platform/upstream/llvm.git
4 years ago[DWARFYAML] Make the ExtLen field of extended opcodes optional.
Xing GUO [Thu, 24 Sep 2020 06:11:47 +0000 (14:11 +0800)]
[DWARFYAML] Make the ExtLen field of extended opcodes optional.

This patch makes the 'ExtLen' field of extended opcodes optional. We
don't need to manually calculate it in the future.

Reviewed By: jhenderson, MaskRay

Differential Revision: https://reviews.llvm.org/D88136

4 years ago[IRSim][NFC] Removing warning from IRSimilarityIdentifier
Andrew Litteken [Thu, 24 Sep 2020 05:26:21 +0000 (00:26 -0500)]
[IRSim][NFC] Removing warning from IRSimilarityIdentifier

4 years agoDebugInfo: Filter DWARFv5 TUs out of the debug_info unit list when CUs requested
David Blaikie [Fri, 18 Sep 2020 18:30:50 +0000 (11:30 -0700)]
DebugInfo: Filter DWARFv5 TUs out of the debug_info unit list when CUs requested

Since DWARFv5 places TUs in debug_info, some of DWARFContext's APIs have
become a bit erroneous, including TUs in the CU list by accident.
Correct that by providing compile_units (& dwo_compile_units) that
filter out the type units from the debug_info units.

Differential Revision: https://reviews.llvm.org/D87935

4 years ago[dwarfdump] Warn for tags with DW_CHILDREN_yes but no children.
Jonas Devlieghere [Thu, 24 Sep 2020 05:00:57 +0000 (22:00 -0700)]
[dwarfdump] Warn for tags with DW_CHILDREN_yes but no children.

Flag DIEs that have DW_CHILDREN_yes set in their abbreviation but don't
actually have any children.

rdar://59809554

Differential revision: https://reviews.llvm.org/D88048

4 years ago[Power10] Implement custom codegen for the vec_replace_elt and vec_replace_unaligned...
Amy Kwan [Wed, 23 Sep 2020 22:14:24 +0000 (17:14 -0500)]
[Power10]  Implement custom codegen for the vec_replace_elt and vec_replace_unaligned builtins.

This patch implements custom codegen for the vec_replace_elt and
vec_replace_unaligned builtins.

These builtins map to the @llvm.ppc.altivec.vinsw and @llvm.ppc.altivec.vinsd
intrinsics depending on the arguments. The main motivation for doing custom
codegen for these intrinsics is because there are float and double versions of
the builtin. Normally, the converting the float to an integer would be done via
fptoui in the IR. This is incorrect as fptoui truncates the value and we must
ensure the value is not truncated. Therefore, we provide custom codegen to utilize
bitcast instead as bitcasts do not truncate.

Differential Revision: https://reviews.llvm.org/D83500

4 years ago[IRSim] Adding structural comparison to IRSimilarityCandidate.
Andrew Litteken [Wed, 23 Sep 2020 16:04:39 +0000 (11:04 -0500)]
[IRSim] Adding structural comparison to IRSimilarityCandidate.

Just because sequences of instructions are similar to one another,
doesn't mean they are doing the same thing.

This introduces a structural check for the IRSimilarityCandidate that
compares two IRSimilarityCandidates against one another, and in each
instruction creates a mapping between the operands and results, or
checks that the existing mapping is valid.  If this check passes, it
means we have structurally similar IRSimilarityCandidates.

Tests for whether the candidates are found in
unittests/Analysis/IRSimilarityIdentifierTest.cpp.

Recommit of: b27db2bb68163fa5bcb4a8f631a305eb5adb44e5 for Differential
URL.

Differential Revision: https://reviews.llvm.org/D86971

4 years agoRevert "[IRSim] Adding structural comparison to IRSimilarityCandidate."
Andrew Litteken [Thu, 24 Sep 2020 03:40:37 +0000 (22:40 -0500)]
Revert "[IRSim] Adding structural comparison to IRSimilarityCandidate."

This reverts commit b27db2bb68163fa5bcb4a8f631a305eb5adb44e5.

4 years ago[IRSim] Adding structural comparison to IRSimilarityCandidate.
Andrew Litteken [Wed, 23 Sep 2020 16:04:39 +0000 (11:04 -0500)]
[IRSim] Adding structural comparison to IRSimilarityCandidate.

Just because sequences of instructions are similar to one another,
doesn't mean they are doing the same thing.

This introduces a structural check for the IRSimilarityCandidate that
compares two IRSimilarityCandidates against one another, and in each
instruction creates a mapping between the operands and results, or
checks that the existing mapping is valid.  If this check passes, it
means we have structurally similar IRSimilarityCandidates.

Tests for whether the candidates are found in
unittests/Analysis/IRSimilarityIdentifierTest.cpp.

4 years ago[lld-macho][NFC] Make `!= nullptr` implicit
Jez Ng [Thu, 24 Sep 2020 03:09:49 +0000 (20:09 -0700)]
[lld-macho][NFC] Make `!= nullptr` implicit

4 years ago[lld-macho][NFC] Refactor syslibroot / library path lookup
Jez Ng [Sat, 19 Sep 2020 03:51:38 +0000 (20:51 -0700)]
[lld-macho][NFC] Refactor syslibroot / library path lookup

* Move computation of systemLibraryRoots into a separate function, so we
  can add more functionality to it without things becoming unwieldy
* Have `getSearchPaths` and related functions return by value instead of
  by output parameter. NRVO should ensure that performance is unaffected.

Reviewed By: #lld-macho, smeenai

Differential Revision: https://reviews.llvm.org/D87959

4 years ago[lld-macho] Support -weak_lx, -weak_library, -weak_framework
Jez Ng [Fri, 18 Sep 2020 18:38:15 +0000 (11:38 -0700)]
[lld-macho] Support -weak_lx, -weak_library, -weak_framework

They cause their corresponding libraries / frameworks to be loaded via
`LC_LOAD_WEAK_DYLIB` instead of `LC_LOAD_DYLIB`.

Reviewed By: #lld-macho, gkm

Differential Revision: https://reviews.llvm.org/D87929

4 years ago[lld-macho] Ignore `-mllvm` and its argument
Jez Ng [Wed, 16 Sep 2020 23:07:21 +0000 (16:07 -0700)]
[lld-macho] Ignore `-mllvm` and its argument

Test Plan:

Reviewed By: #lld-macho, compnerd, MaskRay

Differential Revision: https://reviews.llvm.org/D87803

4 years ago[lld-macho] Emit indirect symbol table
Jez Ng [Sat, 5 Sep 2020 01:02:07 +0000 (18:02 -0700)]
[lld-macho] Emit indirect symbol table

Makes it a little easier to read objdump's disassembly.

Reviewed By: #lld-macho, gkm

Differential Revision: https://reviews.llvm.org/D87178

4 years ago[lld-macho] Implement and test resolution of common symbols
Jez Ng [Tue, 1 Sep 2020 03:57:19 +0000 (20:57 -0700)]
[lld-macho] Implement and test resolution of common symbols

Handle the case where there are both common and non-common definitions
of the same symbol. Add a bunch of tests to ensure compatibility with ld64.

Reviewed By: #lld-macho, gkm

Differential Revision: https://reviews.llvm.org/D86910

4 years ago[lld-macho] Initial support for common symbols
Jez Ng [Fri, 28 Aug 2020 05:54:43 +0000 (22:54 -0700)]
[lld-macho] Initial support for common symbols

On Unix, it is traditionally allowed to write variable definitions without
initialization expressions (such as "int foo;") to header files. These are
called tentative definitions.

The compiler creates common symbols when it sees tentative definitions. When
linking the final binary, if there are remaining common symbols after name
resolution is complete, the linker converts them to regular defined symbols in
a `__common` section.

This diff implements most of that functionality, though we do not yet handle
the case where there are both common and non-common definitions of the same
symbol.

Reviewed By: #lld-macho, gkm

Differential Revision: https://reviews.llvm.org/D86909

4 years ago[GlobalISel][AMDGPU] Lower G_SMULH/G_UMULH
Pushpinder Singh [Mon, 10 Aug 2020 09:47:50 +0000 (05:47 -0400)]
[GlobalISel][AMDGPU] Lower G_SMULH/G_UMULH

Reviewed By: arsenm, foad

Differential Revision: https://reviews.llvm.org/D85653

4 years ago[NFC] Remove unnecessary default constructors
Arthur Eubanks [Thu, 24 Sep 2020 01:52:35 +0000 (18:52 -0700)]
[NFC] Remove unnecessary default constructors

4 years ago[AMDGPU] Consider all SGPR uses as unique in constant bus verify
Carl Ritson [Thu, 24 Sep 2020 01:28:46 +0000 (10:28 +0900)]
[AMDGPU] Consider all SGPR uses as unique in constant bus verify

Fix the verifier so that overlapping SGPR operands are counted
independently.  We cannot assume that overlapping SGPR accesses
only count as a single constant bus use.
The exception is implicit uses which do not add to constant bus
usage (only) when overlapping.

Reviewed By: rampitec

Differential Revision: https://reviews.llvm.org/D87748

4 years agoRevert "[NewPM] Add callbacks to PassBuilder to run before/after parsing a pass"
Arthur Eubanks [Thu, 24 Sep 2020 01:43:13 +0000 (18:43 -0700)]
Revert "[NewPM] Add callbacks to PassBuilder to run before/after parsing a pass"

This reverts commit 111aa4e36614d9a056cf5040d4d7bbfddeb9ebb2.

4 years ago[X86] Parse data32 call in .code16 as CALLpcrel32
Fangrui Song [Thu, 24 Sep 2020 01:36:50 +0000 (18:36 -0700)]
[X86] Parse data32 call in .code16 as CALLpcrel32

Used by kexec-tools (PR46942)
In GNU as, tc-i386.c:output_jump uses 4-byte immediate if a data32 prefix is present.

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D88137

4 years ago[X86] Add TDX instructions.
Freddy Ye [Thu, 24 Sep 2020 01:12:28 +0000 (09:12 +0800)]
[X86] Add TDX instructions.

For more details about these instructions, please refer to the latest TDX document: https://software.intel.com/content/www/us/en/develop/articles/intel-trust-domain-extensions.html

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D88006

4 years agoUtility: ignore OS version on non-Darwin targets in `ArchSpec`
Saleem Abdulrasool [Wed, 23 Sep 2020 20:57:55 +0000 (20:57 +0000)]
Utility: ignore OS version on non-Darwin targets in `ArchSpec`

The OS version field is generally not very helpful for non-Darwin
targets.  On Linux, it identifies the kernel version which moves
out-of-sync with the userspace.  On Windows, this field actually ends up
corresponding to the Visual Studio toolset version instead of the OS
version.  Consider non-Darwin targets without an OS version to be fully
specified.

Differential Revision: https://reviews.llvm.org/D88181
Reviewed By: Jonas Devlieghere, Dave Lee

4 years agoExplicitly specify CMAKE_AR in WinMsvc.cmake
Gwen Mittertreiner [Wed, 23 Sep 2020 23:20:06 +0000 (16:20 -0700)]
Explicitly specify CMAKE_AR in WinMsvc.cmake

As of cmake 3.18, cmake changes how it searches for compilers for
Windows (see
https://gitlab.kitware.com/cmake/cmake/-/commit/55196a1440e26917d40e6a7a3eb8d9fb323fa657)
and now finds llvm-ar instead of llvm-lib as CMAKE_AR. This explicitly
specifies CMAKE_AR as llvm-lib so the correct program is found.

Reviewed By: smeenai

Differential Revision: https://reviews.llvm.org/D88176

4 years ago[CGProfile] don't emit cgprofile entry if called function is dllimport
Zequan Wu [Wed, 23 Sep 2020 01:10:26 +0000 (18:10 -0700)]
[CGProfile] don't emit cgprofile entry if called function is dllimport

Differential Revision: https://reviews.llvm.org/D88127

4 years ago[clangd] Improve bad-RPC-payload error messages slightly
Sam McCall [Wed, 23 Sep 2020 23:51:29 +0000 (01:51 +0200)]
[clangd] Improve bad-RPC-payload error messages slightly

4 years ago[libc++] Fix spurious test failure in -fno-exceptions
Louis Dionne [Wed, 23 Sep 2020 23:44:03 +0000 (19:44 -0400)]
[libc++] Fix spurious test failure in -fno-exceptions

4 years agoFix LLDB tweak in 62a47e994fcf5b73e29547d26cd9676b30cb69a3
Sam McCall [Wed, 23 Sep 2020 23:30:42 +0000 (01:30 +0200)]
Fix LLDB tweak in 62a47e994fcf5b73e29547d26cd9676b30cb69a3

4 years ago[JSON] Add error reporting to fromJSON and ObjectMapper
Sam McCall [Wed, 23 Sep 2020 23:14:12 +0000 (01:14 +0200)]
[JSON] Add error reporting to fromJSON and ObjectMapper

Translating between JSON objects and C++ strutctures is common.
From experience in clangd, fromJSON/ObjectMapper work well and save a lot of
code, but aren't adopted elsewhere at least partly due to total lack of error
reporting beyond "ok"/"bad".

The recently-added error model should be rich enough for most applications.
It requires tracking the path within the root object and reporting local
errors at appropriate places.
To do this, we exploit the fact that the call graph of recursive
parse functions mirror the structure of the JSON itself.
The current path is represented as a linked list of segments, each of which is
on the stack as a parameter. Concretely, fromJSON now looks like:
  bool fromJSON(const Value&, T&, Path);

Beyond the signature change, this is reasonably unobtrusive: building
the path segments is mostly handled by ObjectMapper and the vector<T> fromJSON.
However the root caller of fromJSON must now create a Root object to
store the errors, which is a little clunky.

I've added high-level parse<T>(StringRef) -> Expected<T>, but it's not
general enough to be the primary interface I think (at least, not usable in
clangd).

All existing users (mostly just clangd) are updated in this patch,
making this change backwards-compatible is a bit hairy.

Differential Revision: https://reviews.llvm.org/D88103

4 years ago[libunwind] Optimize dl_iterate_phdr's findUnwindSectionsByPhdr
Ryan Prichard [Wed, 23 Sep 2020 21:25:23 +0000 (14:25 -0700)]
[libunwind] Optimize dl_iterate_phdr's findUnwindSectionsByPhdr

Currently, findUnwindSectionsByPhdr is slightly micro-optimized for the
case where the first callback has the target address, and is otherwise
very inefficient -- it decodes .eh_frame_hdr even when no PT_LOAD
matches the PC. (If the FrameHeaderCache is enabled, then the
micro-optimization only helps the first time unwind info is looked up.)

Instead, it makes more sense to optimize for the case where the
callback *doesn't* find the target address, so search for a PT_LOAD
segment first, and only look for the unwind info section if a matching
PT_LOAD is found.

This change helps on an Android benchmark with 100 shared objects,
where the DSO at the end of the dl_iterate_phdr list throws 10000
exceptions. Assuming the frame cache is disabled, this change cuts
about 30-40% off the benchmark's runtime.

Reviewed By: compnerd, saugustine, #libunwind

Differential Revision: https://reviews.llvm.org/D87881

4 years ago[libunwind] Combine dl_iterate_phdr codepaths for DWARF and EHABI
Ryan Prichard [Wed, 23 Sep 2020 21:25:13 +0000 (14:25 -0700)]
[libunwind] Combine dl_iterate_phdr codepaths for DWARF and EHABI

dl_iterate_phdr is used to search for unwind info provided by either
PT_GNU_EH_FRAME or PT_ARM_EXIDX. Most of the code between the two is
the same, so combine them, and factor out what's different into
checkForUnwindInfoSegment.

Details:
 - The FrameHeaderCache can now be enabled for ARM EHABI.
 - findUnwindSectionsByPhdr now finds the last PT_ARM_EXIDX rather than
   the first. There should only be one segment.
 - The dso_base and text_segment_length fields of UnwindInfoSections
   are now needed for dl_iterate_phdr when using EHABI, to hold the
   low and high PC values for a cache entry.

Reviewed By: compnerd, danielkiss, #libunwind, saugustine

Differential Revision: https://reviews.llvm.org/D87880

4 years ago[JSON] Display errors associated with Paths in context
Sam McCall [Wed, 23 Sep 2020 22:28:52 +0000 (00:28 +0200)]
[JSON] Display errors associated with Paths in context

When an error occurs processing a JSON object, seeing the actual
surrounding data helps. Dumping just the node where the problem
was identified can be too much or too little information.

printErrorContext() shows the error message in its context, as a comment.
JSON values along the path to the broken place are shown in some detail,
the rest of the document is elided. For example:

```
{
  "credentials": [
    {
      "username": /* error: expected string */ 42,
      "password": "secret"
    },
    { ... }
  ]
  "backups": { ... }
}
```

Differential Revision: https://reviews.llvm.org/D88103

4 years ago[NewPM] Add callbacks to PassBuilder to run before/after parsing a pass
Arthur Eubanks [Tue, 22 Sep 2020 16:34:46 +0000 (09:34 -0700)]
[NewPM] Add callbacks to PassBuilder to run before/after parsing a pass

This is in preparation for supporting -debugify-each, which adds a debug
info pass before and after each pass.

Switch VerifyEach to use this.

Reviewed By: ychen

Differential Revision: https://reviews.llvm.org/D88107

4 years ago[NewPM][CGSCC] Handle newly added functions in updateCGAndAnalysisManagerForPass
Arthur Eubanks [Wed, 16 Sep 2020 20:49:31 +0000 (13:49 -0700)]
[NewPM][CGSCC] Handle newly added functions in updateCGAndAnalysisManagerForPass

This seems to fit the CGSCC updates model better than calling
addNewFunctionInto{Ref,}SCC() on newly created/outlined functions.
Now addNewFunctionInto{Ref,}SCC() are no longer necessary.

However, this doesn't work on newly outlined functions that aren't
referenced by the original function. e.g. if a() was outlined into b()
and c(), but c() is only referenced by b() and not by a(), this will
trigger an assert.

This also fixes an issue I was seeing with newly created functions not
having passes run on them.

Ran check-llvm with expensive checks.

Reviewed By: asbirlea

Differential Revision: https://reviews.llvm.org/D87798

4 years ago[NewPM][MSSA] Fix failures under NPM due to -enable-mssa-loop-dependency
Arthur Eubanks [Wed, 23 Sep 2020 01:33:42 +0000 (18:33 -0700)]
[NewPM][MSSA] Fix failures under NPM due to -enable-mssa-loop-dependency

Reviewed By: asbirlea

Differential Revision: https://reviews.llvm.org/D88128

4 years ago[JSON] Facility to track position within an object and report errors.
Sam McCall [Wed, 23 Sep 2020 22:01:45 +0000 (00:01 +0200)]
[JSON] Facility to track position within an object and report errors.

This error model should be rich enough for most applications. It comprises:

- a name for the root object, so the user knows what we're parsing
- a path from the root object to the JSON node most associated with the error
- a local error message

This can be presented as an llvm::Error e.g.
  "expected string at ConfigFile.credentials[0].username"

It's designed to be cheap: Paths are a linked list of lightweight
objects on the stack. No heap allocations unless errors are encountered.

A subsequent commit will make use of this in the JSON-to-object
translation facilities: fromJSON and ObjectMapper.
However it's independent of these and can be used for e.g. validation alone.

Another subsequent commit will support showing the error in its context
within the parsed value.

Differential Revision: https://reviews.llvm.org/D88103

4 years ago[lld-macho] handle options -search_paths_first, -search_dylibs_first
Greg McGary [Sun, 20 Sep 2020 15:37:20 +0000 (08:37 -0700)]
[lld-macho] handle options -search_paths_first, -search_dylibs_first

Differential Revision: https://reviews.llvm.org/D88054

4 years ago[X86] Add a memory clobber to the bittest intrinsic inline asm. Get default clobbers...
Craig Topper [Wed, 23 Sep 2020 21:53:16 +0000 (14:53 -0700)]
[X86] Add a memory clobber to the bittest intrinsic inline asm. Get default clobbers from the target

I believe the inline asm emitted here should have a memory clobber since it writes to memory.

It was also missing the dirflag clobber that we use by default along with flags and fpsr. To avoid missing defaults in the future, get the default list from the target

Differential Revision: https://reviews.llvm.org/D88121

4 years ago[lld-macho] cleanup unimplemented-option warnings
Greg McGary [Wed, 23 Sep 2020 15:53:37 +0000 (08:53 -0700)]
[lld-macho] cleanup unimplemented-option warnings

Remove all spurious `HelpHidden` flags from  `lld/MachO/Options.td`. Add test for `HelpHidden` to `warnIfUnimplementedOption()` so that the empty `// handled elsewhere` case is unnecessary.

Reviewed By: #lld-macho, int3, smeenai

Differential Revision: https://reviews.llvm.org/D88160

4 years ago[JSON] Allow emitting comments in json::OStream
Sam McCall [Wed, 23 Sep 2020 21:33:19 +0000 (23:33 +0200)]
[JSON] Allow emitting comments in json::OStream

This isn't standard JSON, but is a popular extension.
It will be used to show errors in context, rendering pseudo-json for humans.

Differential Revision: https://reviews.llvm.org/D88103

4 years ago[SelectionDAG][GISel] Make LegalizeDAG lower FNEG using integer ops.
Eli Friedman [Wed, 23 Sep 2020 21:10:33 +0000 (14:10 -0700)]
[SelectionDAG][GISel] Make LegalizeDAG lower FNEG using integer ops.

Previously, if a floating-point type was legal, but FNEG wasn't legal,
we would use FSUB.  Instead, we should use integer ops, to preserve the
semantics.  (Alternatively, there's a compiler-rt call we could use, but
there isn't much reason to use that.)

It turns out we actually are still using this obscure codepath in a few
cases: on some targets, we have "legal" floating-point types that don't
actually support any floating-point operations.  In particular, ARM and
AArch64 are using this path.

The implementation for SelectionDAG is pretty simple because we can
reuse the infrastructure from FCOPYSIGN.

See also 9a3dc3e, the corresponding change to type legalization.

Also includes a "bonus" change to STRICT_FSUB legalization, so we can
lower a STRICT_FSUB to a float libcall.

Includes the changes to both LegalizeDAG and GlobalISel so we don't have
inconsistent results in the future.

Fixes https://bugs.llvm.org/show_bug.cgi?id=46792 .

Differential Revision: https://reviews.llvm.org/D84287

4 years ago[AArch64] Expand some vector of i64 reductions on NEON
Cameron McInally [Wed, 23 Sep 2020 20:55:06 +0000 (15:55 -0500)]
[AArch64] Expand some vector of i64 reductions on NEON

With the exception of VECREDUCE_ADD, there are no NEON instructions to support vector of i64 reductions. This patch removes the Custom lowerings for those and adds some test coverage to confirm.

Differential Revision: https://reviews.llvm.org/D88161

4 years agoRecommit [NFC] Refactor DiagnosticBuilder and PartialDiagnostic
Yaxun (Sam) Liu [Wed, 23 Sep 2020 20:16:00 +0000 (16:16 -0400)]
Recommit [NFC] Refactor DiagnosticBuilder and PartialDiagnostic

This recommits 829d14ee0a6aa79c89f7f3d9fcd9d27d3efd2b91.

The patch was reverted due to a regression in some CUDA app
which was thought to be caused by this patch. However, investigation
showed that the regression was due to some other issues, therefore
recommit this patch.

4 years ago[PowerPC] Implement the 128-bit vec_[all|any]_[eq | ne | lt | gt | le | ge] builtins...
Amy Kwan [Wed, 23 Sep 2020 20:46:54 +0000 (16:46 -0400)]
[PowerPC] Implement the 128-bit vec_[all|any]_[eq | ne | lt | gt | le | ge] builtins in Clang/LLVM

This patch implements the vec_[all|any]_[eq | ne | lt | gt | le | ge] builtins for vector signed/unsigned __int128.

Differential Revision: https://reviews.llvm.org/D87910

4 years ago[PowerPC] Implement Vector signed/unsigned __int128 overloads for the comparison...
Albion Fung [Wed, 23 Sep 2020 20:37:48 +0000 (16:37 -0400)]
[PowerPC] Implement Vector signed/unsigned __int128 overloads for the comparison builtins

This patch implements Vector signed/unsigned __int128 overloads for the comparison builtins.

Differential Revision: https://reviews.llvm.org/D87804

4 years agoClean up test file, NFC
Krzysztof Parzyszek [Wed, 23 Sep 2020 19:59:13 +0000 (14:59 -0500)]
Clean up test file, NFC

4 years ago[LoopTerminology][NFC] Fix formatting typo
Stefanos Baziotis [Wed, 23 Sep 2020 19:53:05 +0000 (22:53 +0300)]
[LoopTerminology][NFC] Fix formatting typo

4 years ago[lldb] Normalize paths in new test
Joseph Tremoulet [Wed, 23 Sep 2020 16:20:10 +0000 (09:20 -0700)]
[lldb] Normalize paths in new test

The minidump-sysroot test I added in commit 20f84257 compares two paths
using a string comparison.  This causes the Windows buildbot to fail
because of mismatched forward slashes and backslashes.  Use
os.path.normcase to normalize before comparing.

4 years agoAllow init_priority values <= 100 and > 65535 within system headers.
Aaron Ballman [Wed, 23 Sep 2020 19:24:52 +0000 (15:24 -0400)]
Allow init_priority values <= 100 and > 65535 within system headers.

This also adds some bare-bones documentation for the attribute rather
than leaving it undocumented.

4 years ago[AArch64][SVE] Add lowering for llvm frecpx
Muhammad Asif Manzoor [Wed, 23 Sep 2020 19:22:13 +0000 (15:22 -0400)]
[AArch64][SVE] Add lowering for llvm frecpx

Add the functionality to lower frecpx for passthru variant

Reviewed By: paulwalker-arm

Differential Revision: https://reviews.llvm.org/D88032

4 years agoRevert "[lsan] On Fuchsia, don't use atexit hook for leak checks"
Nikita Popov [Wed, 23 Sep 2020 19:08:15 +0000 (21:08 +0200)]
Revert "[lsan] On Fuchsia, don't use atexit hook for leak checks"

This reverts commit 0caad9fe441d5ee562e96d8b30b5574b492a933a.
This reverts commit c96d0cceb684fa176b51d7df5f4f8370e2c983f4.

Causes linker errors which were not fixed by the subsequent commit
either:

/home/nikic/llvm-project/compiler-rt/lib/asan/asan_rtl.cpp:503: error: undefined reference to '__asan::InstallAtExitCheckLeaks()'

4 years ago[scudo][standalone] Fix tests under ASan/UBSan
Kostya Kortchinsky [Wed, 23 Sep 2020 17:22:15 +0000 (10:22 -0700)]
[scudo][standalone] Fix tests under ASan/UBSan

Fix a potential UB in `appendSignedDecimal` (with -INT64_MIN) by making
it a special case.

Fix the terrible test cases for `isOwned`: I was pretty sloppy on those
and used some stack & static variables, but since `isOwned` accesses
memory prior to the pointer to check for the validity of the Scudo
header, it ended up being detected as some global and stack buffer out
of bounds accesses. So not I am using buffers with enough room so that
the test will not access memory prior to the variables.

With those fixes, the tests pass on the ASan+UBSan Fuchsia build.

Thanks to Roland for pointing those out!

Differential Revision: https://reviews.llvm.org/D88170

4 years agoasan: Use `#if` to test CAN_SANITIZE_LEAKS
Roland McGrath [Wed, 23 Sep 2020 18:41:09 +0000 (11:41 -0700)]
asan: Use `#if` to test CAN_SANITIZE_LEAKS

The `if (0)` isn't necessarily optimized out so as not to create
a link-time reference to LSan runtime functions that might not
exist.  So use explicit conditional compilation instead.

Reviewed By: phosek

Differential Revision: https://reviews.llvm.org/D88173

4 years ago[PowerPC][PCRelative] Thread Local Storage Support for Local Dynamic
Victor Huang [Wed, 23 Sep 2020 18:43:33 +0000 (13:43 -0500)]
[PowerPC][PCRelative] Thread Local Storage Support for Local Dynamic

This patch is the initial support for the Local Dynamic Thread Local Storage
model to produce code sequence and relocation correct to the ABI for the model
when using PC relative memory operations.

Differential Revision: https://reviews.llvm.org/D87721

4 years ago[IRSim] Adding IRSimilarityCandidate that contains a region of IRInstructionData.
Andrew Litteken [Tue, 15 Sep 2020 22:30:31 +0000 (17:30 -0500)]
[IRSim] Adding IRSimilarityCandidate that contains a region of IRInstructionData.

The IRSimilarityCandidate is a container to hold a region of
IRInstructions and offer interfaces for the starting instruction, ending
instruction, parent function, length.  It also assigns a global value
number for each unique instance of a value in the region.

It also contains an interface to compare two IRSimilarity as to whether
they have the same sequence of similar instructions.

Tests for whether the instructions are similar are found in
unittests/Analysis/IRSimilarityIdentifierTest.cpp.

Recommit of: 4944bb190fed8861d4d043eaf45e3c1e12aa2dc5

Differential Revision: https://reviews.llvm.org/D86970

4 years ago[AMDGPU] Make ds fp atomics overloadable
Stanislav Mekhanoshin [Fri, 18 Sep 2020 20:20:00 +0000 (13:20 -0700)]
[AMDGPU] Make ds fp atomics overloadable

Differential Revision: https://reviews.llvm.org/D87947

4 years agoAdd `breakpoint delete --disabled`: deletes all disabled breakpoints.
Jim Ingham [Wed, 23 Sep 2020 01:05:46 +0000 (18:05 -0700)]
Add `breakpoint delete --disabled`: deletes all disabled breakpoints.

Differential Revision: https://reviews.llvm.org/D88129

4 years ago[AArch64][SVE] Fix frame offset calculation when d8 is saved.
Eli Friedman [Tue, 22 Sep 2020 22:05:13 +0000 (15:05 -0700)]
[AArch64][SVE] Fix frame offset calculation when d8 is saved.

If d8 is saved, the fp is not actually adjacent to the SVE
spills/allocations.  Fix the offset calculation to account for this.

Differential Revision: https://reviews.llvm.org/D88117

4 years ago[mlir][OpFormatGen] Update "custom" directives for attributes.
Mike Urbach [Wed, 23 Sep 2020 18:01:39 +0000 (18:01 +0000)]
[mlir][OpFormatGen] Update "custom" directives for attributes.

This tweaks the generated code for parsing attributes with a custom
directive to call `addAttribute` on the `OperationState` directly,
and adds a newline after this call. Previously, the generated code
would call `addAttribute` on the `OperationState` field `attributes`,
which has no such method and fails to compile. Furthermore, the lack
of newline would generate code with incorrectly formatted single line
`if` statements. Added tests for parsing and printing attributes with
a custom directive.

Reviewed By: mehdi_amini

Differential Revision: https://reviews.llvm.org/D87860

4 years ago[gn build] Allow option to build with asan/tsan/ubsan
Arthur Eubanks [Mon, 21 Sep 2020 23:28:21 +0000 (16:28 -0700)]
[gn build] Allow option to build with asan/tsan/ubsan

Reviewed By: thakis

Differential Revision: https://reviews.llvm.org/D88056

4 years ago[lsan] On Fuchsia, don't use atexit hook for leak checks
Roland McGrath [Wed, 23 Sep 2020 01:02:56 +0000 (18:02 -0700)]
[lsan] On Fuchsia, don't use atexit hook for leak checks

Fuchsia's system libraries are instrumented and use the lsan
allocator for internal purposes.  So leak checking needs to run
after all atexit hooks and after the system libraries' internal
exit-time hooks.  The <zircon/sanitizer.h> hook API calls the
__sanitizer_process_exit_hook function at exactly the right time.

Reviewed By: vitalybuka, phosek

Differential Revision: https://reviews.llvm.org/D86171

4 years agoDocument the `--verbatim` flag from arc to update the description for a phabricator...
Mehdi Amini [Wed, 23 Sep 2020 18:00:37 +0000 (18:00 +0000)]
Document the `--verbatim` flag from arc to update the description for a phabricator revision

4 years agoUpdate Phabricator doc to remove the warning on "arc land": tags a properly handled...
Mehdi Amini [Wed, 23 Sep 2020 17:57:27 +0000 (17:57 +0000)]
Update Phabricator doc to remove the warning on "arc land": tags a properly handled server side now

4 years agoRevert "[ms] [llvm-ml] Add support for .radix directive, and accept all radix specifiers"
Eric Astor [Wed, 23 Sep 2020 17:57:43 +0000 (13:57 -0400)]
Revert "[ms] [llvm-ml] Add support for .radix directive, and accept all radix specifiers"

This reverts commit 5dd1b6d612655c9006ba97a8b6487ded80719b48.

4 years ago[SLP] Make HorizontalReduction::getOperationData take an Instruction* instead of...
Craig Topper [Wed, 23 Sep 2020 17:25:00 +0000 (10:25 -0700)]
[SLP] Make HorizontalReduction::getOperationData take an Instruction* instead of a Value*. NFCI

All of the callers already have an Instruction *. Many of them
from a dyn_cast.

Also update the OperationData constructor to use a Instruction&
to remove a dyn_cast and make it clear that the pointer is non-null.

Differential Revision: https://reviews.llvm.org/D88132

4 years ago[X86] Improve demanded bits for X86ISD::BEXTR.
Craig Topper [Wed, 23 Sep 2020 17:20:03 +0000 (10:20 -0700)]
[X86] Improve demanded bits for X86ISD::BEXTR.

If the control is constant we can figure out exactly which bits
of the input are demanded.

Differential Revision: https://reviews.llvm.org/D88072

4 years agoFix include location (accidentally committed a local variation)
Eric Astor [Wed, 23 Sep 2020 17:49:56 +0000 (13:49 -0400)]
Fix include location (accidentally committed a local variation)

4 years ago[TTI] add wrapper for matching vector reduction to reduce code duplication; NFC
Sanjay Patel [Wed, 23 Sep 2020 17:46:26 +0000 (13:46 -0400)]
[TTI] add wrapper for matching vector reduction to reduce code duplication; NFC

I'm not sure what this means, but the order in which we try
the matches makes a difference on at least 1 regression test...

4 years ago[ms] [llvm-ml] Add support for .radix directive, and accept all radix specifiers
Eric Astor [Wed, 23 Sep 2020 17:42:22 +0000 (13:42 -0400)]
[ms] [llvm-ml] Add support for .radix directive, and accept all radix specifiers

Add support for .radix directive, and radix specifiers [yY] (binary), [oOqQ] (octal), and [tT] (decimal).

Also, when lexing MASM integers, require radix specifier; MASM requires that all literals without a radix specifier be treated as in the default radix. (e.g., 0100 = 100)

Reviewed By: thakis

Differential Revision: https://reviews.llvm.org/D87400

4 years agoEnhance TableGen so that backends can produce better error messages.
Paul C. Anagnostopoulos [Thu, 3 Sep 2020 13:41:09 +0000 (09:41 -0400)]
Enhance TableGen so that backends can produce better error messages.

Modify SearchableTableEmitter.cpp to take advantage.
Clean up formatting and capitalization issues.

4 years agoRe-apply https://reviews.llvm.org/D87921, was reverted to triage a PPC bot failure.
Sriraman Tallam [Wed, 23 Sep 2020 16:26:41 +0000 (09:26 -0700)]
Re-apply https://reviews.llvm.org/D87921, was reverted to triage a PPC bot failure.

D87921 was reverted in commit b89059a31347dd09b55a96b99b3dbe38d7749908
as it was causing an unknown llvm PPC bot failure.  Reapplying the patch
after confirming that this is not responsible. Build bot failure:
https://reviews.llvm.org/D87921#2286644  which caused the revert.

The wrong placement of add pass with optimizations led to
-funique-internal-linkage-names being disabled.

Fixed the placement of the MPM.addpass for UniqueInternalLinkageNames to make it
work correctly with -O2 and new pass manager. Updated the tests to explicitly
check O0 and O1.

Differential Revision: https://reviews.llvm.org/D87921

4 years ago[Driver] Check whether Gentoo-specific configuration directory exists
Dmitry Antipov [Wed, 23 Sep 2020 17:21:18 +0000 (20:21 +0300)]
[Driver] Check whether Gentoo-specific configuration directory exists

Check whether /etc/env.d/gcc exists before trying to read from any
file from there. This saves a few OS calls on a non-Gentoo system.

Differential Revision: https://reviews.llvm.org/D87143

4 years agoBreak long line accidentally left in the previous commit
Krzysztof Parzyszek [Wed, 23 Sep 2020 17:23:54 +0000 (12:23 -0500)]
Break long line accidentally left in the previous commit

4 years agoRevert D87970 "[ThinLTO] Avoid temporaries when loading global decl attachment metadata"
Fangrui Song [Wed, 23 Sep 2020 17:23:48 +0000 (10:23 -0700)]
Revert D87970 "[ThinLTO] Avoid temporaries when loading global decl attachment metadata"

This reverts commit ab1b4810b55279bcf6fdd87be74a403440be3991.

It caused an issue in llvm::lto::thinBackend for a -fsanitize=cfi build.

```
AbbrevNo is 0 => "Invalid abbrev number"
0  llvm::BitstreamCursor::getAbbrev (this=0x9db4c8, AbbrevID=4) at llvm/include/llvm/Bitstream/BitstreamReader.h:528
1  0x00007f5f777a6eb4 in llvm::BitstreamCursor::readRecord (this=0x9db4c8, AbbrevID=4, Vals=llvm::SmallVector of Size 0, Capacity 64, Blob=0x7ffcd0e26558) at
usr/local/google/home/maskray/llvm/llvm/lib/Bitstream/Reader/BitstreamReader.cpp:228
2  0x00007f5f796bf633 in llvm::MetadataLoader::MetadataLoaderImpl::lazyLoadOneMetadata (this=0x9db3a0, ID=188, Placeholders=...) at /usr/local/google/home/mas
ray/llvm/llvm/lib/Bitcode/Reader/MetadataLoader.cpp:1091
3  0x00007f5f796c2527 in llvm::MetadataLoader::MetadataLoaderImpl::getMetadataFwdRefOrLoad (this=0x9db3a0, ID=188) at llvm
lib/Bitcode/Reader/MetadataLoader.cpp:668
4  0x00007f5f796bfff3 in llvm::MetadataLoader::getMetadataFwdRefOrLoad (this=0xd31580, Idx=188) at llvm/lib/Bitcode/Reader
MetadataLoader.cpp:2290
5  0x00007f5f79638265 in (anonymous namespace)::BitcodeReader::parseFunctionBody (this=0xd312e0, F=0x9de758) at llvm/lib/B
tcode/Reader/BitcodeReader.cpp:3938
6  0x00007f5f79635d32 in (anonymous namespace)::BitcodeReader::materialize (this=0xd312e0, GV=0x9de758) at llvm/lib/Bitcod
/Reader/BitcodeReader.cpp:5408
7  0x00007f5f7f8dbe3e in llvm::Module::materialize (this=0x9b92c0, GV=0x9de758) at llvm/lib/IR/Module.cpp:442
8  0x00007f5f7f7f8fbe in llvm::GlobalValue::materialize (this=0x9de758) at llvm/lib/IR/Globals.cpp:50
9  0x00007f5f83b9b5f5 in llvm::FunctionImporter::importFunctions (this=0x7ffcd0e2a730, DestModule=..., ImportList=...) at
llvm/lib/Transforms/IPO/FunctionImport.cpp:1182
```

4 years ago[EarlyCSE] Fix crash with expensive checks after D87691
Krzysztof Parzyszek [Wed, 23 Sep 2020 15:48:34 +0000 (10:48 -0500)]
[EarlyCSE] Fix crash with expensive checks after D87691

D87691 reordered some checks, which turned out to be unsafe. More
specifically, when examining a store instruction, the check against
getOrCreateResult should be done before attempting to call
isSameMemGeneration. Otherwise a crash in MSSA walker can occur.

This patch restores the order of these calls to what it was originally.

4 years agoAdd REQUIRES to embed-bitcode-noopt.ll
Mircea Trofin [Wed, 23 Sep 2020 17:12:16 +0000 (10:12 -0700)]
Add REQUIRES to embed-bitcode-noopt.ll

4 years ago[Support/Path] Add path::is_absolute_gnu
Vinicius Tinti [Wed, 23 Sep 2020 15:28:19 +0000 (16:28 +0100)]
[Support/Path] Add path::is_absolute_gnu

Implements IS_ABSOLUTE_PATH from GNU tools.

C++17 is_absolute behavior is different the from the behavior defined by GNU
tools.

According to cppreference.com, C++17 states: "An absolute path is a path
that unambiguously identifies the location of a file without reference
to an additional starting location."

In other words, the rules are:
 1. POSIX style paths with nonempty root directory are absolute.
 2. Windows style paths with nonempty root name and root directory are
    absolute.
 3. No other paths are absolute.

GNU rules are:
 1. Paths starting with a path separator are absolute.
 2. Windows style paths are also absolute if they start with a character
    followed by ':'.
 3. No other paths are absolute.

On Windows style the path "C:\Users\Default" has "C:" as root name and "\"
as root directory.

Hence "C:" on Windows is absolute under GNU rules and not absolute under
C++17 because it has no root directory. Likewise "/" and "\" on Windows are
absolute under GNU and are not absolute under C++17 due to empty root name.

Related to PR46368.

Differential Revision: https://reviews.llvm.org/D87667

4 years agoAdd optimal thread strategy
Dmitry Antipov [Wed, 23 Sep 2020 16:37:50 +0000 (19:37 +0300)]
Add optimal thread strategy

Add an optimal thread strategy to execute specified amount of tasks.
This strategy should prevent us from creating too many threads if we
occasionaly have an unexpectedly small amount of tasks.

Differential Revision: https://reviews.llvm.org/D87765

4 years ago[clang]Test ensuring -fembed-bitcode passed to cc1 captures pre-opt bitcode.
Mircea Trofin [Tue, 22 Sep 2020 20:33:34 +0000 (13:33 -0700)]
[clang]Test ensuring -fembed-bitcode passed to cc1 captures pre-opt bitcode.

This is important to not regress because it allows us to capture pre-optimization
bitcode and options, and replay the full optimization pipeline.

Differential Revision: https://reviews.llvm.org/D88114

4 years ago[MBFIWrapper] Add a new function getBlockProfileCount
Guozhi Wei [Wed, 23 Sep 2020 16:27:34 +0000 (09:27 -0700)]
[MBFIWrapper] Add a new function getBlockProfileCount

MBFIWrapper keeps track of block frequencies of newly created blocks and
modified blocks, modified block frequencies should also impact block profile
count. This class doesn't provide interface getBlockProfileCount, users can only
use the underlying MBFI to query profile count, the underlying MBFI doesn't know
the modifications made in MBFIWrapper, so it either provides stale profile count
for modified block or simply crashes on new blocks.

So this patch add function getBlockProfileCount to class MBFIWrapper to handle
new blocks or modified blocks.

Differential Revision: https://reviews.llvm.org/D87802

4 years agoNFC: Remove dangling dep on MLIRStandardOps.
Stella Laurenzo [Wed, 23 Sep 2020 16:30:13 +0000 (09:30 -0700)]
NFC: Remove dangling dep on MLIRStandardOps.

Was inadvertently left out of D88155.

4 years ago[UpdateTestChecks] Remove bug-exposing test
David Greene [Wed, 23 Sep 2020 16:07:33 +0000 (11:07 -0500)]
[UpdateTestChecks] Remove bug-exposing test

Remove RISCV codegen tests for --include-generated-funcs because apparently
MachineOutliner has a bug on that target that is exposed by expensive-checks.

4 years agoRemove MLIR C-API explicit registration of standard ops.
Stella Laurenzo [Wed, 23 Sep 2020 15:13:30 +0000 (08:13 -0700)]
Remove MLIR C-API explicit registration of standard ops.

* Added mlirRegisterAllDialects() to the python API until a more complete registration design emerges for it.

Differential Revision: https://reviews.llvm.org/D88155

4 years ago[MLIR][NFC] Adopt use of BlockRange in place of ArrayRef<Block *>
Rahul Joshi [Wed, 23 Sep 2020 04:06:25 +0000 (21:06 -0700)]
[MLIR][NFC] Adopt use of BlockRange in place of ArrayRef<Block *>

- Use BlockRange in ODS generated builders as well as other places throughout the code

Differential Revision: https://reviews.llvm.org/D87955

4 years agoUpdate the documentation for the MLIR Dialect class (NFC)
Mehdi Amini [Wed, 23 Sep 2020 16:09:05 +0000 (16:09 +0000)]
Update the documentation for the MLIR Dialect class (NFC)

4 years agoImprove dynamic AST matching diagnostics for conversion errors
Aaron Ballman [Wed, 23 Sep 2020 16:11:07 +0000 (12:11 -0400)]
Improve dynamic AST matching diagnostics for conversion errors

Currently, when marshaling a dynamic AST matchers, we check for the type
and value validity of matcher arguments at the same time for some matchers.
For instance, when marshaling hasAttr("foo"), the argument is first type
checked to ensure it's a string and then checked to see if that string can
locate an attribute with that name. Similar happens for other enumeration
conversions like cast kinds or unary operator kinds. If the type is
correct but the value cannot be looked up, we make a best-effort attempt
to find a nearby name that the user might have meant, but if one cannot
be found, we throw our hands up and claim the types don't match.

This has an unfortunate behavior that when the user enters something of
the correct type but a best guess cannot be located, you get confusing
error messages like:
Incorrect type for arg 1. (Expected = string) != (Actual = String).

This patch splits the argument check into two parts: if the types don't
match, give a type diagnostic. If the type matches but the value cannot
be converted, give a best guess diagnostic or a value could not be
located diagnostic. This addresses PR47057.

4 years ago[MLIR][NFC] Adopt use of TypeRange in build() methods.
Rahul Joshi [Wed, 23 Sep 2020 04:00:11 +0000 (21:00 -0700)]
[MLIR][NFC] Adopt use of TypeRange in build() methods.

- Use TypeRange instead of ArrayRef<Type> where possible.
- Change some of the custom builders to also use TypeRange

Differential Revision: https://reviews.llvm.org/D87944

4 years ago[MLIR] Change default builders generated by TableGen to use TypeRange for result...
Rahul Joshi [Wed, 23 Sep 2020 03:58:30 +0000 (20:58 -0700)]
[MLIR] Change default builders generated by TableGen to use TypeRange for result types

- Change the default builders to use TypeRange instead of ArrayRef<Type>
- Custom builders defined in LinalgStructuredOps now conflict with the default
  separate param ones, but the default collective params one is still needed. Resolve
  this by replicating the collective param builder as a custom builder and skipping
  the generation of default builders for these ops.

Differential Revision: https://reviews.llvm.org/D87926

4 years agoFix regressioin in test dwp-separate-debug-file.cpp
Yaxun (Sam) Liu [Wed, 23 Sep 2020 15:47:21 +0000 (11:47 -0400)]
Fix regressioin in test dwp-separate-debug-file.cpp

4 years ago[mlir] Add insert before/after to list-like constructs in C API
Alex Zinenko [Wed, 23 Sep 2020 13:02:47 +0000 (15:02 +0200)]
[mlir] Add insert before/after to list-like constructs in C API

Blocks in a region and operations in a block are organized in a linked list.
The C API only provides functions to append or to insert elements at the
specified numeric position in the list. The latter is expensive since it
requires to traverse the list. Add insert before/after functionality with low
cost that relies on the iplist elements being convertible to iterators.

Reviewed By: stellaraccident

Differential Revision: https://reviews.llvm.org/D88148

4 years ago[lldb] Remove lldb-perf remenant
Dave Lee [Tue, 22 Sep 2020 23:08:48 +0000 (16:08 -0700)]
[lldb] Remove lldb-perf remenant

Delete a file remaining from the deletion of lldb-perf in D64362.

Differential Revision: https://reviews.llvm.org/D88119

4 years agorecommit [HIP] Fix -gsplit-dwarf option
Yaxun (Sam) Liu [Mon, 14 Sep 2020 18:21:30 +0000 (14:21 -0400)]
recommit [HIP] Fix -gsplit-dwarf option

recommit e50465ecefc964e5700df26fc7e02a673eed085a with fix for
regression in lldb tests.

Two issues:

1. the directory part of original .dwo file was dropped
2. if the stem of the .dwo file contains '.', the last dot
and strings after that were removed

This recommit fixes those two issues.

4 years ago[AArch64] Fix ldst optimization of non-immediate store offset
Andrew Wei [Wed, 23 Sep 2020 14:44:30 +0000 (22:44 +0800)]
[AArch64] Fix ldst optimization of non-immediate store offset

When matching store instruction for ldst opt, we should make sure store instr is in 'reg+imm' form as load instr,
otherwise, it will have assertion in isLdOffsetInRangeOfSt since it will use getImm() directly.

Reviewed By: efriedma

Differential Revision: https://reviews.llvm.org/D87905

4 years agoAdd missing namespace closure comments. NFCI.
Simon Pilgrim [Wed, 23 Sep 2020 14:54:50 +0000 (15:54 +0100)]
Add missing namespace closure comments. NFCI.

Fixes some clang-tidy llvm-namespace-comment warnings.

4 years agoAdd missing namespace closure comment. NFCI.
Simon Pilgrim [Wed, 23 Sep 2020 14:53:00 +0000 (15:53 +0100)]
Add missing namespace closure comment. NFCI.

Fixes clang-tidy llvm-namespace-comment warning.

4 years agoRevert "[AMDGPU] Insert waitcnt after returning from call"
Sebastian Neubauer [Wed, 23 Sep 2020 15:16:39 +0000 (17:16 +0200)]
Revert "[AMDGPU] Insert waitcnt after returning from call"

This reverts commit ca907bfb57d8ad3ec3bcc2cff2abab7b1b933af6.

According to michel.daenzer,
> This completely broke the Mesa radeonsi driver on Navi 14. Xorg +
> xterm come up with major corruption & psychedelic colours.

4 years ago[mlir] Remove unneeded OpBuilder params. NFC.
Jacques Pienaar [Wed, 23 Sep 2020 15:10:16 +0000 (08:10 -0700)]
[mlir] Remove unneeded OpBuilder params. NFC.

These are now automatically prepended.

4 years ago[mlir][ods] Custom builder with no params
Jacques Pienaar [Wed, 23 Sep 2020 14:50:45 +0000 (07:50 -0700)]
[mlir][ods] Custom builder with no params

Incorrect generation of custom build method without any params.

4 years agoNFC: Remove unused variable.
Stella Laurenzo [Wed, 23 Sep 2020 15:00:31 +0000 (08:00 -0700)]
NFC: Remove unused variable.

4 years agoAdd mlir python APIs for creating operations, regions and blocks.
Stella Laurenzo [Mon, 21 Sep 2020 04:25:46 +0000 (21:25 -0700)]
Add mlir python APIs for creating operations, regions and blocks.

* The API is a bit more verbose than I feel like it needs to be. In a follow-up I'd like to abbreviate some things and look in to creating aliases for common accessors.
* There is a lingering lifetime hazard between the module and newly added operations. We have the facilities now to solve for this but I will do that in a follow-up.
* We may need to craft a more limited API for safely referencing successors when creating operations. We need more facilities to really prove that out and should defer for now.

Differential Revision: https://reviews.llvm.org/D87996

4 years agoImplement python iteration over the operation/region/block hierarchy.
Stella Laurenzo [Sun, 20 Sep 2020 05:02:32 +0000 (22:02 -0700)]
Implement python iteration over the operation/region/block hierarchy.

* Removes the half-completed prior attempt at region/block mutation in favor of new approach to ownership.
* Will re-add mutation more correctly in a follow-on.
* Eliminates the detached state on blocks and regions, simplifying the ownership hierarchy.
* Adds both iterator and index based access at each level.

Differential Revision: https://reviews.llvm.org/D87982