the asymmetry: we don't include the (extend (ifneg X)) transform because
it already exists in SimplifySelectCC(), and that is verified in the later
unchanged tests in the signbit-shift.ll files. Without the 'not' op, the
general transform to use a shift is always a win because that's a single
instruction.

Alive proofs:
https://rise4fun.com/Alive/ysli

Name: if pos, get -1
  %c = icmp sgt i16 %x, -1
  %r = sext i1 %c to i16
  =>
  %n = xor i16 %x, -1
  %r = ashr i16 %n, 15

Name: if pos, get 1
  %c = icmp sgt i16 %x, -1
  %r = zext i1 %c to i16
  =>
  %n = xor i16 %x, -1
  %r = lshr i16 %n, 15

Differential Revision: https://reviews.llvm.org/D48970

llvm-svn: 337130

6 years ago[InstSimplify] add fixme comment for PR37776; NFC
Sanjay Patel [Sun, 15 Jul 2018 16:13:58 +0000 (16:13 +0000)]
[InstSimplify] add fixme comment for PR37776; NFC

llvm-svn: 337129

6 years ago[AMDGPU] adjusted test checks because minnum with NaN gets simplified
Sanjay Patel [Sun, 15 Jul 2018 15:14:40 +0000 (15:14 +0000)]
[AMDGPU] adjusted test checks because minnum with NaN gets simplified

This was improved with rL337127, but I missed the failure in this test.
I'm not sure what the expected result will be, so I've generalized it
and added a FIXME comment.

llvm-svn: 337128

6 years ago[InstSimplify] fold minnum/maxnum with NaN arg
Sanjay Patel [Sun, 15 Jul 2018 14:52:16 +0000 (14:52 +0000)]
[InstSimplify] fold minnum/maxnum with NaN arg

This fold is repeated/misplaced in instcombine, but I'm
not sure if it's safe to remove that yet because some
other folds appear to be asserting that the transform
has occurred within instcombine itself.

This isn't the best fix for PR37776, but it probably
hides the bug with the given code example:
https://bugs.llvm.org/show_bug.cgi?id=37776

We have another test to demonstrate the more general bug.

llvm-svn: 337127

6 years ago[InstSimplify] add tests for minnum/maxnum; NFC
Sanjay Patel [Sun, 15 Jul 2018 14:46:48 +0000 (14:46 +0000)]
[InstSimplify] add tests for minnum/maxnum; NFC

This isn't the best fix for PR37776, but it probably
hides the bug with the given code example:
https://bugs.llvm.org/show_bug.cgi?id=37776

We have another test to demonstrate the more general
bug.

llvm-svn: 337126

6 years agoRun thread safety tests with both lock and capability attributes; NFC to the analysis...
Aaron Ballman [Sun, 15 Jul 2018 12:08:52 +0000 (12:08 +0000)]
Run thread safety tests with both lock and capability attributes; NFC to the analysis behavior.

Patch thanks to Aaron Puchert.

llvm-svn: 337125

6 years ago[llvm-mca] Regenerate X86 specific tests. NFC
Andrea Di Biagio [Sun, 15 Jul 2018 11:43:11 +0000 (11:43 +0000)]
[llvm-mca] Regenerate X86 specific tests. NFC

Not all tests were correctly updated by the update script after r336797.

llvm-svn: 337124

6 years ago[llvm-mca][BtVer2] teach how to identify false dependencies on partially written
Andrea Di Biagio [Sun, 15 Jul 2018 11:01:38 +0000 (11:01 +0000)]
[llvm-mca][BtVer2] teach how to identify false dependencies on partially written
registers.

The goal of this patch is to improve the throughput analysis in llvm-mca for the
case where instructions perform partial register writes.

On x86, partial register writes are quite difficult to model, mainly because
different processors tend to implement different register merging schemes in
hardware.

When the code contains partial register writes, the IPC (instructions per
cycles) estimated by llvm-mca tends to diverge quite significantly from the
observed IPC (using perf).

Modern AMD processors (at least, from Bulldozer onwards) don't rename partial
registers. Quoting Agner Fog's microarchitecture.pdf:
" The processor always keeps the different parts of an integer register together.
For example, AL and AH are not treated as independent by the out-of-order
execution mechanism. An instruction that writes to part of a register will
therefore have a false dependence on any previous write to the same register or
any part of it."

This patch is a first important step towards improving the analysis of partial
register updates. It changes the semantic of RegisterFile descriptors in
tablegen, and teaches llvm-mca how to identify false dependences in the presence
of partial register writes (for more details: see the new code comments in
include/Target/TargetSchedule.h - class RegisterFile).

This patch doesn't address the case where a write to a part of a register is
followed by a read from the whole register.  On Intel chips, high8 registers
(AH/BH/CH/DH)) can be stored in separate physical registers. However, a later
(dirty) read of the full register (example: AX/EAX) triggers a merge uOp, which
adds extra latency (and potentially affects the pipe usage).
This is a very interesting article on the subject with a very informative answer
from Peter Cordes:
https://stackoverflow.com/questions/45660139/how-exactly-do-partial-registers-on-haswell-skylake-perform-writing-al-seems-to

In future, the definition of RegisterFile can be extended with extra information
that may be used to identify delays caused by merge opcodes triggered by a dirty
read of a partial write.

Differential Revision: https://reviews.llvm.org/D49196

llvm-svn: 337123

6 years ago[AVR] Document some public functions
Dylan McKay [Sun, 15 Jul 2018 07:24:27 +0000 (07:24 +0000)]
[AVR] Document some public functions

llvm-svn: 337122

6 years ago[TableGen] std::move vectors into TreePatternNode.
Craig Topper [Sun, 15 Jul 2018 06:52:49 +0000 (06:52 +0000)]
[TableGen] std::move vectors into TreePatternNode.

llvm-svn: 337121

6 years ago[TableGen] Remove what seems to be an unnecessary std::map copy.
Craig Topper [Sun, 15 Jul 2018 06:52:48 +0000 (06:52 +0000)]
[TableGen] Remove what seems to be an unnecessary std::map copy.

The comment says the copy was made so it could be destroyed in the following loop, but the original map wasn't used after the loop.

llvm-svn: 337120

6 years ago[X86] Add some optsize patterns for 256-bit X86vzmovl.
Craig Topper [Sun, 15 Jul 2018 06:03:19 +0000 (06:03 +0000)]
[X86] Add some optsize patterns for 256-bit X86vzmovl.

These patterns use VMOVSS/SD. Without optsize we use BLENDI instead.

llvm-svn: 337119

6 years ago[CMake] Use correct variable as header install prefix
Petr Hosek [Sun, 15 Jul 2018 04:09:35 +0000 (04:09 +0000)]
[CMake] Use correct variable as header install prefix

This variable is already set in CMakeLists.txt but it wasn't used
which means that the headers get installed into a wrong location
when the per target runtime directory option is being used.

Differential Revision: https://reviews.llvm.org/D49345

llvm-svn: 337118

6 years ago[CMake] Use libc++ and compiler-rt for sanitizers
Petr Hosek [Sun, 15 Jul 2018 03:11:43 +0000 (03:11 +0000)]
[CMake] Use libc++ and compiler-rt for sanitizers

When building runtimes for Linux as part of Fuchsia toolchain, use
libc++ and compiler-rt for sanitizers.

Differential Revision: https://reviews.llvm.org/D49331

llvm-svn: 337117

6 years ago[CMake] Change the flag to use compiler-rt builtins to boolean
Petr Hosek [Sun, 15 Jul 2018 03:05:20 +0000 (03:05 +0000)]
[CMake] Change the flag to use compiler-rt builtins to boolean

This changes the name and the type to what it was prior to r333037
which matches the name of the flag used in other runtimes: libc++,
libc++abi and libunwind. We don't need the type to be a string since
there's only binary choice between libgcc and compiler-rt unlike in
the case of C++ library where there're multiple options.

Differential Revision: https://reviews.llvm.org/D49325

llvm-svn: 337116

6 years ago[CMake] Pass CMAKE_INSTALL_DO_STRIP to external projects
Petr Hosek [Sun, 15 Jul 2018 02:12:25 +0000 (02:12 +0000)]
[CMake] Pass CMAKE_INSTALL_DO_STRIP to external projects

This is necessary to make install-<target>-stripped work for
external projects such as runtimes.

Differential Revision: https://reviews.llvm.org/D49335

llvm-svn: 337115

6 years ago[TableGen] Add some std::move to the PatternToMatch constructor.
Craig Topper [Sun, 15 Jul 2018 01:10:28 +0000 (01:10 +0000)]
[TableGen] Add some std::move to the PatternToMatch constructor.

The are two vectors passed by value to the constructor. We should be able to move them into the object.

llvm-svn: 337114

6 years ago[llvm-mca] Turn InstructionTables into a Stage.
Matt Davis [Sat, 14 Jul 2018 23:52:50 +0000 (23:52 +0000)]
[llvm-mca] Turn InstructionTables into a Stage.

Summary:
This patch converts the InstructionTables class into a subclass of mca::Stage.  This change allows us to use the Stage's inherited Listeners for event notifications.  This also allows us to create a simple pipeline for viewing the InstructionTables report.

I have been working on a follow on patch that should cleanup addView in InstructionTables.  Right now, addView adds the view to both the Listener list and Views list.  The follow-on patch addresses the fact that we don't really need two lists in this case.  That change is not specific to just InstructionTables, so it will be a separate patch.

Reviewers: andreadb, courbet, RKSimon

Reviewed By: andreadb

Subscribers: tschuett, gbedwell, llvm-commits

Differential Revision: https://reviews.llvm.org/D49329

llvm-svn: 337113

6 years ago[NFC][InstCombine] foldICmpWithLowBitMaskedVal(): update comments.
Roman Lebedev [Sat, 14 Jul 2018 20:08:52 +0000 (20:08 +0000)]
[NFC][InstCombine] foldICmpWithLowBitMaskedVal(): update comments.

All predicates are handled.
There does not seem to be any other possible folds here.
There are some more folds possible with inverted mask though.

llvm-svn: 337112

6 years ago[InstCombine] Fold x & (-1 >> y) s< x to x s> (-1 >> y)
Roman Lebedev [Sat, 14 Jul 2018 20:08:47 +0000 (20:08 +0000)]
[InstCombine] Fold  x & (-1 >> y) s< x  to  x s> (-1 >> y)

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/I3O

This pattern is not commutative!
We must make sure not to fold the commuted version!

llvm-svn: 337111

6 years ago[NFC][InstCombine] Tests for x & (-1 >> y) s< x to x s> (-1 >> y) fold.
Roman Lebedev [Sat, 14 Jul 2018 20:08:42 +0000 (20:08 +0000)]
[NFC][InstCombine] Tests for  x & (-1 >> y) s< x  to  x s> (-1 >> y)  fold.

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/I3O

This pattern is not commutative!
We must make sure not to fold the commuted version!

llvm-svn: 337110

6 years ago[InstCombine] Fold x & (-1 >> y) s>= x to x s<= (-1 >> y)
Roman Lebedev [Sat, 14 Jul 2018 20:08:37 +0000 (20:08 +0000)]
[InstCombine] Fold  x & (-1 >> y) s>= x  to  x s<= (-1 >> y)

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/I3O

This pattern is not commutative!
We must make sure not to fold the commuted version!

llvm-svn: 337109

6 years ago[NFC][InstCombine] Tests for x & (-1 >> y) s>= x to x s<= (-1 >> y) fold.
Roman Lebedev [Sat, 14 Jul 2018 20:08:31 +0000 (20:08 +0000)]
[NFC][InstCombine] Tests for  x & (-1 >> y) s>= x  to  x s<= (-1 >> y)  fold.

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/I3O

This pattern is not commutative!
We must make sure not to fold the commuted version!

llvm-svn: 337108

6 years ago[InstCombine] Fold x s<= x & (-1 >> y) to x s<= (-1 >> y)
Roman Lebedev [Sat, 14 Jul 2018 20:08:26 +0000 (20:08 +0000)]
[InstCombine] Fold  x s<= x & (-1 >> y)  to  x s<= (-1 >> y)

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/I3O

This pattern is not commutative!
We must make sure not to fold the commuted version!

llvm-svn: 337107

6 years ago[NFC][InstCombine] Tests for x s<= x & (-1 >> y) to x s<= (-1 >> y) fold.
Roman Lebedev [Sat, 14 Jul 2018 20:08:21 +0000 (20:08 +0000)]
[NFC][InstCombine] Tests for  x s<= x & (-1 >> y)  to  x s<= (-1 >> y)  fold.

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/I3O

This pattern is not commutative!
We must make sure not to fold the commuted version!

llvm-svn: 337106

6 years ago[InstCombine] Fold x s> x & (-1 >> y) to x s> (-1 >> y)
Roman Lebedev [Sat, 14 Jul 2018 20:08:16 +0000 (20:08 +0000)]
[InstCombine] Fold  x s> x & (-1 >> y)  to  x s> (-1 >> y)

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/I3O

This pattern is not commutative!
We must make sure not to fold the commuted version!

llvm-svn: 337105

6 years ago[NFC][InstCombine] Tests for x s> x & (-1 >> y) to x s> (-1 >> y) fold.
Roman Lebedev [Sat, 14 Jul 2018 20:08:09 +0000 (20:08 +0000)]
[NFC][InstCombine] Tests for  x s> x & (-1 >> y)  to  x s> (-1 >> y)  fold.

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/I3O

This pattern is not commutative!
We must make sure not to fold the commuted version!

llvm-svn: 337104

6 years agoAdd caching when looking up coroutine_traits
Brian Gesiak [Sat, 14 Jul 2018 18:21:44 +0000 (18:21 +0000)]
Add caching when looking up coroutine_traits

Summary:
Currently clang looks up the coroutine_traits ClassTemplateDecl
everytime it looks up the promise type. This is unnecessary
as coroutine_traits doesn't change between promise type lookups.

This diff caches the coroutine_traits lookup.

Patch by Tanoy Sinha!

Test Plan:
I added log statements in the new lookupCoroutineTraits function
to ensure that LookupQualifiedName was only called once even
when multiple coroutines existed in the source file.

Reviewers: modocache, GorNishanov

Reviewed By: modocache

Subscribers: cfe-commits

Differential Revision: https://reviews.llvm.org/D48981

llvm-svn: 337103

6 years ago[InstCombine] Fold x u<= x & C to x u<= C
Roman Lebedev [Sat, 14 Jul 2018 16:44:54 +0000 (16:44 +0000)]
[InstCombine] Fold  x u<= x & C  to  x u<= C

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/Fqp

This pattern is not commutative. But InstSimplify will
already have taken care of the 'commutative' variant.

llvm-svn: 337102

6 years ago[NFC][InstCombine] Tests for x u<= x & C to x u<= C fold.
Roman Lebedev [Sat, 14 Jul 2018 16:44:48 +0000 (16:44 +0000)]
[NFC][InstCombine] Tests for  x u<= x & C  to  x u<= C  fold.

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/Fqp

This pattern is not commutative. But InstSimplify will
already have taken care of the 'commutative' variant.

llvm-svn: 337101

6 years ago[InstCombine] Fold x u> x & C to x u> C
Roman Lebedev [Sat, 14 Jul 2018 16:44:43 +0000 (16:44 +0000)]
[InstCombine] Fold  x u> x & C  to  x u> C

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/JvS

This pattern is not commutative. But InstSimplify will
already have taken care of the 'commutative' variant.

llvm-svn: 337100

6 years ago[NFC][InstCombine] Tests for x u> x & C to x u> C fold.
Roman Lebedev [Sat, 14 Jul 2018 16:44:37 +0000 (16:44 +0000)]
[NFC][InstCombine] Tests for  x u> x & C  to  x u> C  fold.

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/JvS

This pattern is not commutative. But InstSimplify will
already have taken care of the 'commutative' variant.

llvm-svn: 337099

6 years ago[InstCombine] Fold x & (-1 >> y) u< x to x u> (-1 >> y)
Roman Lebedev [Sat, 14 Jul 2018 12:20:16 +0000 (12:20 +0000)]
[InstCombine] Fold  x & (-1 >> y) u< x  to  x u> (-1 >> y)

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/ocb

This pattern is not commutative. But InstSimplify will
already have taken care of the 'commutative' variant.

llvm-svn: 337098

6 years ago[NFC][InstCombine] Tests for x & (-1 >> y) u< x to x u> (-1 >> y)
Roman Lebedev [Sat, 14 Jul 2018 12:20:11 +0000 (12:20 +0000)]
[NFC][InstCombine] Tests for  x & (-1 >> y) u< x  to  x u> (-1 >> y)

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/ocb

This pattern is not commutative. But InstSimplify will
already have taken care of the 'commutative' variant.

llvm-svn: 337097

6 years ago[InstCombine] Fold x & (-1 >> y) u>= x to x u<= (-1 >> y)
Roman Lebedev [Sat, 14 Jul 2018 12:20:06 +0000 (12:20 +0000)]
[InstCombine] Fold  x & (-1 >> y) u>= x  to  x u<= (-1 >> y)

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/azI

This pattern is not commutative. But InstSimplify will
already have taken care of the 'commutative' variant.

llvm-svn: 337096

6 years ago[NFC][InstCombine] Tests for x & (-1 >> y) u>= x to x u<= (-1 >> y)
Roman Lebedev [Sat, 14 Jul 2018 12:20:01 +0000 (12:20 +0000)]
[NFC][InstCombine] Tests for  x & (-1 >> y) u>= x  to  x u<= (-1 >> y)

https://bugs.llvm.org/show_bug.cgi?id=38123
https://rise4fun.com/Alive/azI

This pattern is not commutative. But InstSimplify will
already have taken care of the 'commutative' variant.

llvm-svn: 337095

6 years ago[NFC][InstCombine] Add forgotten variable tests for foldICmpWithLowBitMaskedVal()
Roman Lebedev [Sat, 14 Jul 2018 12:19:56 +0000 (12:19 +0000)]
[NFC][InstCombine] Add forgotten variable tests for foldICmpWithLowBitMaskedVal()

llvm-svn: 337094

6 years agoattempt to get test/COFF/driver.test passing on sanitizer-x86_64-linux-fast; cf r337092
Nico Weber [Sat, 14 Jul 2018 11:47:23 +0000 (11:47 +0000)]
attempt to get test/COFF/driver.test passing on sanitizer-x86_64-linux-fast; cf r337092

llvm-svn: 337093

6 years agoAttempt to get test/tools/llvm-lib/help.test passing on sanitizer-x86_64-linux-fast
Nico Weber [Sat, 14 Jul 2018 11:33:33 +0000 (11:33 +0000)]
Attempt to get test/tools/llvm-lib/help.test passing on sanitizer-x86_64-linux-fast

The bot has a /b directory, so /? matches against that and gets expanded to it.

(Thanks to Hans's r187366, which solved the same problem for clang-cl a while
ago and which saved me much head scratching.)

llvm-svn: 337092

6 years ago[clang-tidy] Force exceptions to be enabled in test
Benjamin Kramer [Sat, 14 Jul 2018 10:48:06 +0000 (10:48 +0000)]
[clang-tidy] Force exceptions to be enabled in test

For targets that have them off by default.

llvm-svn: 337091

6 years ago[MachineOutliner] Check the last instruction from the sequence when updating liveness
Francis Visoiu Mistrih [Sat, 14 Jul 2018 09:40:01 +0000 (09:40 +0000)]
[MachineOutliner] Check the last instruction from the sequence when updating liveness

The MachineOutliner was doing an std::for_each from the call (inserted
before the outlined sequence) to the iterator at the end of the
sequence.

std::for_each needs the iterator past the end, so the last instruction
was not taken into account when propagating the liveness information.

This fixes the machine verifier issue in machine-outliner-disubprogram.ll.

Differential Revision: https://reviews.llvm.org/D49295

llvm-svn: 337090

6 years ago[x86/SLH] Fix an issue where we wouldn't harden any loads if we found
Chandler Carruth [Sat, 14 Jul 2018 09:32:37 +0000 (09:32 +0000)]
[x86/SLH] Fix an issue where we wouldn't harden any loads if we found
no conditions.

This is only valid to do if we're hardening calls and rets with LFENCE
which results in an LFENCE guarding the entire entry block for us.

llvm-svn: 337089

6 years ago[X86] Fix a subtle bug in the custom execution domain fixing for blends.
Craig Topper [Sat, 14 Jul 2018 06:30:30 +0000 (06:30 +0000)]
[X86] Fix a subtle bug in the custom execution domain fixing for blends.

The code tried to find the immediate by using getNumOperands() on the MachineInstr, but there might be implicit-defs after the immediate that get counted.

Instead use getNumOperands() from the instruction description which will only count the operands that are defined in the td file.

llvm-svn: 337088

6 years agoMark __equal_to 's operations as constexpr.
Marshall Clow [Sat, 14 Jul 2018 04:15:19 +0000 (04:15 +0000)]
Mark __equal_to 's operations as constexpr.

llvm-svn: 337087

6 years agolld-link: Add /lib to Options.td so that it appears in lld-link's help output.
Nico Weber [Sat, 14 Jul 2018 04:07:51 +0000 (04:07 +0000)]
lld-link: Add /lib to Options.td so that it appears in lld-link's help output.

https://reviews.llvm.org/D49319

llvm-svn: 337086

6 years agoMark one more __wrap_iter operation as constexpr.
Marshall Clow [Sat, 14 Jul 2018 03:06:11 +0000 (03:06 +0000)]
Mark one more __wrap_iter operation as constexpr.

llvm-svn: 337085

6 years agoGive llvm-lib rudimentary help output.
Nico Weber [Sat, 14 Jul 2018 02:29:44 +0000 (02:29 +0000)]
Give llvm-lib rudimentary help output.

https://reviews.llvm.org/D49318

llvm-svn: 337084

6 years ago[X86] Prefer blendi over movss/sd when avx512 is enabled unless optimizing for size.
Craig Topper [Sat, 14 Jul 2018 02:05:08 +0000 (02:05 +0000)]
[X86] Prefer blendi over movss/sd when avx512 is enabled unless optimizing for size.

AVX512 doesn't have an immediate controlled blend instruction. But blend throughput is still better than movss/sd on SKX.

This commit changes AVX512 to use the AVX blend instructions instead of MOVSS/MOVSD. This constrains the register allocation since it won't be able to use XMM16-31, but hopefully the increased throughput and reduced port 5 pressure makes up for that.

llvm-svn: 337083

6 years agoRevert "[ThinLTO] Ensure we always select the same function copy to import"
Teresa Johnson [Sat, 14 Jul 2018 01:50:14 +0000 (01:50 +0000)]
Revert "[ThinLTO] Ensure we always select the same function copy to import"

This reverts commit r337051.

llvm-svn: 337082

6 years agoRevert "[ThinLTO] Ensure we always select the same function copy to import"
Teresa Johnson [Sat, 14 Jul 2018 01:45:49 +0000 (01:45 +0000)]
Revert "[ThinLTO] Ensure we always select the same function copy to import"

This reverts commits r337050 and r337059. Caused failure in
reverse-iteration bot that needs more investigation.

llvm-svn: 337081

6 years agoRevert "[ThinLTO] Add debug output to test"
Teresa Johnson [Sat, 14 Jul 2018 01:34:06 +0000 (01:34 +0000)]
Revert "[ThinLTO] Add debug output to test"

This reverts commit r337076. Not needed any more.

llvm-svn: 337080

6 years agoRevert "AMDGPU: Fix handling of alignment padding in DAG argument lowering"
Evgeniy Stepanov [Sat, 14 Jul 2018 01:20:53 +0000 (01:20 +0000)]
Revert "AMDGPU: Fix handling of alignment padding in DAG argument lowering"

This reverts commit r337021.

WARNING: MemorySanitizer: use-of-uninitialized-value
    #0 0x1415cd65 in void write_signed<long>(llvm::raw_ostream&, long, unsigned long, llvm::IntegerStyle) /code/llvm-project/llvm/lib/Support/NativeFormatting.cpp:95:7
    #1 0x1415c900 in llvm::write_integer(llvm::raw_ostream&, long, unsigned long, llvm::IntegerStyle) /code/llvm-project/llvm/lib/Support/NativeFormatting.cpp:121:3
    #2 0x1472357f in llvm::raw_ostream::operator<<(long) /code/llvm-project/llvm/lib/Support/raw_ostream.cpp:117:3
    #3 0x13bb9d4 in llvm::raw_ostream::operator<<(int) /code/llvm-project/llvm/include/llvm/Support/raw_ostream.h:210:18
    #4 0x3c2bc18 in void printField<unsigned int, &(amd_kernel_code_s::amd_kernel_code_version_major)>(llvm::StringRef, amd_kernel_code_s const&, llvm::raw_ostream&) /code/llvm-project/llvm/lib/Target/AMDGPU/Utils/AMDKernelCodeTUtils.cpp:78:23
    #5 0x3c250ba in llvm::printAmdKernelCodeField(amd_kernel_code_s const&, int, llvm::raw_ostream&) /code/llvm-project/llvm/lib/Target/AMDGPU/Utils/AMDKernelCodeTUtils.cpp:104:5
    #6 0x3c27ca3 in llvm::dumpAmdKernelCode(amd_kernel_code_s const*, llvm::raw_ostream&, char const*) /code/llvm-project/llvm/lib/Target/AMDGPU/Utils/AMDKernelCodeTUtils.cpp:113:5
    #7 0x3a46e6c in llvm::AMDGPUTargetAsmStreamer::EmitAMDKernelCodeT(amd_kernel_code_s const&) /code/llvm-project/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp:161:3
    #8 0xd371e4 in llvm::AMDGPUAsmPrinter::EmitFunctionBodyStart() /code/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp:204:26

[...]

Uninitialized value was created by an allocation of 'KernelCode' in the stack frame of function '_ZN4llvm16AMDGPUAsmPrinter21EmitFunctionBodyStartEv'
    #0 0xd36650 in llvm::AMDGPUAsmPrinter::EmitFunctionBodyStart() /code/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp:192

llvm-svn: 337079

6 years ago[x86/SLH] Add an assert to catch if we ever end up trying to harden
Chandler Carruth [Sat, 14 Jul 2018 00:52:09 +0000 (00:52 +0000)]
[x86/SLH] Add an assert to catch if we ever end up trying to harden
post-load a register that isn't valid for use with OR or SHRX.

llvm-svn: 337078

6 years ago[llvm-mca] Remove unused InstRef formal from pre and post execute callbacks. NFC.
Matt Davis [Sat, 14 Jul 2018 00:10:42 +0000 (00:10 +0000)]
[llvm-mca] Remove unused InstRef formal from pre and post execute callbacks. NFC.

llvm-svn: 337077

6 years ago[ThinLTO] Add debug output to test
Teresa Johnson [Sat, 14 Jul 2018 00:08:48 +0000 (00:08 +0000)]
[ThinLTO] Add debug output to test

Add -debug-only=function-import to get more information for debugging
reverse-iteration bot failure from r337050.

llvm-svn: 337076

6 years agoRe-apply "[SCEV] Strengthen StrengthenNoWrapFlags (reapply r334428)."
Tim Shen [Fri, 13 Jul 2018 23:58:46 +0000 (23:58 +0000)]
Re-apply "[SCEV] Strengthen StrengthenNoWrapFlags (reapply r334428)."

llvm-svn: 337075

6 years agoAdd a CHECK line for r337072.
Tim Shen [Fri, 13 Jul 2018 23:48:59 +0000 (23:48 +0000)]
Add a CHECK line for r337072.

llvm-svn: 337074

6 years ago[Hexagon] Avoid introducing calls into coalesced range of HVX vector pairs
Krzysztof Parzyszek [Fri, 13 Jul 2018 23:42:29 +0000 (23:42 +0000)]
[Hexagon] Avoid introducing calls into coalesced range of HVX vector pairs

If an HVX vector register is to be coalesced into a vector pair, make
sure that the vector pair will not have a function call in its live range,
unless it already had one. All HVX vector registers are volatile, so
any vector register live across a function call will have to be spilled.

If a vector needs to be spilled, and it's coalesced into a vector pair
then the whole pair will need to be spilled (even if only a part of it is
live), taking extra stack space.

llvm-svn: 337073

6 years ago[LSR] If no Use is interesting, early return.
Tim Shen [Fri, 13 Jul 2018 23:40:00 +0000 (23:40 +0000)]
[LSR] If no Use is interesting, early return.

Summary:
By looking at the callers of getUse(), we can see that even though
IVUsers may offer uses, but they may not be interesting to
LSR. It's possible that none of them is interesting.

Reviewers: sanjoy

Subscribers: jlebar, hiraditya, bixia, llvm-commits

Differential Revision: https://reviews.llvm.org/D49049

llvm-svn: 337072

6 years agoRollback r337070.
Sterling Augustine [Fri, 13 Jul 2018 23:03:15 +0000 (23:03 +0000)]
Rollback r337070.

Someone simultaneously fixed the breakage it was designed to fix.

llvm-svn: 337071

6 years agoUpdate ClangASTContext for the new DependentVector type.
Sterling Augustine [Fri, 13 Jul 2018 22:54:41 +0000 (22:54 +0000)]
Update ClangASTContext for the new DependentVector type.

https://reviews.llvm.org/D49326

llvm-svn: 337070

6 years ago[Documentation] Add missing description for bugprone-exception-escape in Release...
Eugene Zelenko [Fri, 13 Jul 2018 22:53:05 +0000 (22:53 +0000)]
[Documentation] Add missing description for bugprone-exception-escape in Release Notes.

llvm-svn: 337069

6 years ago[UBSan] Followup for silence_unsigned_overflow flag to handle negate overflows.
Max Moroz [Fri, 13 Jul 2018 22:49:06 +0000 (22:49 +0000)]
[UBSan] Followup for silence_unsigned_overflow flag to handle negate overflows.

Summary:
That flag has been introduced in https://reviews.llvm.org/D48660 for
suppressing UIO error messages in an efficient way. The main motivation is to
be able to use UIO checks in builds used for fuzzing as it might provide an
interesting signal to a fuzzing engine such as libFuzzer.

See https://github.com/google/oss-fuzz/issues/910 for more information.

Reviewers: morehouse, kcc

Reviewed By: morehouse

Subscribers: kubamracek, delcypher, #sanitizers, llvm-commits

Differential Revision: https://reviews.llvm.org/D49324

llvm-svn: 337068

6 years ago[X86][SLH] Remove PDEP and PEXT from isDataInvariantLoad
Craig Topper [Fri, 13 Jul 2018 22:41:52 +0000 (22:41 +0000)]
[X86][SLH] Remove PDEP and PEXT from isDataInvariantLoad

Ryzen has something like an 18 cycle latency on these based on Agner's data. AMD's own xls is blank. So it seems like there might be something tricky here.

Agner's data for Intel CPUs indicates these are a single uop there.

Probably safest to remove them. We never generate them without an intrinsic so this should be ok.

Differential Revision: https://reviews.llvm.org/D49315

llvm-svn: 337067

6 years ago[X86][SLH] Add VEX and EVEX conversion instructions to isDataInvariantLoad
Craig Topper [Fri, 13 Jul 2018 22:41:50 +0000 (22:41 +0000)]
[X86][SLH] Add VEX and EVEX conversion instructions to isDataInvariantLoad

-Drop the intrinsic versions of conversion instructions. These should be handled when we do vectors. They shouldn't show up in scalar code.
-Add the float<->double conversions which were missing.
-Add the AVX512 and AVX version of the conversion instructions including the unsigned integer conversions unique to AVX512

Differential Revision: https://reviews.llvm.org/D49313

llvm-svn: 337066

6 years ago[X86][SLH] Regroup the instructions in isDataInvariantLoad a little. NFC
Craig Topper [Fri, 13 Jul 2018 22:41:46 +0000 (22:41 +0000)]
[X86][SLH] Regroup the instructions in isDataInvariantLoad a little. NFC

-Move BSF/BSR to the same group as TZCNT/LZCNT/POPCNT.
-Split some of the bit manipulation instructions away from TZCNT/LZCNT/POPCNT. These are things like 'x & (x - 1)' which are composed of a few simple arithmetic operations. These aren't nearly as complicated/surprising as counting bits.
-Move BEXTR/BZHI into their own group. They aren't like a simple arithmethic op or the bit manipulation instructions. They're more like a shift+and.

Differential Revision: https://reviews.llvm.org/D49312

llvm-svn: 337065

6 years ago[lldb-mi] Make symbol-list-lines.test XFAIL on Windows
Alexander Polyakov [Fri, 13 Jul 2018 22:41:16 +0000 (22:41 +0000)]
[lldb-mi] Make symbol-list-lines.test XFAIL on Windows

It's a temporary decision until we reach out what causes
a failure.

llvm-svn: 337064

6 years agoFix -Wswitch after introduction of clang;:Type::DependentVector in r337036
Fangrui Song [Fri, 13 Jul 2018 22:40:40 +0000 (22:40 +0000)]
Fix -Wswitch after introduction of clang;:Type::DependentVector in r337036

llvm-svn: 337063

6 years ago[docs] Update usage directive for llvm-cov report -show-functions
Vedant Kumar [Fri, 13 Jul 2018 22:39:31 +0000 (22:39 +0000)]
[docs] Update usage directive for llvm-cov report -show-functions

llvm-svn: 337062

6 years agoFix comments which mixed up 'before' and 'after', NFC
Vedant Kumar [Fri, 13 Jul 2018 22:39:31 +0000 (22:39 +0000)]
Fix comments which mixed up 'before' and 'after', NFC

llvm-svn: 337061

6 years agoClarify wording of a doxygen comment, NFC
Vedant Kumar [Fri, 13 Jul 2018 22:39:29 +0000 (22:39 +0000)]
Clarify wording of a doxygen comment, NFC

llvm-svn: 337060

6 years ago[ThinLTO] Require x86 target for new test
Teresa Johnson [Fri, 13 Jul 2018 22:36:22 +0000 (22:36 +0000)]
[ThinLTO] Require x86 target for new test

Should fix non-x86 bot failures for new test from r337050.

llvm-svn: 337059

6 years agoMake these tests c++ tests so they can be skipped on systems that don't support those...
Jim Ingham [Fri, 13 Jul 2018 22:31:59 +0000 (22:31 +0000)]
Make these tests c++ tests so they can be skipped on systems that don't support those tests.

llvm-svn: 337058

6 years ago[X86] Use the correct types in some recently added isel patterns.
Craig Topper [Fri, 13 Jul 2018 22:27:53 +0000 (22:27 +0000)]
[X86] Use the correct types in some recently added isel patterns.

These were supposed to be integer types since we are selecting integer instructions.

Found while preparing to remove these patterns for another patch.

llvm-svn: 337057

6 years agoAMDGPU/GlobalISel: Implement select() for 32-bit @llvm.minnun and @llvm.maxnum
Tom Stellard [Fri, 13 Jul 2018 22:16:03 +0000 (22:16 +0000)]
AMDGPU/GlobalISel: Implement select() for 32-bit @llvm.minnun and @llvm.maxnum

Reviewers: arsenm, nhaehnle

Subscribers: kzhuravl, wdng, yaxunl, rovka, kristof.beyls, dstuttard, tpr, llvm-commits, t-tye

Differential Revision: https://reviews.llvm.org/D46172

llvm-svn: 337056

6 years ago[X86][FastISel] Support uitofp with avx512.
Craig Topper [Fri, 13 Jul 2018 22:09:30 +0000 (22:09 +0000)]
[X86][FastISel] Support uitofp with avx512.

llvm-svn: 337055

6 years ago[Polly][isl] Add neutrally-named accessors to isl list elements and sizes
Philip Pfaffe [Fri, 13 Jul 2018 22:05:01 +0000 (22:05 +0000)]
[Polly][isl] Add neutrally-named accessors to isl list elements and sizes

Summary: This could simplify the isl iterator implementation a lot.

Reviewers: grosser, Meinersbur, bollu

Reviewed By: grosser

Subscribers: pollydev, llvm-commits

Differential Revision: https://reviews.llvm.org/D49019

llvm-svn: 337054

6 years ago[LTO] Fix linking with an alias defined using another alias.
Eli Friedman [Fri, 13 Jul 2018 21:58:55 +0000 (21:58 +0000)]
[LTO] Fix linking with an alias defined using another alias.

When we're linking an alias which will be defined later, we neeed to
build a GlobalAlias, or else we'll crash later in
IRLinker::linkGlobalValueBody.

clang sometimes constructs aliases like this for C++ destructors.

Differential Revision: https://reviews.llvm.org/D49316

llvm-svn: 337053

6 years ago[X86] Correct comment of TEST elimination in BSF/TZCNT
Fangrui Song [Fri, 13 Jul 2018 21:40:08 +0000 (21:40 +0000)]
[X86] Correct comment of TEST elimination in BSF/TZCNT

llvm-svn: 337052

6 years ago[ThinLTO] Ensure we always select the same function copy to import
Teresa Johnson [Fri, 13 Jul 2018 21:35:58 +0000 (21:35 +0000)]
[ThinLTO] Ensure we always select the same function copy to import

Clang change to reflect the FunctionsToImportTy type change
in the llvm changes for D48670.

llvm-svn: 337051

6 years ago[ThinLTO] Ensure we always select the same function copy to import
Teresa Johnson [Fri, 13 Jul 2018 21:35:51 +0000 (21:35 +0000)]
[ThinLTO] Ensure we always select the same function copy to import

In order to always import the same copy of a linkonce function,
even when encountering it with different thresholds (a higher one then a
lower one), keep track of the summary we decided to import.
This ensures that the backend only gets a single definition to import
for each GUID, so that it doesn't need to choose one.

Move the largest threshold the GUID was considered for import into the
current module out of the ImportMap (which is part of a larger map
maintained across the whole index), and into a new map just maintained
for the current module we are computing imports for. This saves some
memory since we no longer have the thresholds maintained across the
whole index (and throughout the in-process backends when doing a normal
non-distributed ThinLTO build), at the cost of some additional
information being maintained for each invocation of ComputeImportForModule
(the selected summary pointer for each import).

There is an additional map lookup for each callee being considered for
importing, however, this was able to subsume a map lookup in the
Worklist iteration that invokes computeImportForFunction. We also are
able to avoid calling selectCallee if we already failed to import at the
same or higher threshold.

I compared the run time and peak memory for the SPEC2006 471.omnetpp
benchmark (running in-process ThinLTO backends), as well as for a large
internal benchmark with a distributed ThinLTO build (so just looking at
the thin link time/memory). Across a number of runs with and without
this change there was no significant change in the time and memory.

(I tried a few other variations of the change but they also didn't
improve time or peak memory).

Reviewers: davidxl

Subscribers: mehdi_amini, inglorion, llvm-commits

Differential Revision: https://reviews.llvm.org/D48670

llvm-svn: 337050