OSDN Git Service

android-x86/external-llvm.git
5 years ago[llvm-pdbutil] Pretty print PDBSymbolUsingNamespace symbols
Aaron Smith [Thu, 11 Oct 2018 21:37:18 +0000 (21:37 +0000)]
[llvm-pdbutil] Pretty print PDBSymbolUsingNamespace symbols

Reviewers: rnk, zturner, llvm-commits

Differential Revision: https://reviews.llvm.org/D52799

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344298 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Type legalize v2f32 loads by using an f64 load and a scalar_to_vector.
Craig Topper [Thu, 11 Oct 2018 20:36:06 +0000 (20:36 +0000)]
[X86] Type legalize v2f32 loads by using an f64 load and a scalar_to_vector.

On 64-bit targets the generic legalize will use an i64 load and a scalar_to_vector for us. But on 32-bit targets i64 isn't legal and the generic legalizer will end up emitting two 32-bit loads. We have DAG combines that try to put those two loads back together with pretty good success.

This patch instead uses f64 to avoid the splitting entirely. I've made it do the same for 64-bit mode for consistency and to keep the load in the fp domain.

There are a few things in here that look like regressions in 32-bit mode, but I believe they bring us closer to the 64-bit mode codegen. And that the 64-bit mode code could be better. I think those issues should be looked at separately.

Differential Revision: https://reviews.llvm.org/D52528

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344291 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WebAssembly][NFC] Remove repetition of Defs = [ARGUMENTS] (fixed)
Thomas Lively [Thu, 11 Oct 2018 20:21:22 +0000 (20:21 +0000)]
[WebAssembly][NFC] Remove repetition of Defs = [ARGUMENTS] (fixed)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344287 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoUpdate test of r344198 to work with release builds.
Warren Ristow [Thu, 11 Oct 2018 20:19:25 +0000 (20:19 +0000)]
Update test of r344198 to work with release builds.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344286 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Hexagon] Restrict compound instructions with constant value.
Sumanth Gundapaneni [Thu, 11 Oct 2018 19:48:15 +0000 (19:48 +0000)]
[Hexagon] Restrict compound instructions with constant value.

Having a constant value operand in the compound instruction
is not always profitable. This patch improves coremark by ~4% on
Hexagon.

Differential Revision: https://reviews.llvm.org/D53152

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344284 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Pipeliner] Use the Index from Topo instead of relying on NodeNum. (NFC)
Sumanth Gundapaneni [Thu, 11 Oct 2018 19:45:07 +0000 (19:45 +0000)]
[Pipeliner] Use the Index from Topo instead of relying on NodeNum. (NFC)

In future, if we may add any new DAG mutations other than artificial dependencies,
the NodeNum may not be valid. Instead the index from topological schedule DAG can be
used as long as we update it with the DAG change.

Differential Revision: https://reviews.llvm.org/D53104

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344283 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Pipeliner] Fix the Schedule DAG topoligical order.
Sumanth Gundapaneni [Thu, 11 Oct 2018 19:42:46 +0000 (19:42 +0000)]
[Pipeliner] Fix the Schedule DAG topoligical order.

This patch updates the DAG change to reflect in the topological ordering
of the nodes.

Differential Revision: https://reviews.llvm.org/D53105

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344282 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WebAssembly] Revert rL344180, which was breaking expensive checks
Thomas Lively [Thu, 11 Oct 2018 18:45:48 +0000 (18:45 +0000)]
[WebAssembly] Revert rL344180, which was breaking expensive checks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344280 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert SymbolFileNativePDB plugin.
Zachary Turner [Thu, 11 Oct 2018 18:45:44 +0000 (18:45 +0000)]
Revert SymbolFileNativePDB plugin.

This was originally causing some test failures on non-Windows
platforms, which required fixes in the compiler and linker.  After
those fixes, however, other tests started failing.  Reverting
temporarily until I can address everything.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344279 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert r344197 "[MC][ELF] compute entity size for explicit sections"
Artem Dergachev [Thu, 11 Oct 2018 18:43:08 +0000 (18:43 +0000)]
Revert r344197 "[MC][ELF] compute entity size for explicit sections"

Revert r344206 "[MC][ELF] Fix section_mergeable_size.ll"

They were causing failures on too many important buildbots for too long.
Please revert eagerly if your fix takes more than a couple of hours to land!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344278 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[PassManager/Sanitizer] Port of AddresSanitizer pass from legacy to new PassManager
Leonard Chan [Thu, 11 Oct 2018 18:31:51 +0000 (18:31 +0000)]
[PassManager/Sanitizer] Port of AddresSanitizer pass from legacy to new PassManager

This patch ports the legacy pass manager to the new one to take advantage of
the benefits of the new PM. This involved moving a lot of the declarations for
`AddressSantizer` to a header so that it can be publicly used via
PassRegistry.def which I believe contains all the passes managed by the new PM.

This patch essentially decouples the instrumentation from the legacy PM such
hat it can be used by both legacy and new PM infrastructure.

Differential Revision: https://reviews.llvm.org/D52739

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344274 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DAG] Fix Big Endian in Load-Store forwarding
Nirav Dave [Thu, 11 Oct 2018 18:28:59 +0000 (18:28 +0000)]
[DAG] Fix Big Endian in Load-Store forwarding

Summary:
Correct offset calculation in load-store forwarding for big-endian
targets.

Reviewers: rnk, RKSimon, waltl

Subscribers: sdardis, nemanjai, hiraditya, jrtc27, atanasyan, jsji, llvm-commits

Differential Revision: https://reviews.llvm.org/D53147

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344272 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Hexagon] Eliminate potential sources of non-determinism in HCE
Krzysztof Parzyszek [Thu, 11 Oct 2018 18:26:02 +0000 (18:26 +0000)]
[Hexagon] Eliminate potential sources of non-determinism in HCE

Also, avoid comparing GUIDs when ordering global addresses, because
source file location can cause different GUID to be calculated. As a
result, a pair of symbols can compare "less" in one directory, but
"greater" in another.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344271 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Restore X86ISelDAGToDAG::matchBEXTRFromAnd. Teach address matching to create...
Craig Topper [Thu, 11 Oct 2018 18:06:07 +0000 (18:06 +0000)]
[X86] Restore X86ISelDAGToDAG::matchBEXTRFromAnd. Teach address matching to create a BEXTR pattern from a (shl (and X, mask >> C1) if C1 can be folded into addressing mode.

This is an alternative to D53080 since I think using a BEXTR for a shifted mask is definitely an improvement when the shl can be absorbed into addressing mode. The other cases I'm less sure about.

We already have several tricks for handling an and of a shift in address matching. This adds a new case for BEXTR.

I've moved the BEXTR matching code back to X86ISelDAGToDAG to allow it to match. I suppose alternatively we could directly emit a X86ISD::BEXTR node that isel could pattern match. But I'm trying to view BEXTR matching as an isel concern so DAG combine can see 'and' and 'shift' operations that are well understood. We did lose a couple cases from tbm_patterns.ll, but I think there are ways to recover that.

I've also put back the manual load folding code in matchBEXTRFromAnd that I removed a few months ago in r324939. This gives us some more freedom to make decisions based on the ability to fold a load. I haven't done anything with that yet.

Differential Revision: https://reviews.llvm.org/D53126

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344270 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoBetter support for POSIX paths in PDBs.
Zachary Turner [Thu, 11 Oct 2018 18:01:55 +0000 (18:01 +0000)]
Better support for POSIX paths in PDBs.

While it doesn't make a *ton* of sense for POSIX paths to be
in PDBs, it's possible to occur in real scenarios involving
cross compilation.

The tools need to be able to handle this, because certain types
of debugging scenarios are possible without a running process
and so don't necessarily require you to be on a Windows system.
These include post-mortem debugging and binary forensics (e.g.
using a debugger to disassemble functions and examine symbols
without running the process).

There's changes in clang, LLD, and lldb in this patch.  After
this the cross-platform disassembly and source-list tests pass
on Linux.

Furthermore, the behavior of LLD can now be summarized by a much
simpler rule than before: Unless you specify /pdbsourcepath and
/pdbaltpath, the PDB ends up with paths that are valid within
the context of the machine that the link is performed on.

Differential Revision: https://reviews.llvm.org/D53149

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344269 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-nm] Fix crash when running with --print-armap on corrupt archives.
Jordan Rupprecht [Thu, 11 Oct 2018 17:55:11 +0000 (17:55 +0000)]
[llvm-nm] Fix crash when running with --print-armap on corrupt archives.

error() in llvm-nm intentionally does not return so that the callee can move on to future files/slices. When printing the archive map, this is not currently handled (the caller assumes that error() returns), so processing continues despite there being an error.

Also, change one return to a break, so that symbols can be printed even if the archive map is corrupt.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344268 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DAGCombiner] move comment closer to the corresponding code; NFC
Sanjay Patel [Thu, 11 Oct 2018 16:07:25 +0000 (16:07 +0000)]
[DAGCombiner] move comment closer to the corresponding code; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344255 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-mca][BtVer2] Add tests for optimizable GPR register moves. NFC
Andrea Di Biagio [Thu, 11 Oct 2018 14:54:54 +0000 (14:54 +0000)]
[llvm-mca][BtVer2] Add tests for optimizable GPR register moves. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344253 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine] Fix SimplifyLibCalls erasing an instruction while IC still had referenc...
Amara Emerson [Thu, 11 Oct 2018 14:51:11 +0000 (14:51 +0000)]
[InstCombine] Fix SimplifyLibCalls erasing an instruction while IC still had references to it.

InstCombine keeps a worklist and assumes that optimizations don't
eraseFromParent() the instruction, which SimplifyLibCalls violates. This change
adds a new callback to SimplifyLibCalls to let clients specify their own hander
for erasing actions.

Differential Revision: https://reviews.llvm.org/D52729

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344251 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AARCH64][FIX] Emit data symbol for constant pool data
Diogo N. Sampaio [Thu, 11 Oct 2018 14:10:32 +0000 (14:10 +0000)]
[AARCH64][FIX] Emit data symbol for constant pool data

The ARM64 elf emitter would omit printing data
symbol for zero filled constant data. This patch
overrides the emitFill method as to enforce that
the symbol is correctly printed.

Differential revision: https://reviews.llvm.org/D53132

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344248 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoGeneralize an IR verifier check to work with non-zero program address spaces
Dylan McKay [Thu, 11 Oct 2018 12:49:50 +0000 (12:49 +0000)]
Generalize an IR verifier check to work with non-zero program address spaces

This commit modifies an existing IR verifier check that
assumes all functions will be located in the default address
space 0.

Rather than using the default paramater value getPointerTo(AddrSpace=0),
explicitly specify the program memory address space from the data layout.

This only affects targets that specify a nonzero address space
in their data layouts. The only in-tree target that does this
is AVR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344243 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine] Demand bits of UMin
David Green [Thu, 11 Oct 2018 11:28:27 +0000 (11:28 +0000)]
[InstCombine] Demand bits of UMin

This is the umin alternative to the umax code from rL344237. We use
DeMorgans law on the umax case to bring us to the same thing on umin,
but using countLeadingOnes, not countLeadingZeros.

Differential Revision: https://reviews.llvm.org/D53036

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344239 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[RISCV] Re-generate test/CodeGen/RISCV/vararg.ll after r344142
Alex Bradbury [Thu, 11 Oct 2018 11:11:58 +0000 (11:11 +0000)]
[RISCV] Re-generate test/CodeGen/RISCV/vararg.ll after r344142

The improved load-store forwarding committed in r344142 broke this test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344238 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine] Demand bits of UMax
David Green [Thu, 11 Oct 2018 11:04:09 +0000 (11:04 +0000)]
[InstCombine] Demand bits of UMax

Use the demanded bits of umax(A,C) to prove we can just use A so long as the
lowest non-zero bit of DemandMask is higher than the highest non-zero bit of C

Differential Revision: https://reviews.llvm.org/D53033

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344237 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine] Add tests for demand bits of min/max. NFC.
David Green [Thu, 11 Oct 2018 10:46:12 +0000 (10:46 +0000)]
[InstCombine] Add tests for demand bits of min/max. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344236 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[tblgen][CodeGenSchedule] Add a check for invalid RegisterFile definitions with zero...
Andrea Di Biagio [Thu, 11 Oct 2018 10:39:03 +0000 (10:39 +0000)]
[tblgen][CodeGenSchedule] Add a check for invalid RegisterFile definitions with zero physical registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344235 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[LV] Use SmallVector instead of DenseMap in calculateRegisterUsage (NFC).
Florian Hahn [Thu, 11 Oct 2018 09:46:25 +0000 (09:46 +0000)]
[LV] Use SmallVector instead of DenseMap in calculateRegisterUsage (NFC).

We assign indices sequentially for seen instructions, so we can just use
a vector and push back the seen instructions. No need for using a
DenseMap.

Reviewers: hsaito, rengolin, nadav, dcaballe

Reviewed By: rengolin

Differential Revision: https://reviews.llvm.org/D53089

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344233 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[LV] Ignore more debug info.
Florian Hahn [Thu, 11 Oct 2018 09:27:24 +0000 (09:27 +0000)]
[LV] Ignore more debug info.

We can avoid doing some unnecessary work by skipping debug instructions
in a few loops. It also helps to ensure debug instructions do not
prevent vectorization, although I do not have any concrete test cases
for that.

Reviewers: rengolin, hsaito, dcaballe, aprantl, vsk

Reviewed By: rengolin, dcaballe

Differential Revision: https://reviews.llvm.org/D53091

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344232 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[gcov] Display the hit counter for the line of a function definition
Calixte Denizet [Thu, 11 Oct 2018 08:53:43 +0000 (08:53 +0000)]
[gcov] Display the hit counter for the line of a function definition

Summary:
Right now there is no hit counter on the line of function.
So the idea is add the line of the function to all the lines covered by the entry block.
Tests in compiler-rt/profile will be fixed in another patch: https://reviews.llvm.org/D49854

Reviewers: marco-c, davidxl

Reviewed By: marco-c

Subscribers: sylvestre.ledru, llvm-commits

Differential Revision: https://reviews.llvm.org/D49853

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344228 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[NFC] Factor out getOrCreateAddRecExpr method
Max Kazantsev [Thu, 11 Oct 2018 08:46:39 +0000 (08:46 +0000)]
[NFC] Factor out getOrCreateAddRecExpr method

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344227 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86][BMI1]: X86DAGToDAGISel: select BEXTR from x & ~(-1 << nbits) pattern
Roman Lebedev [Thu, 11 Oct 2018 07:51:13 +0000 (07:51 +0000)]
[X86][BMI1]: X86DAGToDAGISel: select BEXTR from x & ~(-1 << nbits) pattern

Summary:
As discussed in D48491, we can't really do this in the TableGen,
since we need to produce *two* instructions. This only implements
one single pattern. The other 3 patterns will be in follow-ups.

I'm not sure yet if we want to also fuse shift into here
(i.e `(x >> start) & ...`)

Reviewers: RKSimon, craig.topper, spatel

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D52304

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344224 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[IndVars] Drop "exact" flag from lshr and udiv when substituting their args
Max Kazantsev [Thu, 11 Oct 2018 07:22:26 +0000 (07:22 +0000)]
[IndVars] Drop "exact" flag from lshr and udiv when substituting their args

There is a transform that may replace `lshr (x+1), 1` with `lshr x, 1` in case
if it can prove that the result will be the same. However the initial instruction
might have an `exact` flag set, and it now should be dropped unless we prove
that it may hold. Incorrectly set `exact` attribute may then produce poison.

Differential Revision: https://reviews.llvm.org/D53061
Reviewed By: sanjoy

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344223 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-nm] Include the text "@FILE" in the output of --help
Martin Storsjo [Thu, 11 Oct 2018 06:53:38 +0000 (06:53 +0000)]
[llvm-nm] Include the text "@FILE" in the output of --help

libtool requires this text to be present, in order to conclude that
the tool supports response files. Also add an explicit test of using
response files with llvm-nm.

Differential Revision: https://reviews.llvm.org/D53064

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344222 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[CMake] Temporarily remove the LLVM_ENABLE_IDE option
Chris Bieneman [Thu, 11 Oct 2018 04:06:14 +0000 (04:06 +0000)]
[CMake] Temporarily remove the LLVM_ENABLE_IDE option

All uses of this option have been removed, and the intent is to change the purpose and default value of this option. To prevent it from having impacts on users, this patch temporarily removes the option and purges it from CMake caches. In a few days, once this has propagated to contributors I will re-introduce the option with the new default value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344219 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[CMake] Unconditionally add .h and .td files to target sources
Chris Bieneman [Thu, 11 Oct 2018 04:02:53 +0000 (04:02 +0000)]
[CMake] Unconditionally add .h and .td files to target sources

Previously adding header and table gen files was conditional on using an IDE. Since these files have the `HEADER_FILE_ONLY` attribute applied they are ignored as sources by all non-IDE generators, so there is really no reason not to include them.

Additionally having the CMake always include these files allows the CMake-server to include them in the sources list for targets, which is valuable to anyone using CMake-server integrated tools.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344218 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Coverage] Apply filtered paths to summary
Chris Bieneman [Thu, 11 Oct 2018 04:00:51 +0000 (04:00 +0000)]
[Coverage] Apply filtered paths to summary

Summary:
The script to generate code coverage reports supports passing filter paths to llvm-cov when generating the HTML reports, but doesn't pass those paths to the summary generation as well. This results in a summary report that doesn't match the HTML report.

This patch addresses the problem by also passing the filter paths to the summary report generation.

Reviewers: vsk

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D53110

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344217 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoUse fully qualified namespace name.
Zachary Turner [Thu, 11 Oct 2018 03:42:17 +0000 (03:42 +0000)]
Use fully qualified namespace name.

llvm::detail is not the only namespace named detail.  So if
someone has done a `using namespace llvm::support`, for example,
this will fail with an ambiguous namespace name.  Granted
people generally shouldn't be using large namespaces like that,
but it's common at local function scopes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344216 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WebAssembly][NFC] Use intrinsic dag nodes directly
Thomas Lively [Thu, 11 Oct 2018 00:49:24 +0000 (00:49 +0000)]
[WebAssembly][NFC] Use intrinsic dag nodes directly

Summary: Instead of custom lowering to WebAssemblyISD nodes first.

Reviewers: aheejin, dschuff

Subscribers: sbc100, jgravelle-google, sunfish, llvm-commits

Differential Revision: https://reviews.llvm.org/D53119

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344211 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[MC][ELF] Fix section_mergeable_size.ll
Fangrui Song [Thu, 11 Oct 2018 00:08:59 +0000 (00:08 +0000)]
[MC][ELF] Fix section_mergeable_size.ll

Some targets use %progbits instead of @progbits.

Updating that check with a {{[@%]}}progbits regex to make those bots happy.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344206 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WebAssembly] Saturating float to int intrinsics
Thomas Lively [Thu, 11 Oct 2018 00:01:25 +0000 (00:01 +0000)]
[WebAssembly] Saturating float to int intrinsics

Summary:
Although the saturating float to int instructions are already
emitted from normal IR, the fpto{s,u}i instructions produce poison
values if the argument cannot fit in the result type. These intrinsics
are therefore necessary to get guaranteed defined saturating behavior.

Reviewers: aheejin, dschuff

Subscribers: sbc100, jgravelle-google, sunfish, llvm-commits

Differential Revision: https://reviews.llvm.org/D53004

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344204 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agollvm-c: Add C APIs to access DebugLoc info
Saleem Abdulrasool [Wed, 10 Oct 2018 23:53:12 +0000 (23:53 +0000)]
llvm-c: Add C APIs to access DebugLoc info

Add thin shims to C interface to provide access to DebugLoc info for
Instructions, GlobalVariables and Functions.  Patch by Josh Berdine!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344202 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAdd a flag to remap manglings when reading profile data information.
Richard Smith [Wed, 10 Oct 2018 23:13:47 +0000 (23:13 +0000)]
Add a flag to remap manglings when reading profile data information.

This can be used to preserve profiling information across codebase
changes that have widespread impact on mangled names, but across which
most profiling data should still be usable. For example, when switching
from libstdc++ to libc++, or from the old libstdc++ ABI to the new ABI,
or even from a 32-bit to a 64-bit build.

The user can provide a remapping file specifying parts of mangled names
that should be treated as equivalent (eg, std::__1 should be treated as
equivalent to std::__cxx11), and profile data will be treated as
applying to a particular function if its name is equivalent to the name
of a function in the profile data under the provided equivalences. See
the documentation change for a description of how this is configured.

Remapping is supported for both sample-based profiling and instruction
profiling. We do not support remapping indirect branch target
information, but all other profile data should be remapped
appropriately.

Support is only added for the new pass manager. If someone wants to also
add support for this for the old pass manager, doing so should be
straightforward.

This is the LLVM side of Clang r344199.

Reviewers: davidxl, tejohnson, dlj, erik.pilkington

Subscribers: mehdi_amini, steven_wu, dexonsmith, llvm-commits

Differential Revision: https://reviews.llvm.org/D51249

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344200 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[LTO] Account for overriding lib calls via the alias attribute
Warren Ristow [Wed, 10 Oct 2018 22:54:31 +0000 (22:54 +0000)]
[LTO] Account for overriding lib calls via the alias attribute

Given a library call that is represented as an llvm intrinsic call, but
later transformed to an actual call, if an overriding definition of that
library routine is provided indirectly via an alias, prevent LTO from
eliminating the definition.

This is a fix for PR38547.

Differential Revision: https://reviews.llvm.org/D52836

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344198 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[MC][ELF] compute entity size for explicit sections
Nick Desaulniers [Wed, 10 Oct 2018 22:52:32 +0000 (22:52 +0000)]
[MC][ELF] compute entity size for explicit sections

Summary:
Global variables might declare themselves to be in explicit sections.
Calculate the entity size always to prevent assembler warnings
"entity size for SHF_MERGE not specified" when sections are to be
marked merge-able.

Fixes PR31828.

Reviewers: rnk, echristo

Reviewed By: rnk

Subscribers: llvm-commits, pirama, srhines

Differential Revision: https://reviews.llvm.org/D53056

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344197 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Prevent non-temporal loads from folding into instructions by blocking them...
Craig Topper [Wed, 10 Oct 2018 21:48:34 +0000 (21:48 +0000)]
[X86] Prevent non-temporal loads from folding into instructions by blocking them in X86DAGToDAGISel::IsProfitableToFold rather than with a predicate.

Remove tryFoldVecLoad since tryFoldLoad would call IsProfitableToFold and pick up the new check.

This saves about 5K out of ~600K on the generated isel table.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344189 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[CMake] NFC. Updating documentation on options
Chris Bieneman [Wed, 10 Oct 2018 21:36:12 +0000 (21:36 +0000)]
[CMake] NFC. Updating documentation on options

The Ninja pool options are only supported with the Ninja generator and
should be called out as such.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344188 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoSupport for remapping profile data when symbols change, for sample-based
Richard Smith [Wed, 10 Oct 2018 21:31:01 +0000 (21:31 +0000)]
Support for remapping profile data when symbols change, for sample-based
profiling.

Reviewers: davidxl, tejohnson, dlj, erik.pilkington

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D51248

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344187 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoReplace most users of UnknownSize with LocationSize::unknown(); NFC
George Burgess IV [Wed, 10 Oct 2018 21:28:44 +0000 (21:28 +0000)]
Replace most users of UnknownSize with LocationSize::unknown(); NFC

Moving away from UnknownSize is part of the effort to migrate us to
LocationSizes (e.g. the cleanup promised in D44748).

This doesn't entirely remove all of the uses of UnknownSize; some uses
require tweaks to assume that UnknownSize isn't just some kind of int.
This patch is intended to just be a trivial replacement for all places
where LocationSize::unknown() will Just Work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344186 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoTest commit: fix typo in comment
Armando Montanez [Wed, 10 Oct 2018 21:16:57 +0000 (21:16 +0000)]
Test commit: fix typo in comment

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344185 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoSupport for remapping profile data when symbols change, for
Richard Smith [Wed, 10 Oct 2018 21:09:37 +0000 (21:09 +0000)]
Support for remapping profile data when symbols change, for
instrumentation-based profiling.

Reviewers: davidxl, tejohnson, dlj, erik.pilkington

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D51247

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344184 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agollvm-ar: Darwin archive format fixes.
James Y Knight [Wed, 10 Oct 2018 21:07:02 +0000 (21:07 +0000)]
llvm-ar: Darwin archive format fixes.

* Support writing the DARWIN64 symbol table format.

* In darwin archives, emit a symbol table whenever requested, even
  when there are no members, as the apple linker will abort if given
  an archive without a symbol table.

Added tests for same, and also simplified and moved the GNU 64-bit
symbol table test into archive-symtab.test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344183 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[NFC][X86][AArch64] extract-bits.ll: add tests with constants+storing results.
Roman Lebedev [Wed, 10 Oct 2018 20:50:52 +0000 (20:50 +0000)]
[NFC][X86][AArch64] extract-bits.ll: add tests with constants+storing results.

As noted in https://reviews.llvm.org/D53080#inline-467678,
this *may* get pessimized by that diff.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344182 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine] reverse 'trunc X to <N x i1>' canonicalization; 2nd try
Sanjay Patel [Wed, 10 Oct 2018 20:47:46 +0000 (20:47 +0000)]
[InstCombine] reverse 'trunc X to <N x i1>' canonicalization; 2nd try

Re-trying r344082 because it unintentionally included extra diffs.

Original commit message:
icmp ne (and X, 1), 0 --> trunc X to N x i1

Ideally, we'd do the same for scalars, but there will likely be
regressions unless we add more trunc folds as we're doing here
for vectors.

The motivating vector case is from PR37549:
https://bugs.llvm.org/show_bug.cgi?id=37549

define <4 x float> @bitwise_select(<4 x float> %x, <4 x float> %y, <4 x float> %z, <4 x float> %w) {

  %c = fcmp ole <4 x float> %x, %y
  %s = sext <4 x i1> %c to <4 x i32>
  %s1 = shufflevector <4 x i32> %s, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 1, i32 1>
  %s2 = shufflevector <4 x i32> %s, <4 x i32> undef, <4 x i32> <i32 2, i32 2, i32 3, i32 3>
  %cond = or <4 x i32> %s1, %s2
  %condtr = trunc <4 x i32> %cond to <4 x i1>
  %r = select <4 x i1> %condtr, <4 x float> %z, <4 x float> %w
  ret <4 x float> %r

}

Here's a sampling of the vector codegen for that case using
mask+icmp (current behavior) vs. trunc (with this patch):

AVX before:

vcmpleps        %xmm1, %xmm0, %xmm0
vpermilps       $80, %xmm0, %xmm1 ## xmm1 = xmm0[0,0,1,1]
vpermilps       $250, %xmm0, %xmm0 ## xmm0 = xmm0[2,2,3,3]
vorps   %xmm0, %xmm1, %xmm0
vandps  LCPI0_0(%rip), %xmm0, %xmm0
vxorps  %xmm1, %xmm1, %xmm1
vpcmpeqd        %xmm1, %xmm0, %xmm0
vblendvps       %xmm0, %xmm3, %xmm2, %xmm0

AVX after:

vcmpleps        %xmm1, %xmm0, %xmm0
vpermilps       $80, %xmm0, %xmm1 ## xmm1 = xmm0[0,0,1,1]
vpermilps       $250, %xmm0, %xmm0 ## xmm0 = xmm0[2,2,3,3]
vorps   %xmm0, %xmm1, %xmm0
vblendvps       %xmm0, %xmm2, %xmm3, %xmm0

AVX512f before:

vcmpleps        %xmm1, %xmm0, %xmm0
vpermilps       $80, %xmm0, %xmm1 ## xmm1 = xmm0[0,0,1,1]
vpermilps       $250, %xmm0, %xmm0 ## xmm0 = xmm0[2,2,3,3]
vorps   %xmm0, %xmm1, %xmm0
vpbroadcastd    LCPI0_0(%rip), %xmm1 ## xmm1 = [1,1,1,1]
vptestnmd       %zmm1, %zmm0, %k1
vblendmps       %zmm3, %zmm2, %zmm0 {%k1}

AVX512f after:

vcmpleps        %xmm1, %xmm0, %xmm0
vpermilps       $80, %xmm0, %xmm1 ## xmm1 = xmm0[0,0,1,1]
vpermilps       $250, %xmm0, %xmm0 ## xmm0 = xmm0[2,2,3,3]
vorps   %xmm0, %xmm1, %xmm0
vpslld  $31, %xmm0, %xmm0
vptestmd        %zmm0, %zmm0, %k1
vblendmps       %zmm2, %zmm3, %zmm0 {%k1}

AArch64 before:

fcmge   v0.4s, v1.4s, v0.4s
zip1    v1.4s, v0.4s, v0.4s
zip2    v0.4s, v0.4s, v0.4s
orr     v0.16b, v1.16b, v0.16b
movi    v1.4s, #1
and     v0.16b, v0.16b, v1.16b
cmeq    v0.4s, v0.4s, #0
bsl     v0.16b, v3.16b, v2.16b

AArch64 after:

fcmge   v0.4s, v1.4s, v0.4s
zip1    v1.4s, v0.4s, v0.4s
zip2    v0.4s, v0.4s, v0.4s
orr     v0.16b, v1.16b, v0.16b
bsl     v0.16b, v2.16b, v3.16b

PowerPC-le before:

xvcmpgesp 34, 35, 34
vspltisw 0, 1
vmrglw 3, 2, 2
vmrghw 2, 2, 2
xxlor 0, 35, 34
xxlxor 35, 35, 35
xxland 34, 0, 32
vcmpequw 2, 2, 3
xxsel 34, 36, 37, 34

PowerPC-le after:

xvcmpgesp 34, 35, 34
vmrglw 3, 2, 2
vmrghw 2, 2, 2
xxlor 0, 35, 34
xxsel 34, 37, 36, 0

Differential Revision: https://reviews.llvm.org/D52747

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344181 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WebAssembly][NFC] Remove repetition of Defs = [ARGUMENTS]
Thomas Lively [Wed, 10 Oct 2018 20:40:54 +0000 (20:40 +0000)]
[WebAssembly][NFC] Remove repetition of Defs = [ARGUMENTS]

Summary:
By moving that line into the `I` multiclass.

Reviewers: aheejin

Subscribers: dschuff, sbc100, jgravelle-google, sunfish, llvm-commits

Differential Revision: https://reviews.llvm.org/D53093

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344180 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Move X86DAGToDAGISel::matchBEXTRFromAnd() into X86ISelLowering
Roman Lebedev [Wed, 10 Oct 2018 20:40:12 +0000 (20:40 +0000)]
[X86] Move X86DAGToDAGISel::matchBEXTRFromAnd() into X86ISelLowering

Summary:
As discussed in [[ https://bugs.llvm.org/show_bug.cgi?id=38938 | PR38938 ]],
we fail to emit `BEXTR` if the mask is shifted.
We can't deal with that in `X86DAGToDAGISel` `before the address mode for the inc is selected`,
and we can't really do it in the normal DAGCombine, because we don't have generic `ISD::BitFieldExtract` node,
and if we simply turn the shifted mask into a normal mask + shift-left, it will be folded back.
So it would seem X86ISelLowering is the place to handle this.

This patch only moves the matchBEXTRFromAnd()
from X86DAGToDAGISel to X86ISelLowering.
It does not add support for the 'shifted mask' pattern.

Reviewers: RKSimon, craig.topper, spatel

Reviewed By: RKSimon

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D52426

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344179 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agorevert r344082: [InstCombine] reverse 'trunc X to <N x i1>' canonicalization
Sanjay Patel [Wed, 10 Oct 2018 20:39:39 +0000 (20:39 +0000)]
revert r344082: [InstCombine] reverse 'trunc X to <N x i1>' canonicalization

This commit accidentally included the diffs from D53057.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344178 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DwarfVerifier] Fixed -Wimplicit-fallthrough warning
David Bolvansky [Wed, 10 Oct 2018 20:10:37 +0000 (20:10 +0000)]
[DwarfVerifier] Fixed -Wimplicit-fallthrough warning

Reviewers: JDevlieghere, RKSimon

Reviewed By: JDevlieghere

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D52963

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344176 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WebAssembly][NFC] Use vnot patfrag to simplify v128.not
Thomas Lively [Wed, 10 Oct 2018 19:09:16 +0000 (19:09 +0000)]
[WebAssembly][NFC] Use vnot patfrag to simplify v128.not

Reviewers: aheejin, dschuff

Subscribers: sbc100, jgravelle-google, sunfish, llvm-commits

Differential Revision: https://reviews.llvm.org/D53097

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344175 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[LV] Add a new reduction pattern match
Renato Golin [Wed, 10 Oct 2018 18:49:49 +0000 (18:49 +0000)]
[LV] Add a new reduction pattern match

Adding a new reduction pattern match for vectorizing code similar to TSVC s3111:

for (int i = 0; i < N; i++)
  if (a[i] > b)
    sum += a[i];

This patch adds support for fadd, fsub and fmull, as well as multiple
branches and different (but compatible) instructions (ex. add+sub) in
different branches.

I have forwarded to trunk, added fsub and fmul functionality and
additional tests, but the credit goes to Takahiro, who did most of the
actual work.

Differential Revision: https://reviews.llvm.org/D49168

Patch by Takahiro Miyoshi <takahiro.miyoshi@linaro.org>.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344172 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoReland: [OptRemarks] Add library for parsing optimization remarks
Francis Visoiu Mistrih [Wed, 10 Oct 2018 18:43:42 +0000 (18:43 +0000)]
Reland: [OptRemarks] Add library for parsing optimization remarks

Add a library that parses optimization remarks (currently YAML, so based
on the YAMLParser).

The goal is to be able to provide tools a remark parser that is not
completely dependent on YAML, in case we decide to change the format
later.

It exposes a C API which takes a handler that is called with the remark
structure.

It adds a libLLVMOptRemark.a static library, and it's used in-tree by
the llvm-opt-report tool (from which the parser has been mostly moved
out).

Differential Revision: https://reviews.llvm.org/D52776

Fixed the tests by removing the usage of C++11 strings, which seems not
to be supported by gcc 4.8.4 if they're used as a macro argument.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344171 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Support] Remove redundant qualifiers in YAMLTraits (NFC)
Scott Linder [Wed, 10 Oct 2018 18:14:02 +0000 (18:14 +0000)]
[Support] Remove redundant qualifiers in YAMLTraits (NFC)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344166 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert "[OptRemarks] Add library for parsing optimization remarks"
Francis Visoiu Mistrih [Wed, 10 Oct 2018 18:07:44 +0000 (18:07 +0000)]
Revert "[OptRemarks] Add library for parsing optimization remarks"

This reverts commit 1cc98e6672b6319fdb00b70dd4474aabdadbe193.

Seems to break bots: http://lab.llvm.org:8011/builders/clang-x86_64-linux-abi-test/builds/33398/steps/build-unified-tree/logs/stdio

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344164 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[GlobalISel] Fix the artifact combiner to fold G_IMPLICIT_DEF properly
Volkan Keles [Wed, 10 Oct 2018 18:01:48 +0000 (18:01 +0000)]
[GlobalISel] Fix the artifact combiner to fold G_IMPLICIT_DEF properly

Summary:
GlobalISel generates incorrect code because the legalizer artifact
combiner assumes `G_[SZ]EXT (G_IMPLICIT_DEF)` is equivalent to
`G_IMPLICIT_DEF `.

Replace `G_[SZ]EXT (G_IMPLICIT_DEF)` with 0 because the top bits
will be 0 for G_ZEXT and 0/1 for the G_SEXT.

Reviewers: aditya_nandakumar, dsanders, aemerson, javed.absar

Reviewed By: aditya_nandakumar

Subscribers: rovka, kristof.beyls, llvm-commits

Differential Revision: https://reviews.llvm.org/D52996

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344163 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[OptRemarks] Add library for parsing optimization remarks
Francis Visoiu Mistrih [Wed, 10 Oct 2018 17:58:09 +0000 (17:58 +0000)]
[OptRemarks] Add library for parsing optimization remarks

Add a library that parses optimization remarks (currently YAML, so based
on the YAMLParser).

The goal is to be able to provide tools a remark parser that is not
completely dependent on YAML, in case we decide to change the format
later.

It exposes a C API which takes a handler that is called with the remark
structure.

It adds a libLLVMOptRemark.a static library, and it's used in-tree by
the llvm-opt-report tool (from which the parser has been mostly moved
out).

Differential Revision: https://reviews.llvm.org/D52776

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344162 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[VPlan] Fix CondBit quoting in dumpBasicBlock
Renato Golin [Wed, 10 Oct 2018 17:55:21 +0000 (17:55 +0000)]
[VPlan] Fix CondBit quoting in dumpBasicBlock

Quotes were being printed for VPInstructions but not the rest.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344161 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoChange the timestamp of llvmcache-foo file to meet the thinLTO prune policy
Craig Topper [Wed, 10 Oct 2018 17:37:32 +0000 (17:37 +0000)]
Change the timestamp of llvmcache-foo file to meet the thinLTO prune policy

The case will randomly fail if we test it with command "
while llvm-lit test/tools/gold/X86/cache.ll ; do true; done". It is because the llvmcache-foo file is younger than llvmcache-349F039B8EB076D412007D82778442BED3148C4E and llvmcache-A8107945C65C2B2BBEE8E61AA604C311D60D58D6. But due to timestamp precision reason their timestamp is the same. Given the same timestamp, the file prune policy is to remove bigger size file first, so mostly foo file is removed for its bigger size. And the files size is under threshold after deleting foo file. That's what test case expect.

However sometimes, the precision is enough to measure that timestamp of llvmcache-349F039B8EB076D412007D82778442BED3148C4E and llvmcache-A8107945C65C2B2BBEE8E61AA604C311D60D58D6 are smaller than foo, so llvmcache-349F039B8EB076D412007D82778442BED3148C4E and llvmcache-A8107945C65C2B2BBEE8E61AA604C311D60D58D6 are deleted first. Since the files size is still above the file size threshold after deleting the 2 files, the foo file is also deleted. And then the test case fails, because it expect only one file should be deleted instead of 3.

The fix is to change the timestamp of llvmcache-foo file to meet the thinLTO prune policy.

Patch by Luo Yuanke.

Differential Revision: https://reviews.llvm.org/D52452

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344158 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRelax trivial cast requirements in CallPromotionUtils
Scott Linder [Wed, 10 Oct 2018 16:35:47 +0000 (16:35 +0000)]
Relax trivial cast requirements in CallPromotionUtils

Differential Revision: https://reviews.llvm.org/D52792

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344153 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-exegesis] Fix always true assert
Guillaume Chatelet [Wed, 10 Oct 2018 16:16:43 +0000 (16:16 +0000)]
[llvm-exegesis] Fix always true assert

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344151 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-mca] Minor refactoring in preparation for a patch that will fully fix PR36671...
Andrea Di Biagio [Wed, 10 Oct 2018 16:08:02 +0000 (16:08 +0000)]
[llvm-mca] Minor refactoring in preparation for a patch that will fully fix PR36671. NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344149 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-exegesis][NFC] Pass Instruction instead of bare Opcode
Guillaume Chatelet [Wed, 10 Oct 2018 14:57:32 +0000 (14:57 +0000)]
[llvm-exegesis][NFC] Pass Instruction instead of bare Opcode

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344145 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-mca][BtVer2] Add two more move-elimination tests. NFC
Andrea Di Biagio [Wed, 10 Oct 2018 14:46:54 +0000 (14:46 +0000)]
[llvm-mca][BtVer2] Add two more move-elimination tests. NFC

These should test all the optimizable moves on Jaguar.
A follow-up patch will teach how to recognize these optimizable register moves.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344144 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-exegesis][NFC] Code simplification
Guillaume Chatelet [Wed, 10 Oct 2018 14:22:48 +0000 (14:22 +0000)]
[llvm-exegesis][NFC] Code simplification

Summary: Simplify code by having LLVMState hold the RegisterAliasingTrackerCache.

Reviewers: courbet

Subscribers: tschuett, llvm-commits

Differential Revision: https://reviews.llvm.org/D53078

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344143 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DAGCombine] Improve Load-Store Forwarding
Nirav Dave [Wed, 10 Oct 2018 14:15:52 +0000 (14:15 +0000)]
[DAGCombine] Improve Load-Store Forwarding

Summary:
Extend analysis forwarding loads from preceeding stores to work with
extended loads and truncated stores to the same address so long as the
load is fully subsumed by the store.

Hexagon's swp-epilog-phis.ll and swp-memrefs-epilog1.ll test are
deleted as they've no longer seem to be relevant.

Reviewers: RKSimon, rnk, kparzysz, javed.absar

Subscribers: sdardis, nemanjai, hiraditya, atanasyan, llvm-commits

Differential Revision: https://reviews.llvm.org/D49200

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344142 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[x86] allow single source horizontal op matching (PR39195)
Sanjay Patel [Wed, 10 Oct 2018 13:39:59 +0000 (13:39 +0000)]
[x86] allow single source horizontal op matching (PR39195)

This is intended to restore horizontal codegen to what it looked like before IR demanded elements improved in:
rL343727

As noted in PR39195:
https://bugs.llvm.org/show_bug.cgi?id=39195
...horizontal ops can be worse for performance than a shuffle+regular binop, so I've added a TODO. Ideally, we'd
solve that in a machine instruction pass, but a quicker solution will be adding a 'HasFastHorizontalOp' feature
bit to deal with it here in the DAG.

Differential Revision: https://reviews.llvm.org/D52997

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344141 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoLift VFS from clang to llvm (NFC)
Jonas Devlieghere [Wed, 10 Oct 2018 13:27:25 +0000 (13:27 +0000)]
Lift VFS from clang to llvm (NFC)

This patch moves the virtual file system form clang to llvm so it can be
used by more projects.

Concretely the patch:
 - Moves VirtualFileSystem.{h|cpp} from clang/Basic to llvm/Support.
 - Moves the corresponding unit test from clang to llvm.
 - Moves the vfs namespace from clang::vfs to llvm::vfs.
 - Formats the lines affected by this change, mostly this is the result of
   the added llvm namespace.

RFC on the mailing list:
http://lists.llvm.org/pipermail/llvm-dev/2018-October/126657.html

Differential revision: https://reviews.llvm.org/D52783

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344140 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-exegesis] Fix function return generation so it doesn't return register 0
John Brawn [Wed, 10 Oct 2018 13:03:23 +0000 (13:03 +0000)]
[llvm-exegesis] Fix function return generation so it doesn't return register 0

When fillMachineFunction generates a return on targets without a return opcode
(such as AArch64) it should pass an empty set of registers as the return
registers, not 0 which means register number zero.

Differential Revision: https://reviews.llvm.org/D53074

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344139 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[TargetLowering] SimplifyDemandedBits - rename demanded mask args. NFCI.
Simon Pilgrim [Wed, 10 Oct 2018 13:00:49 +0000 (13:00 +0000)]
[TargetLowering] SimplifyDemandedBits - rename demanded mask args. NFCI.

Help stop bugs like rL343935 by making the 'original' DemandedBits arg more obviously not the mask that is actually used.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344138 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-exegesis][NFC] Fix typo
Guillaume Chatelet [Wed, 10 Oct 2018 12:58:40 +0000 (12:58 +0000)]
[llvm-exegesis][NFC] Fix typo

Reviewers: courbet

Subscribers: tschuett, llvm-commits

Differential Revision: https://reviews.llvm.org/D53075

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344137 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[TargetLowering] SimplifyDemandedBits - pull out repeated getOperands. NFCI.
Simon Pilgrim [Wed, 10 Oct 2018 12:32:13 +0000 (12:32 +0000)]
[TargetLowering] SimplifyDemandedBits - pull out repeated getOperands. NFCI.

Part of a minor cleanup to make all the switch statements more consistent prior to improving vector support.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344136 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert "[DebugInfo][Dexter] Unreachable line stepped onto after SimplifyCFG."
Carlos Alberto Enciso [Wed, 10 Oct 2018 12:09:34 +0000 (12:09 +0000)]
Revert "[DebugInfo][Dexter] Unreachable line stepped onto after SimplifyCFG."

This reverts commit r344120.

It was causing buildbot failures.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344135 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[TableGen] fix assert in !cast when used out of definition in a multiclass
Valery Pykhtin [Wed, 10 Oct 2018 10:52:57 +0000 (10:52 +0000)]
[TableGen] fix assert in !cast when used out of definition in a multiclass

Differential Revision: https://reviews.llvm.org/D53068

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344134 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[TargetLowering] Add root node back to work list after successful SimplifyDemandedBit...
Simon Pilgrim [Wed, 10 Oct 2018 10:44:15 +0000 (10:44 +0000)]
[TargetLowering] Add root node back to work list after successful SimplifyDemandedBits/SimplifyDemandedVectorElts

Similar to what already happens in the DAGCombiner wrappers, this patch adds the root nodes back onto the worklist if the DCI wrappers' SimplifyDemandedBits/SimplifyDemandedVectorElts were successful.

Differential Revision: https://reviews.llvm.org/D53026

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344132 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-exegesis] Fix broken build.
Guillaume Chatelet [Wed, 10 Oct 2018 10:09:42 +0000 (10:09 +0000)]
[llvm-exegesis] Fix broken build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344131 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-exegesis][NFC] Simplify code now that Instruction has more semantic
Guillaume Chatelet [Wed, 10 Oct 2018 09:45:17 +0000 (09:45 +0000)]
[llvm-exegesis][NFC] Simplify code now that Instruction has more semantic

Reviewers: courbet

Subscribers: tschuett, llvm-commits

Differential Revision: https://reviews.llvm.org/D53065

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344130 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[SystemZ] Temporarily disable high VFs with integer div/rem.
Jonas Paulsson [Wed, 10 Oct 2018 09:30:29 +0000 (09:30 +0000)]
[SystemZ]  Temporarily disable high VFs with integer div/rem.

Until mischeduler is clever enough to avoid spilling in a vectorized loop
with many (scalar) DLRs it is better to avoid high vectorization factors (8
and above).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344129 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix an ordering bug in the scalarizer.
Neil Henning [Wed, 10 Oct 2018 09:27:45 +0000 (09:27 +0000)]
Fix an ordering bug in the scalarizer.

I've added a new test case that causes the scalarizer to try and use
dead-and-erased values - caused by the basic blocks not being in
domination order within the function. To fix this, instead of iterating
through the blocks in function order, I walk them in reverse post order.

Differential Revision: https://reviews.llvm.org/D52540

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344128 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-exegesis] Remove unused variable, add more semantic to Instruction.
Guillaume Chatelet [Wed, 10 Oct 2018 09:12:36 +0000 (09:12 +0000)]
[llvm-exegesis] Remove unused variable, add more semantic to Instruction.

Reviewers: courbet

Subscribers: tschuett, llvm-commits

Differential Revision: https://reviews.llvm.org/D53062

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344127 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DebugInfo][Dexter] Unreachable line stepped onto after SimplifyCFG.
Carlos Alberto Enciso [Wed, 10 Oct 2018 08:29:55 +0000 (08:29 +0000)]
[DebugInfo][Dexter] Unreachable line stepped onto after SimplifyCFG.

When SimplifyCFG changes the PHI node into a select instruction, the debug line records becomes ambiguous. It causes the debugger to display unreachable source lines.

Differential Revision: https://reviews.llvm.org/D52887

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344120 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Remove FeatureRTM from Skylake processor list
Craig Topper [Wed, 10 Oct 2018 07:43:35 +0000 (07:43 +0000)]
[X86] Remove FeatureRTM from Skylake processor list

Summary:
There are a LOT of Skylakes and later without TSX-NI. Examples:
- SKL: https://ark.intel.com/products/136863/Intel-Core-i3-8121U-Processor-4M-Cache-up-to-3-20-GHz-
- KBL: https://ark.intel.com/products/97540/Intel-Core-i7-7560U-Processor-4M-Cache-up-to-3-80-GHz-
- KBL-R: https://ark.intel.com/products/149091/Intel-Core-i7-8565U-Processor-8M-Cache-up-to-4-60-GHz-
- CNL: https://ark.intel.com/products/136863/Intel-Core-i3-8121U-Processor-4M-Cache-up-to-3_20-GHz

This feature seems to be present only on high-end desktop and server
chips (I can't find any SKX without). This commit leaves it disabled
for all processors, but can be re-enabled for specific builds with
-mrtm.

Patch by Thiago Macieira

Reviewers: erichkeane, craig.topper

Reviewed By: craig.topper

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D53041

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344116 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[SystemZ] Take better care when computing needed vector registers in TTI.
Jonas Paulsson [Wed, 10 Oct 2018 07:36:27 +0000 (07:36 +0000)]
[SystemZ]  Take better care when computing needed vector registers in TTI.

A new function getNumVectorRegs() is better to use for the number of needed
vector registers instead of getNumberOfParts(). This is to make sure that the
number of vector registers (and typically operations) required for a vector
type is accurate.

getNumberOfParts() which was previously used works by splitting the vector
type until it is legal gives incorrect results for types with a non
power of two number of elements (rare).

A new static function getScalarSizeInBits() that also checks for a pointer
type and returns 64U for it since otherwise it gets a value of 0). Used in a
few places where Ty may be pointer.

Review: Ulrich Weigand

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344115 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Analysis] Make LocationSizes carry an 'imprecise' bit
George Burgess IV [Wed, 10 Oct 2018 06:39:40 +0000 (06:39 +0000)]
[Analysis] Make LocationSizes carry an 'imprecise' bit

There are places where we need to merge multiple LocationSizes of
different sizes into one, and get a sensible result.

There are other places where we want to optimize aggressively based on
the value of a LocationSizes (e.g. how can a store of four bytes be to
an area of storage that's only two bytes large?)

This patch makes LocationSize hold an 'imprecise' bit to note whether
the LocationSize can be treated as an upper-bound and lower-bound for
the size of a location, or just an upper-bound.

This concludes the series of patches leading up to this. The most recent
of which is r344108.

Fixes PR36228.

Differential Revision: https://reviews.llvm.org/D44748

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344114 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[NFC] Make a variable const
Max Kazantsev [Wed, 10 Oct 2018 04:19:38 +0000 (04:19 +0000)]
[NFC] Make a variable const

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344113 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[PowerPC][NFC] Add a test case for extract and store patterns
Nemanja Ivanovic [Wed, 10 Oct 2018 04:18:35 +0000 (04:18 +0000)]
[PowerPC][NFC] Add a test case for extract and store patterns

An upcoming patch will change the codegen for these patterns. This test case is
added now so that the patch can show the differences in codegen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344112 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AVR] Fix the 'call.ll' CodeGen test
Dylan McKay [Wed, 10 Oct 2018 03:21:42 +0000 (03:21 +0000)]
[AVR] Fix the 'call.ll' CodeGen test

Commit r343851 changed the format of the generated instructions.

An unnecessary load has been removed. Previously, a value would be moved
from r24 into a temporary register just to be copied into r30 before the
indirect call. Now, codegen immediately loads r24 into r30, saving a
MOVW instruction.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344111 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[PowerPC] Fix the assert of ISD::SIGN_EXTEND_INREG when type is v2i16 and v2i8
QingShan Zhang [Wed, 10 Oct 2018 02:33:48 +0000 (02:33 +0000)]
[PowerPC] Fix the assert of ISD::SIGN_EXTEND_INREG when type is v2i16 and v2i8
For ISD::SIGN_EXTEND_INREG operation of v2i16 and v2i8 types will cause assert because they are registered as custom operation.
So that the type legalization phase will enter the custom hook, which do not handle ISD::SIGN_EXTEND_INREG operation and fall throw into unreachable assert.

Patch By: wuzish (Zixuan Wu)
Differential Revision: https://reviews.llvm.org/D52449

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344109 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Analysis] Make LocationSize pretty-printing more descriptive
George Burgess IV [Wed, 10 Oct 2018 01:35:22 +0000 (01:35 +0000)]
[Analysis] Make LocationSize pretty-printing more descriptive

This is the third patch in a series intended to make
https://reviews.llvm.org/D44748 more easily reviewable. Please see that
patch for more context. The second being r344013.

The intent is to make the output of printing a LocationSize more
precise. The main motivation for this is that we plan to add a bit to
distinguish whether a given LocationSize is an upper-bound or is
precise; making that information available in pretty-printing is nice.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344108 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WebAssembly] Fix fneg lowering
Thomas Lively [Wed, 10 Oct 2018 01:09:09 +0000 (01:09 +0000)]
[WebAssembly] Fix fneg lowering

Summary:
Subtraction from zero and floating point negation do not have the same
semantics, so fix lowering.

Reviewers: aheejin, dschuff

Subscribers: sbc100, jgravelle-google, sunfish, llvm-commits

Differential Revision: https://reviews.llvm.org/D52948

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344107 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WebAssembly] Improve comments for SIMD instruction definitions
Heejin Ahn [Wed, 10 Oct 2018 01:04:02 +0000 (01:04 +0000)]
[WebAssembly] Improve comments for SIMD instruction definitions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344106 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[sancov] Generalize the code to get the previous instruction to multiple architectures
George Karpenkov [Wed, 10 Oct 2018 00:57:24 +0000 (00:57 +0000)]
[sancov] Generalize the code to get the previous instruction to multiple architectures

sancov subtracts one from the address to get the previous instruction,
which makes sense on x86_64, but not on other platforms.
This change ensures that the offset is correct for different platforms.
The logic for computing the offset is copied from sanitizer_common.

Differential Revision: https://reviews.llvm.org/D53039

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344103 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[opt] Change the parameter of OptTable::PrintHelp from Name to Usage and don't append...
Fangrui Song [Wed, 10 Oct 2018 00:15:31 +0000 (00:15 +0000)]
[opt] Change the parameter of OptTable::PrintHelp from Name to Usage and don't append "[options] <inputs>"

Summary:
Before, "[options] <inputs>" is unconditionally appended to the `Name` parameter. It is more flexible to change its semantic to `Usage` and let user customize the usage line.

% llvm-objcopy
...
USAGE: llvm-objcopy <input> [ <output> ] [options] <inputs>

With this patch:

% llvm-objcopy
...
USAGE: llvm-objcopy input [output]

Reviewers: rupprecht, alexshap, jhenderson

Reviewed By: rupprecht

Subscribers: jakehehrlich, mehdi_amini, steven_wu, dexonsmith, llvm-commits

Differential Revision: https://reviews.llvm.org/D51009

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@344097 91177308-0d34-0410-b5e6-96231b3b80d8