platform/upstream/llvm.git
3 years ago[lldb] Fix test_exec_root of API tests
Pavel Labath [Mon, 22 Mar 2021 14:13:15 +0000 (15:13 +0100)]
[lldb] Fix test_exec_root of API tests

lit has grown a feature where it stores the runtimes of all tests.
Normally, these times should be stored in the build directory, but
because our API tests have set test_exec_root to point to the source
tree, it has ended up polluting our checkout and led to the
.lit_test_times.txt being committed to the repository.

Delete this file, and adjust the exec root of API tests. I've also
needed to adjust the root of Shell tests, in order to avoid the two
overlapping.

3 years ago[AArch64][SVE] Test more types in sve-fixed-length-subvector.ll
Joe Ellis [Wed, 17 Mar 2021 11:40:12 +0000 (11:40 +0000)]
[AArch64][SVE] Test more types in sve-fixed-length-subvector.ll

Previously only the i32 type was tested. Now, the {i,f}{16,32,64} types
are tested.

The v8{i,f}16 cases lower differently to the other cases, which is worth
defending. The lowering for the other cases is currently identical, but
probably worth having for the better coverage.

Differential Revision: https://reviews.llvm.org/D98690

3 years ago[X86][AVX] Add missing AVX1 PMULDQ combine tests
Simon Pilgrim [Mon, 22 Mar 2021 12:39:16 +0000 (12:39 +0000)]
[X86][AVX] Add missing AVX1 PMULDQ combine tests

Yet another case of update_llc_test_checks.py not reporting when a RUN doesn't have any matching prefixes

3 years ago[RISCV][NFC] Add test of stack slot sizes of large split arguments
Luís Marques [Mon, 22 Mar 2021 13:36:22 +0000 (13:36 +0000)]
[RISCV][NFC] Add test of stack slot sizes of large split arguments

Illustrates bug 49500 <https://bugs.llvm.org/show_bug.cgi?id=49500>.

3 years ago[mlir][Linalg] Fix linalg on tensor fusion
Nicolas Vasilache [Mon, 22 Mar 2021 12:34:45 +0000 (12:34 +0000)]
[mlir][Linalg] Fix linalg on tensor fusion

- Drop unnecessary occurrences of rewriter.eraseOp: dead linalg ops on tensors should be cleaned up by DCE.
- reimplement the part of Linalg on fusion that constructs the body and block arguments: the previous implementation had too much magic. Instead this spells out all cases explicitly and asserts / introduces TODOs for incorrect cases.

As a consequence, we can use the default traversal order for this pattern.

Differential Revision: https://reviews.llvm.org/D99070

3 years ago[clang][ASTImporter] Add import API for 'const Type *' (NFC).
Balázs Kéri [Mon, 22 Mar 2021 10:50:44 +0000 (11:50 +0100)]
[clang][ASTImporter] Add import API for 'const Type *' (NFC).

There was only an `Import` function for `QualType` but not for `Type`.
For correct import of some AST nodes where not `QualType` is used
an import of `Type *` is needed. (It is the case with
`FieldDecl::getCapturedVLAType`.)

Reviewed By: shafik, teemperor, martong

Differential Revision: https://reviews.llvm.org/D98951

3 years ago[libc++] Run ninja with --verbose
Louis Dionne [Fri, 19 Mar 2021 23:26:15 +0000 (16:26 -0700)]
[libc++] Run ninja with --verbose

This makes it easier to see what exact build commands are used.

Differential Revision: https://reviews.llvm.org/D98992

3 years ago[cmake] Disable GCC 9's -Wpessimizing-move
Martin Storsjö [Fri, 19 Mar 2021 11:45:52 +0000 (13:45 +0200)]
[cmake] Disable GCC 9's -Wpessimizing-move

Similar to the existing code for disabling GCC's -Wredudant-move,
also check for the -Wpessimizing-move option and disable it if
possible.

This silences another bunch of noisy warnings when building LLVM
with GCC 9.

As noted for -Wredundant-move, the code can't be fixed to silence the
warnings while retaining support for older compilers.

Differential Revision: https://reviews.llvm.org/D98942

3 years ago[llvm-objcopy][Support] move writeToOutput helper function to Support.
Alexey Lapshin [Thu, 4 Mar 2021 09:51:30 +0000 (12:51 +0300)]
[llvm-objcopy][Support] move writeToOutput helper function to Support.

writeToOutput function is useful when it is necessary to create different kinds
of streams(based on stream name) and when we need to use a temporary file
while writing(which would be renamed into the resulting file in a success case).
This patch moves the writeToStream helper into the Support library.

Differential Revision: https://reviews.llvm.org/D98426

3 years ago[analyzer] Don't include private gtest headers
Dmitri Gribenko [Mon, 22 Mar 2021 12:30:28 +0000 (13:30 +0100)]
[analyzer] Don't include private gtest headers

3 years ago[lit] Do not forget test times for tests that weren't executed
Roman Lebedev [Mon, 22 Mar 2021 12:24:52 +0000 (15:24 +0300)]
[lit] Do not forget test times for tests that weren't executed

Even though we have read the times before,
we intentionally forget about it for performance reasons.
But that means we also forget all the times for the tests
that weren't executed this time. This is mildly inconvenient.

So, when recording the new times, first re-read the old times,
and update times for the tests that were executed,
thus preserving all original times, too.

3 years ago[NFC][lit] Extract 'test time' reading/writing into standalone functions
Roman Lebedev [Mon, 22 Mar 2021 09:05:04 +0000 (12:05 +0300)]
[NFC][lit] Extract 'test time' reading/writing into standalone functions

Simply refactor code into reusable functions,
to allow read_test_times() to be reused later.

3 years ago[NFC][lit] Add a test showing that timing data for tests not executed is lost
Roman Lebedev [Mon, 22 Mar 2021 10:14:29 +0000 (13:14 +0300)]
[NFC][lit] Add a test showing that timing data for tests not executed is lost

I.e. when you first run lit on a directory, and then on a single test,
the timing knowledge about anything else other than that single test
is lost. This isn't right.

3 years ago[NFCI][lit] Unbreak more lit self-tests after D98179
Roman Lebedev [Mon, 22 Mar 2021 11:44:30 +0000 (14:44 +0300)]
[NFCI][lit] Unbreak more lit self-tests after D98179

All of these depend on the order of tests, so if one runs them twice,
the tests within them will naturally be reordered
using the previous run times, which breaks them.

3 years ago[NFC][lit] discovery: find_tests_for_inputs: avoid py warning when no suites found
Roman Lebedev [Mon, 22 Mar 2021 09:26:07 +0000 (12:26 +0300)]
[NFC][lit] discovery: find_tests_for_inputs: avoid py warning when no suites found

If lit was run on a directory that contained no suites,
then naturally suite[0] will not be there,
and that line would cause python warnings.

So just predicate it with a check that it is there in the first place.

3 years ago[ConstraintElimination] Add gep tests without inbounds.
Florian Hahn [Mon, 15 Mar 2021 11:22:50 +0000 (11:22 +0000)]
[ConstraintElimination] Add gep tests without inbounds.

Add a set of interesting test cases for GEPs without inbounds for
upcoming patches.

3 years ago[LLDB] XFAIL dwarf5-debug_line-file-index.s on arm-linux
Muhammad Omair Javaid [Mon, 22 Mar 2021 12:03:48 +0000 (17:03 +0500)]
[LLDB] XFAIL dwarf5-debug_line-file-index.s on arm-linux

Tests dwarf5-debug_line-file-index.s fails on arm-linux-gnueabihf.
Bug # 49678 has been filed against it.

3 years ago[IR] Add vscale_range IR function attribute
Bradley Smith [Wed, 3 Mar 2021 13:53:30 +0000 (13:53 +0000)]
[IR] Add vscale_range IR function attribute

This attribute represents the minimum and maximum values vscale can
take. For now this attribute is not hooked up to anything during
codegen, this will be added in the future when such codegen is
considered stable.

Additionally hook up the -msve-vector-bits=<x> clang option to emit this
attribute.

Differential Revision: https://reviews.llvm.org/D98030

3 years ago[OpenCL] Support template parameters for as_type
Sven van Haastregt [Mon, 22 Mar 2021 11:59:05 +0000 (11:59 +0000)]
[OpenCL] Support template parameters for as_type

Implement the TreeTransform for AsTypeExpr.  Split `BuildAsTypeExpr`
out of `ActOnAsTypeExpr`, such that we can call the Build method from
the TreeTransform.

Fixes PR47979.

Differential Revision: https://reviews.llvm.org/D98855

3 years ago[clangd] Replace usages of dummy with more descriptive words
Kadir Cetinkaya [Mon, 22 Mar 2021 10:18:18 +0000 (11:18 +0100)]
[clangd] Replace usages of dummy with more descriptive words

Dummy is a word with inappropriate associations. This patch updates the
references to it in clangd code base with more precise ones.

The only user-visible change is the default variable name used when extracting a
variable. It will be named as `placeholder` from now on.

Differential Revision: https://reviews.llvm.org/D99065

3 years ago[clang][flang] Moke the definition of `-module-dir` restricted to Flang
Andrzej Warzynski [Sat, 20 Mar 2021 15:26:46 +0000 (15:26 +0000)]
[clang][flang] Moke the definition of `-module-dir` restricted to Flang

`-module-dir` is a Flang specific option and should not be visible in
Clang. This patch adds `FlangOnlyOption` flag to its definition. This
way Clang will know that it should reject it and skip it when generating
output for `clang -help`.

The definition of `-module-dir` is moved next to other Flang options.
As `-J` is an alias for `-module-dir`, it has to be moved as well (the
alias cannot be defined before the original option). As `gfortran` mode
is effectively no longer supported (*), `-J` is claimed as Flang only
option.

This is a follow-up of a post-commit review for
https://reviews.llvm.org/D95448.

* https://reviews.llvm.org/rG6a75496836ea14bcfd2f4b59d35a1cad4ac58cee

Differential Revision: https://reviews.llvm.org/D99018

3 years ago[AArch64] Add some float -> int -> float conversion patterns
Sjoerd Meijer [Fri, 19 Mar 2021 14:16:17 +0000 (14:16 +0000)]
[AArch64] Add some float -> int -> float conversion patterns

This adds some conversion match patterns for which we want to keep the int
values in FP registers using the corresponding NEON instructions (not the FP
instructions) to avoid more costly int <-> fp register transfers.

Differential Revision: https://reviews.llvm.org/D98956

3 years ago[analyzer][solver] Redesign constraint ranges data structure
Valeriy Savchenko [Fri, 24 Jul 2020 11:13:31 +0000 (14:13 +0300)]
[analyzer][solver] Redesign constraint ranges data structure

ImmutableSet doesn't seem like the perfect fit for the RangeSet
data structure.  It is good for saving memory in a persistent
setting, but not for the case when the population of the container
is tiny.  This commit replaces RangeSet implementation and
redesigns the most common operations to be more efficient.

Differential Revision: https://reviews.llvm.org/D86465

3 years ago[llvm-jitlink] Fix Windows build after 4a8161fe40cc
Stefan Gränitz [Mon, 22 Mar 2021 10:41:59 +0000 (11:41 +0100)]
[llvm-jitlink] Fix Windows build after 4a8161fe40cc

3 years ago[ConstraintElimination] Add multi-dimension GEP tests.
Florian Hahn [Mon, 22 Mar 2021 10:09:19 +0000 (10:09 +0000)]
[ConstraintElimination] Add multi-dimension GEP tests.

Add a set of interesting test cases with multi-dimensional GEPs for
upcoming patches.

3 years ago[llvm-jitlink] Add diagnostic output and port executor to getaddrinfo(3) as well
Stefan Gränitz [Mon, 22 Mar 2021 10:17:11 +0000 (11:17 +0100)]
[llvm-jitlink] Add diagnostic output and port executor to getaddrinfo(3) as well

Add diagnostic output for TCP connections on both sides, llvm-jitlink and llvm-jitlink-executor.
Port the executor to use getaddrinfo(3) as well. This makes the code more symmetric and seems to be the recommended way for implementing the server side.

Reviewed By: rzurob

Differential Revision: https://reviews.llvm.org/D98581

3 years ago[llvm-jitlink] Fix use of getaddrinfo(3) when connecting remote executor via TCP...
Stefan Gränitz [Mon, 22 Mar 2021 10:18:49 +0000 (11:18 +0100)]
[llvm-jitlink] Fix use of getaddrinfo(3) when connecting remote executor via TCP socket

Since llvm-jitlink moved from gethostbyname to getaddrinfo in D95477, it seems to no longer connect to llvm-jitlink-executor via TCP. I can reproduce this behavior on both, Debian 10 and macOS 10.15.7:

```
> llvm-jitlink-executor listen=localhost:10819
--
> llvm-jitlink --oop-executor-connect=localhost:10819 /path/to/obj.o
Failed to resolve localhost:10819
```

Reviewed By: rzurob

Differential Revision: https://reviews.llvm.org/D98579

3 years ago[OpenCL] Use -fdeclare-opencl-builtins for some tests
Sven van Haastregt [Mon, 22 Mar 2021 09:46:28 +0000 (09:46 +0000)]
[OpenCL] Use -fdeclare-opencl-builtins for some tests

This speeds up the test running times, as the large `opencl-c.h`
header no longer needs to be parsed.

3 years agoMake clangd CompletionModel usable even with non-standard (but supported) layout
serge-sans-paille [Mon, 22 Mar 2021 09:05:25 +0000 (10:05 +0100)]
Make clangd CompletionModel usable even with non-standard (but supported) layout

llvm supports specifying a non-standard layout where each project lies in its
own place. Do not assume a fixed layout and use the appropriate cmake variable
instead.

Differential Revision: https://reviews.llvm.org/D96787

3 years ago[NFC] Simpler and faster key computation for getSubtargetImpl memoization
serge-sans-paille [Mon, 22 Mar 2021 08:52:39 +0000 (09:52 +0100)]
[NFC] Simpler and faster key computation for getSubtargetImpl memoization

There's no use in computing a large key that's only used for a memoization
optimization.

3 years ago[mlir] Add an option to still use bottom-up traversal
Adrian Kuegel [Mon, 22 Mar 2021 08:42:57 +0000 (09:42 +0100)]
[mlir] Add an option to still use bottom-up traversal

GreedyPatternRewriteDriver was changed from bottom-up traversal to top-down traversal. Not all passes work yet with that change for traversal order. To give some time for fixing, add an option to allow to switch back to bottom-up traversal. Use this option in FusionOfTensorOpsPass which fails otherwise.

Differential Revision: https://reviews.llvm.org/D99059

3 years ago[Driver] -m32: Add /usr/include/i386-linux-gnu for Debian
Fangrui Song [Mon, 22 Mar 2021 08:27:06 +0000 (01:27 -0700)]
[Driver] -m32: Add /usr/include/i386-linux-gnu for Debian

3 years ago[docs] GettingInvolved: split out flang and openmp meeting series
Kristof Beyls [Sat, 20 Mar 2021 08:09:39 +0000 (09:09 +0100)]
[docs] GettingInvolved: split out flang and openmp meeting series

Split out the flang and openmp meeting series, as each has a separate
canonical page where the information is maintained.
As part of that, also call out the alias analysis series separately as
it doesn't seem to be relevant for just flang.

Differential Revision: https://reviews.llvm.org/D99012

3 years ago[ELF][docs] Add line breaks
Yang Fan [Mon, 22 Mar 2021 08:08:47 +0000 (16:08 +0800)]
[ELF][docs] Add line breaks

3 years ago[analyzer][solver] Fix infeasible constraints (PR49642)
Valeriy Savchenko [Fri, 19 Mar 2021 14:00:00 +0000 (17:00 +0300)]
[analyzer][solver] Fix infeasible constraints (PR49642)

Additionally, this patch puts an assertion checking for feasible
constraints in every place where constraints are assigned to states.

Differential Revision: https://reviews.llvm.org/D98948

3 years ago[lldb] Use CompileUnit::ResolveSymbolContext in SymbolFileDWARF
Kim-Anh Tran [Thu, 18 Mar 2021 20:32:39 +0000 (21:32 +0100)]
[lldb] Use CompileUnit::ResolveSymbolContext in SymbolFileDWARF

SymbolFileDWARF::ResolveSymbolContext is currently unaware that in DWARF5 the primary file is specified at file index 0. As a result it misses to correctly resolve the symbol context for the primary file when DWARF5 debug data is used and the primary file is only specified at index 0.

This change makes use of CompileUnit::ResolveSymbolContext to resolve the symbol context. The ResolveSymbolContext in CompileUnit has been previously already updated to reflect changes in DWARF5
and contains a more readable version. It can resolve more, but will also do a bit more work than
SymbolFileDWARF::ResolveSymbolContext (getting the Module, and going through SymbolFileDWARF::ResolveSymbolContextForAddress), however, it's mostly directed by $resolve_scope
what will be resolved, and ensures that code is easier to maintain if there's only one path.

Reviewed By: labath

Differential Revision: https://reviews.llvm.org/D98619

3 years ago[Driver] Gnu.cpp: remove obsoleted i386 triple detection from end-of-life distributio...
Fangrui Song [Mon, 22 Mar 2021 07:23:54 +0000 (00:23 -0700)]
[Driver] Gnu.cpp: remove obsoleted i386 triple detection from end-of-life distribution versions

This saves 16 openat syscalls for `clang a.cc` on x86_64.

3 years ago[clangd] Fix linker error when linking clang-index-server with shared libraries
Nathan Ridge [Mon, 22 Mar 2021 05:30:18 +0000 (01:30 -0400)]
[clangd] Fix linker error when linking clang-index-server with shared libraries

Fixes https://github.com/clangd/clangd/issues/723

Differential Revision: https://reviews.llvm.org/D99049

3 years ago[PowerPC] Enable redundant TOC save removal on AIX
Qiu Chaofan [Mon, 22 Mar 2021 06:29:22 +0000 (14:29 +0800)]
[PowerPC] Enable redundant TOC save removal on AIX

Reviewed By: shchenz

Differential Revision: https://reviews.llvm.org/D97039

3 years ago[Driver] Clean up Debian multiarch /usr/include/<triplet> madness
Fangrui Song [Mon, 22 Mar 2021 05:40:38 +0000 (22:40 -0700)]
[Driver] Clean up Debian multiarch /usr/include/<triplet> madness

Debian multiarch additionally adds /usr/include/<triplet> and somehow
Android borrowed the idea. (Note /usr/<triplet>/include is already an
include dir...). On Debian, we should just assume a GCC installation is
available and use its triple.

3 years agoFix extraneous context parameter in templated helper function.
Stella Laurenzo [Mon, 22 Mar 2021 04:58:17 +0000 (04:58 +0000)]
Fix extraneous context parameter in templated helper function.

(missed in lattner's overall updates related to D99028)

3 years ago[X86] Pass to transform tdpbf16ps intrinsics to scalar operation.
Bing1 Yu [Mon, 22 Mar 2021 01:48:59 +0000 (09:48 +0800)]
[X86] Pass to transform tdpbf16ps intrinsics to scalar operation.

In previous patch https://reviews.llvm.org/D93594, we only scalarize tilezero, tileload, tilestore and tiledpbssd. In this patch we scalarize tdpbf16ps intrinsic.

Reviewed By: pengfei

Differential Revision: https://reviews.llvm.org/D96110

3 years ago[IndVars] Sharpen context in eliminateIVComparison
Max Kazantsev [Mon, 22 Mar 2021 04:07:32 +0000 (11:07 +0700)]
[IndVars] Sharpen context in eliminateIVComparison

When eliminating comparisons, we can use common dominator of
all its users as context. This gives better results when ICMP is not
computed right before the branch that uses it.

Differential Revision: https://reviews.llvm.org/D98924
Reviewed By: lebedev.ri

3 years ago[JITLink][ELF/x86-64] Add support for R_X86_64_GOTPC64 and R_X86_64_GOT64.
Lang Hames [Mon, 22 Mar 2021 02:58:08 +0000 (19:58 -0700)]
[JITLink][ELF/x86-64] Add support for R_X86_64_GOTPC64 and R_X86_64_GOT64.

Start adding support for ELF x86-64 large code model, PIC relocations.

3 years ago[libc] Add a target "install-llvmlibc" to install LLVM libc static archive.
Siva Chandra [Sat, 20 Mar 2021 04:50:48 +0000 (04:50 +0000)]
[libc] Add a target "install-llvmlibc" to install LLVM libc static archive.

3 years ago[JITLink] Start laying the groundwork for ELF x86-64 large code model support.
Lang Hames [Sun, 21 Mar 2021 03:22:40 +0000 (20:22 -0700)]
[JITLink] Start laying the groundwork for ELF x86-64 large code model support.

Introduces DefineExternalSectionStartAndEndSymbols.h, which defines a template
for a JITLink pass that transforms external symbols meeting a user-supplied
predicate into defined symbols pointing at the start and end of a Section
identified by the predicate. JITLink.h is updated with a new makeAbsolute
function to support this pass.

Also renames BasicGOTAndStubsBuilder to PerGraphGOTAndPLTStubsBuilder -- the new
name better describes the intent of this GOT and PLT stubs builder, and will
help to distinguish it from future GOT and PLT stub builders that build entries
that may be shared between multiple graphs.

3 years ago[JITLink][ELF/x86-64] Add Delta32, NegDelta32, NegDelta64 support.
Lang Hames [Mon, 22 Mar 2021 00:20:07 +0000 (17:20 -0700)]
[JITLink][ELF/x86-64] Add Delta32, NegDelta32, NegDelta64 support.

These were missing, but are used in eh-frame section support.

3 years ago[ASTMatcher] Add AST Matcher support for C++20 coroutine keywords
Chuanqi Xu [Mon, 22 Mar 2021 02:25:32 +0000 (10:25 +0800)]
[ASTMatcher] Add AST Matcher support for C++20 coroutine keywords

Summary: Try to enable the support for C++20 coroutine keywords for AST
Matchers.

Reviewers: sammccall, njames93, aaron.ballman

Differential Revision: https://reviews.llvm.org/D96316

3 years ago[X86][AMX] Add test cases for AMX load/store lowering.
Luo, Yuanke [Sun, 21 Mar 2021 02:58:57 +0000 (10:58 +0800)]
[X86][AMX] Add test cases for AMX load/store lowering.

Differential Revision: https://reviews.llvm.org/D99030

3 years ago[Driver] Detect Debian hack g++-multiarch-incdir.diff to simplify addLibStdCXXInclude...
Fangrui Song [Mon, 22 Mar 2021 00:33:30 +0000 (17:33 -0700)]
[Driver] Detect Debian hack g++-multiarch-incdir.diff to simplify addLibStdCXXIncludePaths call sites

3 years ago[test] Add test for cross compiling on Linux
Fangrui Song [Sun, 21 Mar 2021 22:37:35 +0000 (15:37 -0700)]
[test] Add test for cross compiling on Linux

3 years ago[test] Delete obsoleted debian_multiarch_tree and ubuntu_13.04_multiarch_tree
Fangrui Song [Sun, 21 Mar 2021 22:23:49 +0000 (15:23 -0700)]
[test] Delete obsoleted debian_multiarch_tree and ubuntu_13.04_multiarch_tree

They are quite outdated. Delete them to avoid unnecessary test churn.

3 years agoUpdate examples post OwningRewritePatternList change
Jacques Pienaar [Sun, 21 Mar 2021 22:15:34 +0000 (15:15 -0700)]
Update examples post OwningRewritePatternList change

3 years agoRevert "[lld-macho] Implement -dependency_info (partially - more opcodes needed)"
Nico Weber [Sun, 21 Mar 2021 20:35:38 +0000 (16:35 -0400)]
Revert "[lld-macho] Implement -dependency_info (partially - more opcodes needed)"

This reverts commit c53a1322f329e29446c7625da423f58f09ec1a55.
Test only passes depending on build dir having a lexicographically later name
than the source dir, and doesn't link on mac/win. See
https://reviews.llvm.org/D98559#2640265 onward.

3 years ago[clang][Codegen] EmitBranchOnBoolExpr(): emit prof branch counts even at -O0
Roman Lebedev [Sun, 21 Mar 2021 20:22:41 +0000 (23:22 +0300)]
[clang][Codegen] EmitBranchOnBoolExpr(): emit prof branch counts even at -O0

This restores the original behaviour before i unadvertedly broke it in
e3a470162738871bba982416748ae5f5e3572947 and clang/test/Profile/ caught it.

3 years ago[clang][CodeGen] Lower Likelihood attributes to @llvm.expect intrin instead of branch...
Roman Lebedev [Sun, 21 Mar 2021 19:13:47 +0000 (22:13 +0300)]
[clang][CodeGen] Lower Likelihood attributes to @llvm.expect intrin instead of branch weights

08196e0b2e1f8aaa8a854585335c17ba479114df exposed LowerExpectIntrinsic's
internal implementation detail in the form of
LikelyBranchWeight/UnlikelyBranchWeight options to the outside.

While this isn't incorrect from the results viewpoint,
this is suboptimal from the layering viewpoint,
and causes confusion - should transforms also use those weights,
or should they use something else, D98898?

So go back to status quo by making LikelyBranchWeight/UnlikelyBranchWeight
internal again, and fixing all the code that used it directly,
which currently is only clang codegen, thankfully,
to emit proper @llvm.expect intrinsics instead.

3 years agoRevert "[BranchProbability] move options for 'likely' and 'unlikely'"
Roman Lebedev [Sun, 21 Mar 2021 16:55:21 +0000 (19:55 +0300)]
Revert "[BranchProbability] move options for 'likely' and 'unlikely'"

Upon reviewing D98898 i've come to realization that these are
implementation detail of LowerExpectIntrinsicPass,
and they should not be exposed to outside of it.

This reverts commit ee8b53815ddf6f6f94ade0068903cd5ae843fafa.

3 years ago[Driver] Gnu.cpp: fix libstdc++ search path for multilib
Fangrui Song [Sun, 21 Mar 2021 19:01:44 +0000 (12:01 -0700)]
[Driver] Gnu.cpp: fix libstdc++ search path for multilib

With this change, on Debian x86-64 (with a MULTILIB_OSDIRNAMES local patch
../lib64 -> ../lib; this does not matter because /usr/lib64/crt{1,i,n}.o do not exist),
`clang++ --target=aarch64-linux-gnu a.cc -Wl,--dynamic-linker=/usr/aarch64-linux-gnu/lib/ld-linux-aarch64.so.1 -Wl,-rpath,/usr/aarch64-linux-gnu/lib`
built executable can run under qemu-user. Previously this failed with
`/usr/lib/gcc-cross/aarch64-linux-gnu/10/../../../../include/c++/10/iostream:38:10: fatal error: 'bits/c++config.h' file not found`

On Arch Linux, due to the MULTILIB_OSDIRNAMES patch and the existence of
/usr/lib64/crt{1,i,n}.o, clang driver may pick
/usr/lib64/crt{1,i,n}.o and cause a linker error. -B can work around the problem.
`clang++ --target=aarch64-linux-gnu -B /usr/aarch64-linux-gnu/lib a.cc -Wl,--dynamic-linker=/usr/aarch64-linux-gnu/lib/ld-linux-aarch64.so.1 -Wl,-rpath,/usr/aarch64-linux-gnu/lib64:/usr/aarch64-linux-gnu/lib`

3 years ago[lld-macho] Implement -dependency_info (partially - more opcodes needed)
Vy Nguyen [Fri, 12 Mar 2021 22:40:37 +0000 (17:40 -0500)]
[lld-macho] Implement -dependency_info (partially - more opcodes needed)

Bug: https://bugs.llvm.org/show_bug.cgi?id=49278
The flag is not well documented, so this implementation is based on observed behaviour.

When specified, `-dependency_info <path>` produced a text file containing information pertaining to the current linkage, such as input files, output file, linker version, etc.

This file's layout is also not documented, but it seems to be a series of null ('\0') terminated strings in the form `<op code><path>`

`<op code>` could be:
   `0x00` : linker version
   `0x10` : input
   `0x11` : files not found(??)
   `0x40` : output

`<path>` : is the file path, except for the linker-version case.

(??) This part is a bit unclear. I think it means all the files the linker attempted to look at, but could not find.

Differential Revision: https://reviews.llvm.org/D98559

3 years ago[DAGCombiner] Minor compile time improvement to (sext_in_reg (sign_extend_vector_inre...
Craig Topper [Sun, 21 Mar 2021 17:44:31 +0000 (10:44 -0700)]
[DAGCombiner] Minor compile time improvement to (sext_in_reg (sign_extend_vector_inreg x)) optimization.

Don't bother calling ComputeNumSignBits if N00Bits < ExtVTBits. No
matter what answer we get back this will be true:
(N00Bits - DAG.ComputeNumSignBits(N00, DemandedSrcElts)) < ExtVTBits)

So we might as well save the computation. This makes the code more
consistent with the similar (sext_in_reg (sext x)) handling above.

3 years ago[ValueTracking] Improve mul handling in isKnownNonEqual()
Nikita Popov [Sun, 21 Mar 2021 17:36:20 +0000 (18:36 +0100)]
[ValueTracking] Improve mul handling in isKnownNonEqual()

X != X * C is true if:
 * C is not 0 or 1
 * X is not 0
 * mul is nsw or nuw

Proof: https://alive2.llvm.org/ce/z/uwF29z

This is motivated by one of the cases in D98422.

3 years ago[ValueTracking] Add more tests for isKnownNonEqual() of mul (NFC)
Nikita Popov [Sun, 21 Mar 2021 17:14:43 +0000 (18:14 +0100)]
[ValueTracking] Add more tests for isKnownNonEqual() of mul (NFC)

This is for the case of (x * C) == x, rather than the
(x * C1) == (x * C2) variant that we already cover.

3 years agoRemove the extraneous MLIRContext argument from populateWithGenerated. NFC.
Chris Lattner [Sun, 21 Mar 2021 17:38:35 +0000 (10:38 -0700)]
Remove the extraneous MLIRContext argument from populateWithGenerated.  NFC.

3 years agoMIR: Fix missing serialization for HasTailCall
Matt Arsenault [Sun, 21 Mar 2021 16:00:55 +0000 (12:00 -0400)]
MIR: Fix missing serialization for HasTailCall

3 years agoAMDGPU: Fix allowing immediates for tail call pseudo.
Matt Arsenault [Sun, 14 Mar 2021 20:59:34 +0000 (16:59 -0400)]
AMDGPU: Fix allowing immediates for tail call pseudo.

The pseudo was using SSrc_b64, so it allowed folding immediates into
the destination operand for a tail call to null. However, this is not
a valid operand for the s_setpc_b64 this will be lowered to. Avoids
printing the operand as an invalid immediate.

Avoids a regression when tail calls are enabled in GlobalISel (somehow
tail calls to null get deleted in the DAG).

3 years ago[ShapeDialect] Silence a build warning, NFC
Chris Lattner [Sun, 21 Mar 2021 17:10:38 +0000 (10:10 -0700)]
[ShapeDialect] Silence a build warning, NFC

mlir/lib/Dialect/Shape/IR/Shape.cpp:573:26: warning: loop variable 'shape' is always a copy because the range of type '::mlir::Operation::operand_range' (aka 'mlir::OperandRange') does not return a reference [-Wrange-loop-analysis]
        for (const auto &shape : shapes()) {
                         ^

3 years agoChange OwningRewritePatternList to carry an MLIRContext with it.
Chris Lattner [Sat, 20 Mar 2021 23:29:41 +0000 (16:29 -0700)]
Change OwningRewritePatternList to carry an MLIRContext with it.

This updates the codebase to pass the context when creating an instance of
OwningRewritePatternList, and starts removing extraneous MLIRContext
parameters.  There are many many more to be removed.

Differential Revision: https://reviews.llvm.org/D99028

3 years agoReapply [ConstantFold] Handle vectors in ConstantFoldLoadThroughBitcast()
Nikita Popov [Sat, 6 Mar 2021 11:14:48 +0000 (12:14 +0100)]
Reapply [ConstantFold] Handle vectors in ConstantFoldLoadThroughBitcast()

There seems to be an impedance mismatch between what the type
system considers an aggregate (structs and arrays) and what
constants consider an aggregate (structs, arrays and vectors).

Adjust the type check to consider vectors as well. The previous
version of the patch dropped the type check entirely, but it
turns out that getAggregateElement() does require the constant
to be an aggregate in some edge cases: For Poison/Undef the
getNumElements() API is called, without checking in advance that
we're dealing with an aggregate. Possibly the implementation should
avoid doing that, but for now I'm adding an assert so the next
person doesn't fall into this trap.

3 years ago[InstSimplify] Add load of undef aggregate test (NFC)
Nikita Popov [Sun, 21 Mar 2021 16:40:17 +0000 (17:40 +0100)]
[InstSimplify] Add load of undef aggregate test (NFC)

To make sure this doesn't crash the following commit.

3 years ago[InstSimplify] Regenerate test checks (NFC)
Nikita Popov [Sun, 21 Mar 2021 16:32:14 +0000 (17:32 +0100)]
[InstSimplify] Regenerate test checks (NFC)

3 years ago[InstSimplify] Add additional select operand replacement tests (NFC)
Nikita Popov [Sun, 21 Mar 2021 12:45:23 +0000 (13:45 +0100)]
[InstSimplify] Add additional select operand replacement tests (NFC)

This tests for binops with identity elements.

3 years ago[InstSimplify] Clean up SimplifyReplacedWithOp implementation (NFCI)
Nikita Popov [Sun, 21 Mar 2021 12:32:24 +0000 (13:32 +0100)]
[InstSimplify] Clean up SimplifyReplacedWithOp implementation (NFCI)

Replace Op with RepOp up-front, and then always work with the new
operands, rather than checking for replacement in various places.

3 years agoGlobalISel: Avoid unnecessary truncation to i64
Matt Arsenault [Sat, 20 Mar 2021 17:42:17 +0000 (13:42 -0400)]
GlobalISel: Avoid unnecessary truncation to i64

We can just directly pass through the APInt to create a new constant.

3 years agoAMDGPU/GlobalISel: Enable CSE in pre-legalizer combiner
Matt Arsenault [Sat, 20 Mar 2021 16:53:58 +0000 (12:53 -0400)]
AMDGPU/GlobalISel: Enable CSE in pre-legalizer combiner

3 years ago[DAG] Limit (sext_in_reg (zero_extend_vector_inreg x)) to exact sign extension
Simon Pilgrim [Sun, 21 Mar 2021 14:00:59 +0000 (14:00 +0000)]
[DAG] Limit (sext_in_reg (zero_extend_vector_inreg x)) to exact sign extension

As commented by @craig.topper on rG1ba5c550d418, we can't guarantee that we'll be extending zero bits, just sign bit. So, revert to the old code for zero_extend_vector_inreg cases.

3 years ago[lld-macho][nfc] Format Options.td
Jez Ng [Sun, 21 Mar 2021 05:10:04 +0000 (01:10 -0400)]
[lld-macho][nfc] Format Options.td

Summary: A good chunk of it was mis-indented. Fixed by using the
formatting settings from llvm/utils/vim.

3 years ago[X86][AVX] ComputeNumSignBitsForTargetNode - add X86ISD::VBROADCAST handling for...
Simon Pilgrim [Sun, 21 Mar 2021 12:22:51 +0000 (12:22 +0000)]
[X86][AVX] ComputeNumSignBitsForTargetNode - add X86ISD::VBROADCAST handling for scalar sources

The target shuffle code handles vector sources, but X86ISD::VBROADCAST can also accept a scalar source for splatting.

Added as an extension to PR49658

3 years ago[X86] Add 'mulhs' variant of PR49658 test case
Simon Pilgrim [Sun, 21 Mar 2021 12:08:53 +0000 (12:08 +0000)]
[X86] Add 'mulhs' variant of PR49658 test case

3 years ago[ARM] VINS f16 pattern
David Green [Sun, 21 Mar 2021 12:00:06 +0000 (12:00 +0000)]
[ARM] VINS f16 pattern

This adds an extra pattern for inserting an f16 into a odd vector lane
via an VINS. If the dual-insert-lane pattern does not happen to apply,
this can help with some simple cases.

Differential Revision: https://reviews.llvm.org/D95471

3 years ago[RISCV] remove redundant instruction when eliminate frame index
luxufan [Fri, 19 Mar 2021 09:02:28 +0000 (17:02 +0800)]
[RISCV] remove redundant instruction when eliminate frame index

The reason for generating mv a0, a0 instruction is when the stack object offset is large then int<12>. To deal this situation, in the elimintateFrameIndex function, it will
create a virtual register, which needs the register scavenger to scavenge it. If the machine instruction that contains the stack object and the opcode is ADDI(the addi
was generated by frameindexNode), and then this instruction's destination register was the same as the register that was generated by the register scavenger, then the
mv a0, a0 was generated. So to eliminnate this instruction, in the eliminateFrameIndex function, if the instrution opcode is ADDI, then the virtual register can't be created.

Differential Revision: https://reviews.llvm.org/D92479

3 years ago[X86][AVX] computeKnownBitsForTargetNode - add X86ISD::VBROADCAST handling for scalar...
Simon Pilgrim [Sun, 21 Mar 2021 10:40:57 +0000 (10:40 +0000)]
[X86][AVX] computeKnownBitsForTargetNode - add X86ISD::VBROADCAST handling for scalar sources

The target shuffle code handles vector sources, but X86ISD::VBROADCAST can also accept a scalar source for splatting.

Suggested by @craig.topper on PR49658

3 years ago[X86] Add PR49658 test case
Simon Pilgrim [Sun, 21 Mar 2021 10:16:55 +0000 (10:16 +0000)]
[X86] Add PR49658 test case

3 years ago[X86] computeKnownBitsForTargetNode - add X86ISD::PMULUDQ handling
Simon Pilgrim [Sun, 21 Mar 2021 09:57:20 +0000 (09:57 +0000)]
[X86] computeKnownBitsForTargetNode - add X86ISD::PMULUDQ handling

Reuse the existing KnownBits multiplication code to handle what is effectively a ISD::UMUL_LOHI varient

3 years ago[Driver] Linux.cpp: add -internal-isystem lib/../$triple/include
Fangrui Song [Sun, 21 Mar 2021 07:56:03 +0000 (00:56 -0700)]
[Driver] Linux.cpp: add -internal-isystem lib/../$triple/include

With this change, for `#include <ar.h>`, `clang --target=aarch64-linux-gnu`
will read `/usr/lib/gcc/aarch64-linux-gnu/10/../../../../aarch64-linux-gnu/include/ar.h`
(on Debian gcc->gcc-cross)
instead of `/usr/include/ar.h`. Some glibc headers (e.g. gnu/stubs.h) are different across architectures.

3 years ago[Driver] Gnu.cpp: drop an unneeded special rule related to sysroot
Fangrui Song [Sun, 21 Mar 2021 04:37:49 +0000 (21:37 -0700)]
[Driver] Gnu.cpp: drop an unneeded special rule related to sysroot

3 years ago[Driver] Gnu.cpp: drop an unneeded special rule related to sysroot
Fangrui Song [Sun, 21 Mar 2021 04:32:55 +0000 (21:32 -0700)]
[Driver] Gnu.cpp: drop an unneeded special rule related to sysroot

Seem unnecessary to diverge from GCC here.
Beside, lib/../$OSLibDir can be considered closer to the GCC
installation then the system root. The comment should not apply.

3 years ago[Driver] Gnu.cpp: remove unneeded -L detection hack for -mx32
Fangrui Song [Sun, 21 Mar 2021 03:12:45 +0000 (20:12 -0700)]
[Driver] Gnu.cpp: remove unneeded -L detection hack for -mx32

Removing the hack actually improves our compatibility with gcc -mx32.

3 years ago[Driver] Gnu.cpp: remove unneeded -L detection for libc++
Fangrui Song [Sun, 21 Mar 2021 01:56:40 +0000 (18:56 -0700)]
[Driver] Gnu.cpp: remove unneeded -L detection for libc++

If clang is installed in the system, the other -L suffice;
otherwise $ccc_install_dir/../lib below suffices.

3 years ago[Driver] Gnu.cpp: remove unneeded -L lib/gcc/$triple/$version/../../../$triple
Fangrui Song [Sun, 21 Mar 2021 01:50:14 +0000 (18:50 -0700)]
[Driver] Gnu.cpp: remove unneeded -L lib/gcc/$triple/$version/../../../$triple

After path resolution, it duplicates a subsequent -L entry. The entry below
(lib/gcc/$triple/$version/../../../../$OSLibDir) usually does not exist (e.g.
Arch Linux; Debian cross gcc). When it exists, it typically just has ld.so (e.g.
Debian native gcc) which cannot cause collision. Removing the -L (similar to
reordering it) is therefore justified.

3 years ago[RISCV] Add test case to show a case where (mul (and X, 0xffffffff), (and Y, 0xffffff...
Craig Topper [Sun, 21 Mar 2021 00:43:30 +0000 (17:43 -0700)]
[RISCV] Add test case to show a case where (mul (and X, 0xffffffff), (and Y, 0xffffffff)) optimization does not improve code.

If the mul add two users, one of which was a sext.w, the mul
would also be selected to a MULW before our pattern runs. This
causes the ANDs to now be used by the already selected MULW and
the mul we still need to select. They are unneeded on the MULW
since MULW only reads the lower bits. So they get selected to
SLLI+SRLI for the MULW use. The use for the
(mul (and X, 0xffffffff), (and Y, 0xffffffff)) manages to reuse
the SLLI.

The end result is increased register pressure and no improvement
to how soon we can start the MULW.

3 years ago[Canonicalizer] Process regions top-down instead of bottom up & reuse existing constants.
Chris Lattner [Sat, 20 Mar 2021 04:22:15 +0000 (21:22 -0700)]
[Canonicalizer] Process regions top-down instead of bottom up & reuse existing constants.

This reapplies b5d9a3c / https://reviews.llvm.org/D98609 with a one line fix in
processExistingConstants to skip() when erasing a constant we've already seen.

Original commit message:

 1) Change the canonicalizer to walk the function in top-down order instead of
    bottom-up order.  This composes well with the "top down" nature of constant
    folding and simplification, reducing iterations and re-evaluation of ops in
    simple cases.
 2) Explicitly enter existing constants into the OperationFolder table before
    canonicalizing.  Previously we would "constant fold" them and rematerialize
    them, wastefully recreating a bunch fo constants, which lead to pointless
    memory traffic.

Both changes together provide a 33% speedup for canonicalize on some mid-size
CIRCT examples.

One artifact of this change is that the constants generated in normal pattern
application get inserted at the top of the function as the patterns are applied.
Because of this, we get "inverted" constants more often, which is an aethetic
change to the IR but does permute some testcases.

Differential Revision: https://reviews.llvm.org/D99006

3 years agoRevert "[IRSim] Adding basic implementation of llvm-sim."
Andrew Litteken [Sat, 20 Mar 2021 23:03:02 +0000 (18:03 -0500)]
Revert "[IRSim] Adding basic implementation of llvm-sim."

Causing build errors on the Windows Buildbots.

This reverts commit 5155dff2784a47583d432d796b7cf47a0bed9f20.

3 years ago[RISCV] Update comment in RISCVInstrInfoM.td
Jessica Clarke [Sat, 20 Mar 2021 22:35:40 +0000 (22:35 +0000)]
[RISCV] Update comment in RISCVInstrInfoM.td

Missed in 07ed62b7d551.

3 years ago[RISCV] Disable (mul (and X, 0xffffffff), (and Y, 0xffffffff)) optimization when...
Craig Topper [Sat, 20 Mar 2021 22:14:46 +0000 (15:14 -0700)]
[RISCV] Disable (mul (and X, 0xffffffff), (and Y, 0xffffffff)) optimization when Zba is enabled.

This optimization is trying to save SRLI instructions needed to
implement the ANDs. If we have zext.w we won't save anything.
Because we don't check that the multiply is the only user of the
AND we might even increase instruction count.

3 years ago[RISCV] Add Zba command lines to xaluo.ll. NFC
Craig Topper [Sat, 20 Mar 2021 22:09:15 +0000 (15:09 -0700)]
[RISCV] Add Zba command lines to xaluo.ll. NFC

Some of the patterns end up with 32 to 64 bit zero extends on RV64
which can be handled by zext.w.

3 years ago[test] Delete "-internal-isystem" "/usr/local/include"
Fangrui Song [Sat, 20 Mar 2021 22:24:02 +0000 (15:24 -0700)]
[test] Delete "-internal-isystem" "/usr/local/include"

3 years ago[RISCV] Add isel pattern to optimize (mul (and X, 0xffffffff), (and Y, 0xffffffff...
Craig Topper [Sat, 20 Mar 2021 19:34:06 +0000 (12:34 -0700)]
[RISCV] Add isel pattern to optimize (mul (and X, 0xffffffff), (and Y, 0xffffffff)) on RV64

This patterns computes the full 64 bit product of a 32x32 unsigned
multiply. This requires a two pairs of SLLI+SRLI to zero the
upper 32 bits of the inputs.

We can do better than this by using two SLLI to move the lower
bits to the upper bits then use MULHU to compute the product. This
is the high half of a full 64x64 product. Since we put 32 0s in the lower
bits of the inputs we know the 128-bit product will have zeros in the
lower 64 bits. So the upper 64 bits, which MULHU computes, will contain
the original 64 bit product we were after.

The same trick would work for (mul (sext_inreg X, i32), (sext_inreg Y, i32))
using MULHS, but sext_inreg is sext.w which is already one instruction so we
wouldn't save anything.

Differential Revision: https://reviews.llvm.org/D99026

3 years ago[IRSim] Adding basic implementation of llvm-sim.
Andrew Litteken [Thu, 17 Sep 2020 20:43:40 +0000 (15:43 -0500)]
[IRSim] Adding basic implementation of llvm-sim.

This is a similarity visualization tool that accepts a Module and
passes it to the IRSimilarityIdentifier.  The resulting SimilarityGroups
are output in a JSON file.

Tests are found in test/tools/llvm-sim and check for the file not found,
a bad module, and that the JSON is created correctly.

Reviewers: paquette, jroelofs, MaskRay

Recommit of: 15645d044bcfe2a0f63156048b302f997a717688 to fix linking
errors.

Differential Revision: https://reviews.llvm.org/D86974

3 years ago[AIX] Update rpath for BUILD_SHARED_LIBS
Jinsong Ji [Sat, 20 Mar 2021 03:48:48 +0000 (03:48 +0000)]
[AIX] Update rpath for BUILD_SHARED_LIBS

BUILD_SHARED_LIBS build llvm component as shared library,
which can reduce the size a lot.

Normally, the binary use ORIGIN../lib to load component libraries,
unfortunatly, ORIGIN is not supported by AIX ld.

We hardcoded the build lib and install lib path in rpath for now
to enable BUILD_SHARED_LIBS build.

Understand that this is not perfect solution,
we can update this when we find better solution.

Reviewed By: hubert.reinterpretcast

Differential Revision: https://reviews.llvm.org/D98901

3 years ago[test] Fix Driver/gcc-toolchain.cpp if CLANG_DEFAULT_RTLIB is compiler-rt
Fangrui Song [Sat, 20 Mar 2021 20:24:49 +0000 (13:24 -0700)]
[test] Fix Driver/gcc-toolchain.cpp if CLANG_DEFAULT_RTLIB is compiler-rt