platform/upstream/llvm.git
5 years agoAdd missing error checking code intended for r347687
Reid Kleckner [Tue, 27 Nov 2018 19:14:11 +0000 (19:14 +0000)]
Add missing error checking code intended for r347687

llvm-svn: 347690

5 years agoRevert "[RISCV] Mark unit tests as "requires: riscv-registered-target""
Mandeep Singh Grang [Tue, 27 Nov 2018 19:13:52 +0000 (19:13 +0000)]
Revert "[RISCV] Mark unit tests as "requires: riscv-registered-target""

This reverts commit 1a6a0c9ea2716378d55858c11adf5941608531f8.

llvm-svn: 347689

5 years ago[RISCV] Mark unit tests as "requires: riscv-registered-target"
Mandeep Singh Grang [Tue, 27 Nov 2018 19:13:13 +0000 (19:13 +0000)]
[RISCV] Mark unit tests as "requires: riscv-registered-target"

Summary: Some of these tests break if the RISCV backend has not been built.

Reviewers: asb, apazos, sabuasal

Reviewed By: sabuasal

Subscribers: rbar, johnrusso, simoncook, sabuasal, niosHD, kito-cheng, shiva0217, jrtc27, zzheng, edward-jones, rogfer01, MartinMosbeck, brucehoult, the_o, rkruppe, PkmX, jocewei, cfe-commits

Differential Revision: https://reviews.llvm.org/D54816

llvm-svn: 347688

5 years ago[PDB] Add symbol records in bulk
Reid Kleckner [Tue, 27 Nov 2018 19:00:23 +0000 (19:00 +0000)]
[PDB] Add symbol records in bulk

Summary:
This speeds up linking clang.exe/pdb with /DEBUG:GHASH by 31%, from
12.9s to 9.8s.

Symbol records are typically small (16.7 bytes on average), but we
processed them one at a time. CVSymbol is a relatively "large" type. It
wraps an ArrayRef<uint8_t> with a kind an optional 32-bit hash, which we
don't need. Before this change, each DbiModuleDescriptorBuilder would
maintain an array of CVSymbols, and would write them individually with a
BinaryItemStream.

With this change, we now add symbols that happen to appear contiguously
in bulk. For each .debug$S section (roughly one per function), we
allocate two copies, one for relocation, and one for realignment
purposes. For runs of symbols that go in the module stream, which is
most symbols, we now add them as a single ArrayRef<uint8_t>, so the
vector DbiModuleDescriptorBuilder is roughly linear in the number of
.debug$S sections (O(# funcs)) instead of the number of symbol records
(very large).

Some stats on symbol sizes for the curious:
  PDB size: 507M
  sym bytes: 316,508,016
  sym count:  18,954,971
  sym byte avg: 16.7

As future work, we may be able to skip copying symbol records in the
linker for realignment purposes if we make LLVM write them aligned into
the object file. We need to double check that such symbol records are
still compatible with link.exe, but if so, it's definitely worth doing,
since my profile shows we spend 500ms in memcpy in the symbol merging
code. We could potentially cut that in half by saving a copy.
Alternatively, we could apply the relocations *after* we iterate the
symbols. This would require some careful re-engineering of the
relocation processing code, though.

Reviewers: zturner, aganea, ruiu

Subscribers: hiraditya, llvm-commits

Differential Revision: https://reviews.llvm.org/D54554

llvm-svn: 347687

5 years ago[TableGen] Preprocessing support
Vyacheslav Zakharin [Tue, 27 Nov 2018 18:57:43 +0000 (18:57 +0000)]
[TableGen] Preprocessing support

Differential Revision: https://reviews.llvm.org/D54926

llvm-svn: 347686

5 years ago[ASTImporter] Added Import functions for transition to new API.
Balazs Keri [Tue, 27 Nov 2018 18:36:31 +0000 (18:36 +0000)]
[ASTImporter] Added Import functions for transition to new API.

Summary:
These Import_New functions should be used in the ASTImporter,
and the old Import functions should not be used. Later the
Import_New should be renamed to Import again and the old Import
functions must be removed. But this can happen only after LLDB
was updated to use the new Import interface.

This commit is only about introducing the new Import_New
functions. These are not implemented now, only calling the old
Import ones.

Reviewers: shafik, rsmith, a_sidorin, a.sidorin

Reviewed By: a_sidorin

Subscribers: spyffe, a_sidorin, gamesh411, shafik, rsmith, dkrupp, martong, Szelethus, cfe-commits

Differential Revision: https://reviews.llvm.org/D53751

llvm-svn: 347685

5 years ago[X86] Replace an APInt that is guaranteed to be 8-bits with just an 'unsigned'
Craig Topper [Tue, 27 Nov 2018 18:24:56 +0000 (18:24 +0000)]
[X86] Replace an APInt that is guaranteed to be 8-bits with just an 'unsigned'

We're already mixing this APInt with other 'unsigned' variables. This allows us to use regular comparison operators instead of needing to use APInt::ult or APInt::uge. And it removes a later conversion from APInt to unsigned.

I might be adding another combine to this function and this will probably simplify the logic required for that.

llvm-svn: 347684

5 years ago[PartialInliner] Make PHIs free in cost computation.
Florian Hahn [Tue, 27 Nov 2018 18:17:27 +0000 (18:17 +0000)]
[PartialInliner] Make PHIs free in cost computation.

InlineCost also treats them as free and the current implementation
can cause assertion failures if PHI nodes are moved outside the region
from entry BBs to the region.

It also updates the code to use the instructionsWithoutDebug iterator.

Reviewers: davidxl, davide, vsk, graham-yiu-huawei

Reviewed By: davidxl

Differential Revision: https://reviews.llvm.org/D54748

llvm-svn: 347683

5 years ago[X86] Add -march=cascadelake support in clang.
Craig Topper [Tue, 27 Nov 2018 18:05:14 +0000 (18:05 +0000)]
[X86] Add -march=cascadelake support in clang.

This is skylake-avx512 with the addition of avx512vnni ISA.

Patch by Jianping Chen

Differential Revision: https://reviews.llvm.org/D54792

llvm-svn: 347682

5 years ago[X86] Add cascade lake arch in X86 target.
Craig Topper [Tue, 27 Nov 2018 18:05:00 +0000 (18:05 +0000)]
[X86] Add cascade lake arch in X86 target.

This is skylake-avx512 with the addition of avx512vnni ISA.

Patch by Jianping Chen

Differential Revision: https://reviews.llvm.org/D54785

llvm-svn: 347681

5 years agoRemove SaveAndRestore and SmallVectorImpl from lld/Common/LLVM.h. NFC.
Rui Ueyama [Tue, 27 Nov 2018 17:54:15 +0000 (17:54 +0000)]
Remove SaveAndRestore and SmallVectorImpl from lld/Common/LLVM.h. NFC.

llvm-svn: 347680

5 years agoDocumentation: add \file markup as needed.
James Dennett [Tue, 27 Nov 2018 17:53:03 +0000 (17:53 +0000)]
Documentation: add \file markup as needed.

This makes Doxygen correctly associate the doc comment with the current
file rather than adding to the documentation for namespace llvm.

llvm-svn: 347679

5 years agoMake a member function non-member. NFC.
Rui Ueyama [Tue, 27 Nov 2018 17:47:24 +0000 (17:47 +0000)]
Make a member function non-member. NFC.

llvm-svn: 347678

5 years agoFix linker option for -fprofile-arcs -ftest-coverage
Marco Castelluccio [Tue, 27 Nov 2018 17:31:08 +0000 (17:31 +0000)]
Fix linker option for -fprofile-arcs -ftest-coverage

Summary:
Linux toolchain accidentally added "-u__llvm_runtime_variable" when "-fprofile-arcs -ftest-coverage", this is not added when "--coverage" option is used.
Using "-u__llvm_runtime_variable" generates an empty default.profraw file while an application built with  "-fprofile-arcs -ftest-coverage" is running.

Reviewers: calixte, marco-c, sylvestre.ledru

Reviewed By: marco-c

Subscribers: vsk, cfe-commits

Differential Revision: https://reviews.llvm.org/D54195

llvm-svn: 347677

5 years agoRevert "[clang] - Simplify tools::SplitDebugName."
Jonas Toth [Tue, 27 Nov 2018 17:28:38 +0000 (17:28 +0000)]
Revert "[clang] - Simplify tools::SplitDebugName."

This reverts commit r347035 as it introduced assertion failures under
certain conditions. More information can be found here:
https://reviews.llvm.org/rL347035

llvm-svn: 347676

5 years ago[clangd] textDocument/SymbolInfo extension
Jan Korous [Tue, 27 Nov 2018 16:40:46 +0000 (16:40 +0000)]
[clangd] textDocument/SymbolInfo extension

New method returning symbol info for given source position.

Differential Revision: https://reviews.llvm.org/D54799

rdar://problem/46050281

llvm-svn: 347675

5 years ago[clangd][NFC] Move SymbolID to a separate file
Jan Korous [Tue, 27 Nov 2018 16:40:34 +0000 (16:40 +0000)]
[clangd][NFC] Move SymbolID to a separate file

Prerequisity for textDocument/SymbolInfo

Differential Revision: https://reviews.llvm.org/D54799

llvm-svn: 347674

5 years agoCatch up with EvaluateAsInt() clang API change.
Davide Italiano [Tue, 27 Nov 2018 16:40:21 +0000 (16:40 +0000)]
Catch up with EvaluateAsInt() clang API change.

llvm-svn: 347673

5 years agoImplement P1085R2 - Should Span be Regular?. This consists entirely of deletions
Marshall Clow [Tue, 27 Nov 2018 16:35:41 +0000 (16:35 +0000)]
Implement P1085R2 - Should Span be Regular?. This consists entirely of deletions

llvm-svn: 347672

5 years ago[clang-tidy] Ignore bool -> single bit bitfield conversion in readability-implicit...
Malcolm Parsons [Tue, 27 Nov 2018 16:23:39 +0000 (16:23 +0000)]
[clang-tidy] Ignore bool -> single bit bitfield conversion in readability-implicit-bool-conversion

Summary: There is no ambiguity / information loss in this conversion

Reviewers: alexfh, aaron.ballman, hokein

Reviewed By: alexfh

Subscribers: xazax.hun, cfe-commits

Differential Revision: https://reviews.llvm.org/D54941

llvm-svn: 347671

5 years ago[Demangle] remove itaniumFindTypesInMangledName
Pavel Labath [Tue, 27 Nov 2018 16:11:24 +0000 (16:11 +0000)]
[Demangle] remove itaniumFindTypesInMangledName

Summary:
This (very specialized) function was added to enable an LLDB use case.
Now that a more generic interface (overriding of parser functions -
D52992)  is available, and LLDB has been converted to use that (D54074),
the function is unused and can be removed.

Reviewers: erik.pilkington, sgraenitz, rsmith

Subscribers: mgorny, hiraditya, christof, libcxx-commits, llvm-commits

Differential Revision: https://reviews.llvm.org/D54893

llvm-svn: 347670

5 years ago[clangd] Put direct headers into srcs section.
Kadir Cetinkaya [Tue, 27 Nov 2018 16:08:53 +0000 (16:08 +0000)]
[clangd] Put direct headers into srcs section.

Summary:
Currently, there's no way of knowing about header files
using compilation database, since it doesn't contain header files as entries.

Using this information, restoring from cache using compile commands becomes
possible instead of doing directory traversal. Also, we can issue indexing
actions for out-of-date headers even if source files depending on them haven't
changed.

Reviewers: sammccall

Subscribers: ilya-biryukov, ioeric, MaskRay, jkorous, arphaman, cfe-commits

Differential Revision: https://reviews.llvm.org/D54817

llvm-svn: 347669

5 years agotravis: Add cmake build
Jan Vesely [Tue, 27 Nov 2018 16:07:21 +0000 (16:07 +0000)]
travis: Add cmake build

Reviewer: Aaron Watry

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 347668

5 years agoAdd cmake build system
Jan Vesely [Tue, 27 Nov 2018 16:07:19 +0000 (16:07 +0000)]
Add cmake build system

Add cmake support for CLC and ll asm language,
the latter includes clang preprocessing stage.
Add ctests to check for external function calls.

v2: fix typos, style

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Acked-by: Aaron Watry <awatry@gmail.com>
Tested-by: Aaron Watry <awatry@gmail.com>
Acked-by: Vedran Miletić <vedran@miletic.net>
llvm-svn: 347667

5 years agor600: Remove empty OVERRIDES file
Jan Vesely [Tue, 27 Nov 2018 16:01:16 +0000 (16:01 +0000)]
r600: Remove empty OVERRIDES file

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewer: Aaron Watry
llvm-svn: 347666

5 years agoamdgcn: Consolidate atomic minmax helpers
Jan Vesely [Tue, 27 Nov 2018 16:01:13 +0000 (16:01 +0000)]
amdgcn: Consolidate atomic minmax helpers

Removes most overrides
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewer: Aaron Watry

llvm-svn: 347665

5 years agoconfigure: Add target specific asm rule.
Jan Vesely [Tue, 27 Nov 2018 16:01:10 +0000 (16:01 +0000)]
configure: Add target specific asm rule.

Run the file through target specific preprocessing stage.
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewer: Aaron Watry

llvm-svn: 347664

5 years agoconfigure: provide llvm_as helper variable
Jan Vesely [Tue, 27 Nov 2018 16:01:01 +0000 (16:01 +0000)]
configure: provide llvm_as helper variable

Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewer: Aaron Watry
llvm-svn: 347663

5 years ago[llvm-mca] pass -dispatch-stats flag to a couple of tests. NFC
Andrea Di Biagio [Tue, 27 Nov 2018 15:56:00 +0000 (15:56 +0000)]
[llvm-mca] pass -dispatch-stats flag to a couple of tests. NFC

This change is in preparation for a patch that fixes PR36666.

llvm-mca currently doesn't know if a buffered processor resource describes a
load or store queue. So, any dynamic dispatch stall caused by the lack of
load/store queue entries is normally reported as a generic SCHEDULER stall. See for
example the -dispatch-stats output from the two tests modified by this patch.

In future, processor models will be able to tag processor resources that are
used to describe load/store queues. That information would then be used by
llvm-mca to correctly classify dynamic dispatch stalls caused by the lack of
tokens in the LS.

llvm-svn: 347662

5 years ago[x86] regenerate checks; NFC
Sanjay Patel [Tue, 27 Nov 2018 15:52:17 +0000 (15:52 +0000)]
[x86] regenerate checks; NFC

llvm-svn: 347661

5 years agoMove time cast to SymbolFileDWARFDebugMap
Jonas Devlieghere [Tue, 27 Nov 2018 15:25:58 +0000 (15:25 +0000)]
Move time cast to SymbolFileDWARFDebugMap

When trying to fix the bots we expected that the cast would be needed in
different places. Ultimately it turned out only the
SymbolFileDWARFDebugMap was affected so, as Pavel correctly notes, it
makes more sense to do the cast just there instead of in teh FS.

llvm-svn: 347660

5 years ago[AMDGPU] Disable DAG combine at -O0
Stanislav Mekhanoshin [Tue, 27 Nov 2018 15:13:37 +0000 (15:13 +0000)]
[AMDGPU] Disable DAG combine at -O0

Differential Revision: https://reviews.llvm.org/D54358

llvm-svn: 347659

5 years agoDerive builtin return type from its definition
Marco Antognini [Tue, 27 Nov 2018 14:54:58 +0000 (14:54 +0000)]
Derive builtin return type from its definition

Summary:
Prior to this patch, OpenCL code such as the following would attempt to create
a BranchInst with a non-bool argument:

    if (enqueue_kernel(get_default_queue(), 0, nd, ^(void){})) /* ... */

This patch is a follow up on a similar issue with pipe builtin
operations. See commit r280800 and https://bugs.llvm.org/show_bug.cgi?id=30219.

This change, while being conservative on non-builtin functions,
should set the type of expressions invoking builtins to the
proper type, instead of defaulting to `bool` and requiring
manual overrides in Sema::CheckBuiltinFunctionCall.

In addition to tests for enqueue_kernel, the tests are extended to
check other OpenCL builtins.

Reviewers: Anastasia, spatel, rsmith

Reviewed By: Anastasia

Subscribers: kristina, cfe-commits, svenvh

Differential Revision: https://reviews.llvm.org/D52879

llvm-svn: 347658

5 years agoRevert r347419 "Update call to EvaluateAsInt() to the new syntax."
Hans Wennborg [Tue, 27 Nov 2018 14:15:37 +0000 (14:15 +0000)]
Revert r347419 "Update call to EvaluateAsInt() to the new syntax."

It's pre-requisite was reverted in r347656.

llvm-svn: 347657

5 years agoRevert r347417 "Re-Reinstate 347294 with a fix for the failures."
Hans Wennborg [Tue, 27 Nov 2018 14:01:40 +0000 (14:01 +0000)]
Revert r347417 "Re-Reinstate 347294 with a fix for the failures."

This caused a miscompile in Chrome (see crbug.com/908372) that's
illustrated by this small reduction:

  static bool f(int *a, int *b) {
    return !__builtin_constant_p(b - a) || (!(b - a));
  }

  int arr[] = {1,2,3};

  bool g() {
    return f(arr, arr + 3);
  }

  $ clang -O2 -S -emit-llvm a.cc -o -

g() should return true, but after r347417 it became false for some reason.

This also reverts the follow-up commits.

r347417:
> Re-Reinstate 347294 with a fix for the failures.
>
> Don't try to emit a scalar expression for a non-scalar argument to
> __builtin_constant_p().
>
> Third time's a charm!

r347446:
> The result of is.constant() is unsigned.

r347480:
> A __builtin_constant_p() returns 0 with a function type.

r347512:
> isEvaluatable() implies a constant context.
>
> Assume that we're in a constant context if we're asking if the expression can
> be compiled into a constant initializer. This fixes the issue where a
> __builtin_constant_p() in a compound literal was diagnosed as not being
> constant, even though it's always possible to convert the builtin into a
> constant.

r347531:
> A "constexpr" is evaluated in a constant context. Make sure this is reflected
> if a __builtin_constant_p() is a part of a constexpr.

llvm-svn: 347656

5 years ago[clangd] Prevent thread starvation in tests on loaded systems.
Sam McCall [Tue, 27 Nov 2018 12:09:13 +0000 (12:09 +0000)]
[clangd] Prevent thread starvation in tests on loaded systems.

Summary:
Background index deliberately runs low-priority, but for tests this may stop
them making progress.

Reviewers: kadircet

Subscribers: ilya-biryukov, ioeric, MaskRay, jkorous, arphaman, jfb, cfe-commits

Differential Revision: https://reviews.llvm.org/D54938

llvm-svn: 347655

5 years ago[libclang] Fix clang_Cursor_getNumArguments and clang_Cursor_getArgument for CXXConst...
Ivan Donchevskii [Tue, 27 Nov 2018 12:02:39 +0000 (12:02 +0000)]
[libclang] Fix clang_Cursor_getNumArguments and clang_Cursor_getArgument for CXXConstructExpr

Constructors have the same methods for arguments as call expressions.
Let's provide a way to get their arguments the same way.

Differential Revision: https://reviews.llvm.org/D54934

llvm-svn: 347654

5 years agoInstCombine: add comment explaining malloc deletion. NFC.
Tim Northover [Tue, 27 Nov 2018 11:08:14 +0000 (11:08 +0000)]
InstCombine: add comment explaining malloc deletion. NFC.

I tried to change this, not quite realising the logic behind what we
were doing. Hopefully this comment will help the next person to come
along.

llvm-svn: 347653

5 years ago[clang-tidy] Avoid inconsistent notes in readability-container-size-empty
Alexander Kornienko [Tue, 27 Nov 2018 10:53:44 +0000 (10:53 +0000)]
[clang-tidy] Avoid inconsistent notes in readability-container-size-empty

When a warning is issued in a template instantiation, the check would previously
use template arguments in a note, which would result in inconsistent or
duplicate warnings (depending on how deduplication was done). This patch removes
template arguments from the note.

llvm-svn: 347652

5 years ago[clang-tidy] Minor fixes in a test
Alexander Kornienko [Tue, 27 Nov 2018 10:53:38 +0000 (10:53 +0000)]
[clang-tidy] Minor fixes in a test

Use CHECK-FIXES where it was intended instead of CHECK-MESSAGES. Fixed compiler
warnings to pacify YouCompleteMe.

llvm-svn: 347651

5 years ago[ELF] - Fix R_AARCH64_ADR_GOT_PAGE, R_AARCH64_LD64_GOT_LO12 handling against IFUNC...
George Rimar [Tue, 27 Nov 2018 10:30:46 +0000 (10:30 +0000)]
[ELF] - Fix R_AARCH64_ADR_GOT_PAGE, R_AARCH64_LD64_GOT_LO12 handling against IFUNC symbols.

This is https://bugs.llvm.org/show_bug.cgi?id=38074.

The issue is that when calling a function, LLD generates a
.got entry that points to the IFUNC resolver function when
instead, it should use the PLT entries properly for
handling the IFUNC.

So we should create a got entry that points to PLT entry,
which itself loads the value from
.got.plt, relocated with R_*_IRELATIVE to make things work.
Patch do that.

Differential revision: https://reviews.llvm.org/D54314

llvm-svn: 347650

5 years ago[AArch64] Cortex-a53-843419 erratum should not apply to relaxed TLS.
Peter Smith [Tue, 27 Nov 2018 10:17:35 +0000 (10:17 +0000)]
[AArch64] Cortex-a53-843419 erratum should not apply to relaxed TLS.

The changes to the instructions performed by TLS relaxation and the errata
patching are performed with relocations. As these are applied so late the
errata scanning won't see the changes in the section data made by the TLS
relaxation. This can lead to a TLS relaxed sequence being patched when it
doesn't need to be.

The fix checks to see if there is a R_RELAX_TLS_IE_TO_LE instruction at the
same address as the ADRP as this indicates the presence of a relaxation
of a sequence that might get recognised as a patch.

Differential Revision: https://reviews.llvm.org/D54854

llvm-svn: 347649

5 years ago[ASTImporter] Typedef import brings in the complete type
Gabor Marton [Tue, 27 Nov 2018 09:51:36 +0000 (09:51 +0000)]
[ASTImporter] Typedef import brings in the complete type

Summary:
When we already have an incomplete underlying type of a typedef in the
"To" context, and the "From" context has the same typedef, but the
underlying type is complete, then the imported type should be complete.

Fixes an assertion in CTU analysis of Xerces:
Assertion `DD && "queried property of class with no definition"' failed.
This assert is happening in the analyzer engine, because that attempts
to query an underlying type of a typedef, which happens to be
incomplete.

Reviewers: a_sidorin, a.sidorin

Subscribers: rnkovacs, dkrupp, Szelethus, cfe-commits

Differential Revision: https://reviews.llvm.org/D53693

llvm-svn: 347648

5 years ago[ELF] - Implement -z nodefaultlib
George Rimar [Tue, 27 Nov 2018 09:48:17 +0000 (09:48 +0000)]
[ELF] - Implement -z nodefaultlib

This is https://bugs.llvm.org//show_bug.cgi?id=38978

Spec says that:
"Objects may be built with the -z nodefaultlib option to
suppress any search of the default locations at runtime.
Use of this option implies that all the dependencies of an
object can be located using its runpaths.
Without this option, which is the most common case, no
matter how you augment the runtime linker's library
search path, its last element is always /usr/lib for 32-bit
objects and /usr/lib/64 for 64-bit objects."

The patch implements this option.

Differential revision: https://reviews.llvm.org/D54577

llvm-svn: 347647

5 years ago[CMake] Add a missing case of TO_CMAKE_PATH
Martin Storsjo [Tue, 27 Nov 2018 09:23:15 +0000 (09:23 +0000)]
[CMake] Add a missing case of TO_CMAKE_PATH

This fixes building sanitizers for mingw natively.

llvm-svn: 347646

5 years ago[COFF] Generate a codeview build id signature for MinGW even when not creating a PDB
Martin Storsjo [Tue, 27 Nov 2018 09:20:55 +0000 (09:20 +0000)]
[COFF] Generate a codeview build id signature for MinGW even when not creating a PDB

GNU ld, which doesn't generate PDBs, can optionally generate a
build id by passing the --build-id option. LLD's MinGW frontend knows
about this option but ignores it, as I had falsely assumed that LLD
already generated build IDs even in those cases.

If debug info is requested and no PDB path is set, generate a
build id signature as a hash of the binary itself. This allows
associating a binary to a minidump, even if debug info isn't
written in PDB form by the linker.

Differential Revision: https://reviews.llvm.org/D54828

llvm-svn: 347645

5 years agoAdd missing REQUIRES: asserts
Max Kazantsev [Tue, 27 Nov 2018 07:51:18 +0000 (07:51 +0000)]
Add missing REQUIRES: asserts

llvm-svn: 347644

5 years ago[X86] Add test cases for vector shifts of v2i32/v2i16/v4i16/v2i8/v4i8/v8i8 with promo...
Craig Topper [Tue, 27 Nov 2018 07:20:19 +0000 (07:20 +0000)]
[X86] Add test cases for vector shifts of v2i32/v2i16/v4i16/v2i8/v4i8/v8i8 with promotion legalization and widening legalization. NFC

llvm-svn: 347643

5 years ago[X86] Use getUnpackl/getUnpackh instead of directly creating UNPCKL/UNPCKH nodes.
Craig Topper [Tue, 27 Nov 2018 06:24:56 +0000 (06:24 +0000)]
[X86] Use getUnpackl/getUnpackh instead of directly creating UNPCKL/UNPCKH nodes.

llvm-svn: 347642

5 years ago[LoopSimplifyCFG] Turn on term folding after underlying bug fixed
Max Kazantsev [Tue, 27 Nov 2018 06:19:42 +0000 (06:19 +0000)]
[LoopSimplifyCFG] Turn on term folding after underlying bug fixed

llvm-svn: 347641

5 years ago[LoopSimplifyCFG] Fix corner case with duplicating successors
Max Kazantsev [Tue, 27 Nov 2018 06:17:21 +0000 (06:17 +0000)]
[LoopSimplifyCFG] Fix corner case with duplicating successors

It fixes a bug that doesn't update Phi inputs of the only live successor that
is in the list of block's successors more than once.

Thanks @uabelho for finding this.

Differential Revision: https://reviews.llvm.org/D54849
Reviewed By: anna

llvm-svn: 347640

5 years ago[gn build] Merge r347530 to gn.
Nico Weber [Tue, 27 Nov 2018 06:04:49 +0000 (06:04 +0000)]
[gn build] Merge r347530 to gn.

llvm-svn: 347639

5 years agoMove a file I forgot to move in r347636.
Nico Weber [Tue, 27 Nov 2018 05:49:08 +0000 (05:49 +0000)]
Move a file I forgot to move in r347636.

llvm-svn: 347638

5 years agoAdd support for the Dylan language to ClangASTContext
Bruce Mitchener [Tue, 27 Nov 2018 05:37:27 +0000 (05:37 +0000)]
Add support for the Dylan language to ClangASTContext

Summary:
This change adds eLanguageTypeDylan to the set of languages supported
by ClangASTContext. Debug info generated by the Open Dylan compiler's
LLVM back-end was designed to be compatible with C debug info.

Patch by Peter Housel.

Reviewers: clayborg

Reviewed By: clayborg

Subscribers: brucem, lldb-commits, aprantl

Differential Revision: https://reviews.llvm.org/D54886

llvm-svn: 347637

5 years ago[gn build] Create abi-breaking.h, config.h, llvm-config.h, and add a build file for...
Nico Weber [Tue, 27 Nov 2018 05:19:17 +0000 (05:19 +0000)]
[gn build] Create abi-breaking.h, config.h, llvm-config.h, and add a build file for llvm/lib/Support.

The comments at the top of
llvm/utils/gn/secondary/llvm/include/llvm/Config/BUILD.gn and
llvm/utils/gn/build/write_cmake_config.py should explain the main bits
happening in this patch. The main parts here are that these headers are
generated at build time, not gn time, and that currently they don't do any
actual feature checks but just hardcode most things based on the current OS,
which seems to work well enough. If this stops being enough, the feature checks
should each be their own action writing the result to somewhere, and the config
write step should depend on those checks (so that they can run in parallel and
as part of the build) -- utils/llvm/gn/README.rst already has some more words
on that in "Philosophy".

(write_cmake_config.py is also going to be used to write clang's
clang/include/clang/Config/config.h)

This also adds a few files for linking to system libraries in a consistent way
if needed in llvm/utils/gn/build/libs (and moves pthread to that model).0

I'm also adding llvm/utils/gn/secondary/llvm/lib/Target/targets.gni in this
patch because $native_arch is needed for writing llvm-config.h -- the rest of
it will be used later, when the build files for llvm/lib/Target get added. That
file describes how to select which archs to build.

As a demo, also add a build file for llvm-undname and make it the default build
target (it depends on everything that can currently be built).

Differential Revision: https://reviews.llvm.org/D54678

llvm-svn: 347636

5 years ago[clangd] NFC: Prefer `isa<>` to `dyn_cast<>` to do the checking.
Henry Wong [Tue, 27 Nov 2018 04:27:00 +0000 (04:27 +0000)]
[clangd] NFC: Prefer `isa<>` to `dyn_cast<>` to do the checking.

Summary: Prefer `isa<>` to `dyn_cast<>` when there only need a checking.

Reviewers: ilya-biryukov, MaskRay

Reviewed By: ilya-biryukov, MaskRay

Subscribers: ioeric, MaskRay, jkorous, arphaman, kadircet, cfe-commits, MTC

Differential Revision: https://reviews.llvm.org/D54878

llvm-svn: 347635

5 years ago[docs] UBSan and ASan are supported on Windows
Reid Kleckner [Tue, 27 Nov 2018 03:55:15 +0000 (03:55 +0000)]
[docs] UBSan and ASan are supported on Windows

Also fix a bullet list.

Fixes PR39775

llvm-svn: 347633

5 years ago[X86] Prevent DAG combine from folding a bitcast from vXi1 to iX with a store on...
Craig Topper [Tue, 27 Nov 2018 02:57:27 +0000 (02:57 +0000)]
[X86] Prevent DAG combine from folding a bitcast from vXi1 to iX with a store on pre-AVX512 targets.

If we fold the bitcast into the store we'll end up creating a truncating store to vXi1 that will get scalarized. Instead allow the bitcast to be turned into a movmsk.

We probably need to do something if the store itself is a vXi1 type, but I'll leave that til a testcase appears.

llvm-svn: 347632

5 years ago[X86] Add a bunch of test cases for storing a scalar bitcasted from a vXi1 type.
Craig Topper [Tue, 27 Nov 2018 02:57:23 +0000 (02:57 +0000)]
[X86] Add a bunch of test cases for storing a scalar bitcasted from a vXi1 type.

Currently a store combine will absorb the bitcast before our combine that turns bitcasts into movmsk gets a chance to run. This results in a store being created with a vXi1 type. Type legalization then promotes the input type and makes this a truncating store. Then we badly scalarize this store.

Currently we avoid this on v8i1->i8 bitcasts due to an incompletely qualified(per the original intention) check in isLoadBitCastBeneficial. An easy fix is to disable this for all vXi1->iX bitcasts on pre-avx512 targets. We'll still generate terrible code if the IR explicitly contains a store of vXi1 without a bitcast. We could probably solve that by just turning all stores of vXi1 into (store (iX (bitcast))) as an early DAG combine.

llvm-svn: 347631

5 years agoRevert r347627 "[MS] Push fewer DeclContexts for delayed template parsing"
Reid Kleckner [Tue, 27 Nov 2018 02:54:17 +0000 (02:54 +0000)]
Revert r347627 "[MS] Push fewer DeclContexts for delayed template parsing"

It broke the Windows self-host:
http://lab.llvm.org:8011/builders/clang-x64-windows-msvc/builds/1799/steps/stage%202%20build/logs/stdio

I can build
lib/CodeGen/CMakeFiles/LLVMCodeGen.dir/MachinePostDominators.cpp.obj to
repro.

llvm-svn: 347630

5 years ago[analyzer][PlistMacroExpansion] Part 3.: Macro arguments are expanded
Kristof Umann [Tue, 27 Nov 2018 02:28:23 +0000 (02:28 +0000)]
[analyzer][PlistMacroExpansion] Part 3.: Macro arguments are expanded

This part focuses on expanding macro arguments.

Differential Revision: https://reviews.llvm.org/D52795

llvm-svn: 347629

5 years agoRevert "[clang][slh] add attribute for speculative load hardening"
Zola Bridges [Tue, 27 Nov 2018 02:22:00 +0000 (02:22 +0000)]
Revert "[clang][slh] add attribute for speculative load hardening"

until I figure out why the build is failing or timing out

***************************

Summary:
The prior diff had to be reverted because there were two tests
that failed. I updated the two tests in this diff

clang/test/Misc/pragma-attribute-supported-attributes-list.test
clang/test/SemaCXX/attr-speculative-load-hardening.cpp

LLVM IR already has an attribute for speculative_load_hardening. Before
this commit, when a user passed the -mspeculative-load-hardening flag to
Clang, every function would have this attribute added to it. This Clang
attribute will allow users to opt into SLH on a function by function
basis.

This can be applied to functions and Objective C methods.

Reviewers: chandlerc, echristo, kristof.beyls, aaron.ballman

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D54915

This reverts commit a5b3c232d1e3613f23efbc3960f8e23ea70f2a79.
(r347617)

llvm-svn: 347628

5 years ago[MS] Push fewer DeclContexts for delayed template parsing
Reid Kleckner [Tue, 27 Nov 2018 02:21:51 +0000 (02:21 +0000)]
[MS] Push fewer DeclContexts for delayed template parsing

Only push the outermost record as a DeclContext when parsing a function
body. See the comments in Sema::getContainingDC about the way the parser
pushes contexts. This is intended to match the behavior the parser
normally displays where it parses all method bodies from all nested
classes at the end of the outermost class, when all nested classes are
complete.

Fixes PR38460.

llvm-svn: 347627

5 years ago[stack-safety] Update comment
Vitaly Buka [Tue, 27 Nov 2018 01:56:44 +0000 (01:56 +0000)]
[stack-safety] Update comment

llvm-svn: 347626

5 years ago[stack-safety] Fix and uncomment assert
Vitaly Buka [Tue, 27 Nov 2018 01:56:35 +0000 (01:56 +0000)]
[stack-safety] Fix and uncomment assert

llvm-svn: 347625

5 years ago[stack-safety] Fix build on gcc 5.4
Vitaly Buka [Tue, 27 Nov 2018 01:56:26 +0000 (01:56 +0000)]
[stack-safety] Fix build on gcc 5.4

llvm-svn: 347624

5 years ago[COFF] Add exported functions to gfids table for /guard:cf
Reid Kleckner [Tue, 27 Nov 2018 01:50:17 +0000 (01:50 +0000)]
[COFF] Add exported functions to gfids table for /guard:cf

Summary:
MSVC does this, and we should to.

The .gfids table is a table of RVAs, so it's impossible for a DLL to
indicate that an imported symbol is address taken. Therefore, exports
appear to be listed as address taken by the DLL that exports them.

This fixes an issue that Firefox ran into here:
https://bugzilla.mozilla.org/show_bug.cgi?id=1485016#c12

In Firefox, the export directive came from a .def file, but we need to
do this for any kind of export.

Reviewers: dmajor, hans, amccarth, alex

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D54723

llvm-svn: 347623

5 years agoFix filtering of sanitizer_common unittest architectures on Darwin.
Kuba Mracek [Tue, 27 Nov 2018 01:37:08 +0000 (01:37 +0000)]
Fix filtering of sanitizer_common unittest architectures on Darwin.

llvm-svn: 347622

5 years ago[WebAssembly] Remove `using` statements from header files. NFC.
Sam Clegg [Tue, 27 Nov 2018 01:08:16 +0000 (01:08 +0000)]
[WebAssembly] Remove `using` statements from header files. NFC.

Differential Revision: https://reviews.llvm.org/D54758

llvm-svn: 347621

5 years ago[Common] Threads: use function_ref instead of std::function
Fangrui Song [Tue, 27 Nov 2018 00:26:50 +0000 (00:26 +0000)]
[Common] Threads: use function_ref instead of std::function

The benefits are:

a) Performance (very minor): it removes a heap allocation of the std::function constructor (template<class F> function(F f))
b) Clarity: it suggests that the callable's lifetime should end after the callee returns. Such callback is widely used in llvm. lld also uses it a lot.

Reviewers: ruiu, rnk, pcc

Reviewed By: ruiu

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D54813

llvm-svn: 347620

5 years agoRevert "[CMake] Streamline code signing for debugserver and pass entitlements to...
Davide Italiano [Tue, 27 Nov 2018 00:25:49 +0000 (00:25 +0000)]
Revert "[CMake] Streamline code signing for debugserver and pass entitlements to extended llvm_codesign"

It breaks the lldb cmake bots.

llvm-svn: 347619

5 years ago[InstCombine] add tests for rotate/bswap equality; NFC
Sanjay Patel [Tue, 27 Nov 2018 00:08:21 +0000 (00:08 +0000)]
[InstCombine] add tests for rotate/bswap equality; NFC

llvm-svn: 347618

5 years ago[clang][slh] add attribute for speculative load hardening
Zola Bridges [Tue, 27 Nov 2018 00:03:44 +0000 (00:03 +0000)]
[clang][slh] add attribute for speculative load hardening

Summary:
The prior diff had to be reverted because there were two tests
that failed. I updated the two tests in this diff

clang/test/Misc/pragma-attribute-supported-attributes-list.test
clang/test/SemaCXX/attr-speculative-load-hardening.cpp

----- Summary from Previous Diff (Still Accurate) -----

LLVM IR already has an attribute for speculative_load_hardening. Before
this commit, when a user passed the -mspeculative-load-hardening flag to
Clang, every function would have this attribute added to it. This Clang
attribute will allow users to opt into SLH on a function by function basis.

This can be applied to functions and Objective C methods.

Reviewers: chandlerc, echristo, kristof.beyls, aaron.ballman

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D54915

llvm-svn: 347617

5 years agoFix debug build break
JF Bastien [Mon, 26 Nov 2018 23:48:47 +0000 (23:48 +0000)]
Fix debug build break

Comment out an assertion from D54543 which failed with error: no member named 'Range' in '(anonymous namespace)::PassAsArgInfo'.

llvm-svn: 347616

5 years ago[FileSystem] Ignore nanoseconds when comparing oso_mod_time
Jonas Devlieghere [Mon, 26 Nov 2018 23:40:52 +0000 (23:40 +0000)]
[FileSystem] Ignore nanoseconds when comparing oso_mod_time

After a recent change in LLVM the TimePoint encoding become more
precise, exceeding the precision of the TimePoint obtained from the
DebugMap. This patch adds a flag to the GetModificationTime helper in
the FileSystem to return the modification time with less precision.

Thanks to Davide for bisecting this failure on the LLDB bots.

llvm-svn: 347615

5 years agoNotify the linker when a TU compiled with split-stack has a function without a prologue.
Sterling Augustine [Mon, 26 Nov 2018 23:26:31 +0000 (23:26 +0000)]
Notify the linker when a TU compiled with split-stack has a function without a prologue.

More context here: https://go-review.googlesource.com/c/go/+/148819/

llvm-svn: 347614

5 years agoRemove trailing empty line
Vitaly Buka [Mon, 26 Nov 2018 23:17:52 +0000 (23:17 +0000)]
Remove trailing empty line

llvm-svn: 347613

5 years ago[stack-safety] Analysis documentation
Vitaly Buka [Mon, 26 Nov 2018 23:16:07 +0000 (23:16 +0000)]
[stack-safety] Analysis documentation

Summary:
Basic documentation of the Stack Safety Analysis.
It will be improved during review and upstream of an implementation.

Reviewers: kcc, eugenis, vlad.tsyrklevich, glider

Reviewed By: vlad.tsyrklevich

Subscribers: arphaman, llvm-commits

Differential Revision: https://reviews.llvm.org/D53336

llvm-svn: 347612

5 years ago[stack-safety] Inter-Procedural Analysis implementation
Vitaly Buka [Mon, 26 Nov 2018 23:05:58 +0000 (23:05 +0000)]
[stack-safety] Inter-Procedural Analysis implementation

Summary:
IPA is implemented as module pass which produce map from Function or Alias to
StackSafetyInfo for a single function.

From prototype by Evgenii Stepanov and Vlad Tsyrklevich.

Reviewers: eugenis, vlad.tsyrklevich, pcc, glider

Subscribers: hiraditya, mgrang, llvm-commits

Differential Revision: https://reviews.llvm.org/D54543

llvm-svn: 347611

5 years ago[stack-safety] Empty local passes for Stack Safety Global Analysis
Vitaly Buka [Mon, 26 Nov 2018 23:05:48 +0000 (23:05 +0000)]
[stack-safety] Empty local passes for Stack Safety Global Analysis

Reviewers: eugenis, vlad.tsyrklevich

Subscribers: hiraditya, llvm-commits

Differential Revision: https://reviews.llvm.org/D54541

llvm-svn: 347610

5 years agoAArch64ISelLowering: Remove a return-of-assignment to allow NRVO
David Blaikie [Mon, 26 Nov 2018 22:57:18 +0000 (22:57 +0000)]
AArch64ISelLowering: Remove a return-of-assignment to allow NRVO

Patch by Arthur O'Dwyer!

llvm-svn: 347609

5 years agoRemove duplicate _LIBCPP_INLINE_VISIBILITY attributes.
Eric Fiselier [Mon, 26 Nov 2018 22:51:35 +0000 (22:51 +0000)]
Remove duplicate _LIBCPP_INLINE_VISIBILITY attributes.

This attribute should appear only on the first declaration. This
patch cleans up <string> by removing the attribute on redeclarations.

llvm-svn: 347608

5 years agoAdd new passes to X86 pipeline tests
Mircea Trofin [Mon, 26 Nov 2018 22:49:17 +0000 (22:49 +0000)]
Add new passes to X86 pipeline tests

Summary: Fixes test failures introduced by rL347596.

Reviewers: davidxl

Reviewed By: davidxl

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D54916

llvm-svn: 347607

5 years ago[X86] Add dependency from X86 to ProfileData after rL347596
Fangrui Song [Mon, 26 Nov 2018 22:16:19 +0000 (22:16 +0000)]
[X86] Add dependency from X86 to ProfileData after rL347596

llvm-svn: 347606

5 years ago[ICP] Remove incompatible attributes at indirect-call promoted callsites.
Xin Tong [Mon, 26 Nov 2018 22:03:52 +0000 (22:03 +0000)]
[ICP] Remove incompatible attributes at indirect-call promoted callsites.

Summary:
Removing ncompatible attributes at indirect-call promoted callsites, not removing it results in
at least a IR verification error.

Reviewers: davidxl, xur, mssimpso

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D54913

llvm-svn: 347605

5 years ago[InstCombine] add helper function to reduce code duplication; NFC
Sanjay Patel [Mon, 26 Nov 2018 22:00:41 +0000 (22:00 +0000)]
[InstCombine] add helper function to reduce code duplication; NFC

llvm-svn: 347604

5 years ago[stack-safety] Local analysis implementation
Vitaly Buka [Mon, 26 Nov 2018 21:57:59 +0000 (21:57 +0000)]
[stack-safety] Local analysis implementation

Summary:
Analysis produces StackSafetyInfo which contains information with how allocas
and parameters were used in functions.

From prototype by Evgenii Stepanov and  Vlad Tsyrklevich.

Reviewers: eugenis, vlad.tsyrklevich, pcc, glider

Subscribers: hiraditya, llvm-commits

Differential Revision: https://reviews.llvm.org/D54504

llvm-svn: 347603

5 years ago[stack-safety] Empty local passes for Stack Safety Local Analysis
Vitaly Buka [Mon, 26 Nov 2018 21:57:47 +0000 (21:57 +0000)]
[stack-safety] Empty local passes for Stack Safety Local Analysis

Reviewers: eugenis, vlad.tsyrklevich

Subscribers: mgorny, hiraditya, llvm-commits

Differential Revision: https://reviews.llvm.org/D54502

llvm-svn: 347602

5 years ago[cfi] Help sanstats to find binary if they are not at the original location
Vitaly Buka [Mon, 26 Nov 2018 21:48:55 +0000 (21:48 +0000)]
[cfi] Help sanstats to find binary if they are not at the original location

Summary:
By default sanstats search binaries at the same location where they were when
stats was collected. Sometime you can not print report immediately or you need
to move post-processing to another workstation. To support this use-case when
original binary is missing sanstats will fall-back to directory with sanstats
file.

Reviewers: pcc

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D53857

llvm-svn: 347601

5 years ago[cfi] Make sanstats print address of the check
Vitaly Buka [Mon, 26 Nov 2018 21:48:45 +0000 (21:48 +0000)]
[cfi] Make sanstats print address of the check

Summary: Help with off-line symbolization or other type debugging.

Reviewers: pcc

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D53606

llvm-svn: 347600

5 years ago[AArch64] Refactor the scheduling predicates (3/3) (NFC)
Evandro Menezes [Mon, 26 Nov 2018 21:47:46 +0000 (21:47 +0000)]
[AArch64] Refactor the scheduling predicates (3/3) (NFC)

Refactor the scheduling predicates based on `MCInstPredicate`.  In this
case, `AArch64InstrInfo::hasExtendedReg()`.

Differential revision: https://reviews.llvm.org/D54822

llvm-svn: 347599

5 years ago[AArch64] Refactor the scheduling predicates (2/3) (NFC)
Evandro Menezes [Mon, 26 Nov 2018 21:47:41 +0000 (21:47 +0000)]
[AArch64] Refactor the scheduling predicates (2/3) (NFC)

Refactor the scheduling predicates based on `MCInstPredicate`.  In this
case, `AArch64InstrInfo::hasShiftedReg()`.

Differential revision: https://reviews.llvm.org/D54820

llvm-svn: 347598

5 years ago[AArch64] Refactor the scheduling predicates (1/3) (NFC)
Evandro Menezes [Mon, 26 Nov 2018 21:47:28 +0000 (21:47 +0000)]
[AArch64] Refactor the scheduling predicates (1/3) (NFC)

Refactor the scheduling predicates based on `MCInstPredicate`.  In this
case, `AArch64InstrInfo::isScaledAddr()`

Differential revision: https://reviews.llvm.org/D54777

llvm-svn: 347597

5 years agoSupport for inserting profile-directed cache prefetches
Mircea Trofin [Mon, 26 Nov 2018 21:36:18 +0000 (21:36 +0000)]
Support for inserting profile-directed cache prefetches

Summary:
Support for profile-driven cache prefetching (X86)

This change is part of a larger system, consisting of a cache prefetches recommender, create_llvm_prof (https://github.com/google/autofdo), and LLVM.

A proof of concept recommender is DynamoRIO's cache miss analyzer. It processes memory access traces obtained from a running binary and identifies patterns in cache misses. Based on them, it produces a csv file with recommendations. The expectation is that, by leveraging such recommendations, we can reduce the amount of clock cycles spent waiting for data from memory. A microbenchmark based on the DynamoRIO analyzer is available as a proof of concept: https://goo.gl/6TM2Xp.

The recommender makes prefetch recommendations in terms of:

* the binary offset of an instruction with a memory operand;
* a delta;
* and a type (nta, t0, t1, t2)

meaning: a prefetch of that type should be inserted right before the instrution at that binary offset, and the prefetch should be for an address delta away from the memory address the instruction will access.

For example:

0x400ab2,64,nta

and assuming the instruction at 0x400ab2 is:

movzbl (%rbx,%rdx,1),%edx

means that the recommender determined it would be beneficial for a prefetchnta instruction to be inserted right before this instruction, as such:

prefetchnta 0x40(%rbx,%rdx,1)
movzbl (%rbx, %rdx, 1), %edx

The workflow for prefetch cache instrumentation is as follows (the proof of concept script details these steps as well):

1. build binary, making sure -gmlt -fdebug-info-for-profiling is passed. The latter option will enable the X86DiscriminateMemOps pass, which ensures instructions with memory operands are uniquely identifiable (this causes ~2% size increase in total binary size due to the additional debug information).

2. collect memory traces, run analysis to obtain recommendations (see above-referenced DynamoRIO demo as a proof of concept).

3. use create_llvm_prof to convert recommendations to reference insertion locations in terms of debug info locations.

4. rebuild binary, using the exact same set of arguments used initially, to which -mllvm -prefetch-hints-file=<file> needs to be added, using the afdo file obtained at step 3.

Note that if sample profiling feedback-driven optimization is also desired, that happens before step 1 above. In this case, the sample profile afdo file that was used to produce the binary at step 1 must also be included in step 4.

The data needed by the compiler in order to identify prefetch insertion points is very similar to what is needed for sample profiles. For this reason, and given that the overall approach (memory tracing-based cache recommendation mechanisms) is under active development, we use the afdo format as a syntax for capturing this information. We avoid confusing semantics with sample profile afdo data by feeding the two types of information to the compiler through separate files and compiler flags. Should the approach prove successful, we can investigate improvements to this encoding mechanism.

Reviewers: davidxl, wmi, craig.topper

Reviewed By: davidxl, wmi, craig.topper

Subscribers: davide, danielcdh, mgorny, aprantl, eraman, JDevlieghere, llvm-commits

Differential Revision: https://reviews.llvm.org/D54052

llvm-svn: 347596

5 years agoAMDGPU: Record SGPR spills when restoring too
Matt Arsenault [Mon, 26 Nov 2018 21:28:40 +0000 (21:28 +0000)]
AMDGPU: Record SGPR spills when restoring too

It's possible in some cases to have a restore present
without a corresponding spill. Due to an apparent bug
in D54366 <https://reviews.llvm.org/D54366>, only the
restore for a register was emitted. It's probably
always a bug for this to happen, but due to how SGPR
spilling is implemented, this makes the issues appear
worse than it is.

llvm-svn: 347595

5 years agoELF: ICF: Include contents of referenced sections in initial partitioning hash. NFCI.
Peter Collingbourne [Mon, 26 Nov 2018 21:15:47 +0000 (21:15 +0000)]
ELF: ICF: Include contents of referenced sections in initial partitioning hash. NFCI.

On my machine this reduced median link time of lld-speed-test/chrome
from 2.68s to 2.41s. It also reduces link time of Chrome for Android
with a prototype compiler change that causes the compiler to create
large numbers of identical (modulo relocations) sections from >15
minutes to a few seconds.

Differential Revision: https://reviews.llvm.org/D54773

llvm-svn: 347594

5 years ago[LegalizeVectorTypes][X86][ARM][AArch64][PowerPC] Don't use SplitVecOp_TruncateHelper...
Craig Topper [Mon, 26 Nov 2018 21:12:39 +0000 (21:12 +0000)]
[LegalizeVectorTypes][X86][ARM][AArch64][PowerPC] Don't use SplitVecOp_TruncateHelper for FP_TO_SINT/UINT.

SplitVecOp_TruncateHelper tries to promote the result type while splitting FP_TO_SINT/UINT. It then concatenates the result and introduces a truncate to the original result type. But it does this without inserting the AssertZExt/AssertSExt that the regular result type promotion would insert. Nor does it turn FP_TO_UINT into FP_TO_SINT the way normal result type promotion for these operations does. This is bad on X86 which doesn't support FP_TO_SINT until AVX512.

This patch disables the use of SplitVecOp_TruncateHelper for these operations and just lets normal promotion handle it. I've tweaked a couple things in X86ISelLowering to avoid a few obvious regressions there. I believe all the changes on X86 are improvements. The other targets look neutral.

Differential Revision: https://reviews.llvm.org/D54906

llvm-svn: 347593

5 years ago[ThinLTO] Consolidate cache key computation between new/old LTO APIs
Teresa Johnson [Mon, 26 Nov 2018 20:40:37 +0000 (20:40 +0000)]
[ThinLTO] Consolidate cache key computation between new/old LTO APIs

Summary:
The old legacy LTO API had a separate cache key computation, which was
a subset of the cache key computation in the new LTO API (from what I
can tell this is largely just because certain features such as CFI,
dsoLocal, etc are only utilized via the new LTO API). However, having
separate computations is unnecessary (much of the code is duplicated),
and can lead to bugs when adding new optimizations if both cache
computation algorithms aren't updated properly - it's much easier to
maintain if we have a single facility.

This patch refactors the old LTO API code to use the cache key
computation from the new LTO API. To do this, we set up an lto::Config
object and fill in the fields that the old LTO was hashing (the others
will just use the defaults).

There are two notable changes:
- I added a Freestanding flag to the LTO Config. Currently this is only
used by the legacy LTO API. In the patch that added it (D30791) I had
asked about adding it to the new LTO API, but it looks like that was not
addressed. This should probably be discussed as a follow up to this
change, as it is orthogonal.
- The legacy LTO API had some code that was hashing the GUID of all
preserved symbols defined in the module. I looked back at the history of
this (which was added with the original hashing in the legacy LTO API in
D18494), and there is a comment in the review thread that it was added
in preparation for future internalization. We now do the internalization
of course, and that is handled in the new LTO API cache key computation
by hashing the recorded linkage type of all defined globals. Therefore I
didn't try to move over and keep the preserved symbols handling.

Reviewers: steven_wu, pcc

Subscribers: mehdi_amini, inglorion, eraman, dexonsmith, dang, llvm-commits

Differential Revision: https://reviews.llvm.org/D54635

llvm-svn: 347592

5 years ago[SelectionDAG] Teach BaseIndexOffset::match to unwrap the base after looking through...
Craig Topper [Mon, 26 Nov 2018 20:16:33 +0000 (20:16 +0000)]
[SelectionDAG] Teach BaseIndexOffset::match to unwrap the base after looking through an add/or

We might find a target specific node that needs to be unwrapped after we look through an add/or. Otherwise we get inconsistent results if one pointer is just X86WrapperRIP and the other is (add X86WrapperRIP, C)

Differential Revision: https://reviews.llvm.org/D54818

llvm-svn: 347591

5 years ago[X86] Add test case for D54818
Craig Topper [Mon, 26 Nov 2018 20:16:31 +0000 (20:16 +0000)]
[X86] Add test case for D54818

llvm-svn: 347590