platform/upstream/llvm.git
4 years ago[gcov] Temporarily unsupport host-byteorder-big-endian
Fangrui Song [Sun, 10 May 2020 17:26:10 +0000 (10:26 -0700)]
[gcov] Temporarily unsupport host-byteorder-big-endian

llvm-cov gcov does not support host-byteorder-big-endian yet.

4 years ago[AssumeBundles] Prevent generation of some redundant assumes
Tyker [Sun, 10 May 2020 16:20:34 +0000 (18:20 +0200)]
[AssumeBundles] Prevent generation of some redundant assumes

Summary: with this patch the assume salvageKnowledge will not generate assume if all knowledge is already available in an assume with valid context. assume bulider can also in some cases update an existing assume with better information.

Reviewers: jdoerfert

Reviewed By: jdoerfert

Subscribers: hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D78014

4 years ago[gcov] Don't skip leading zeros when reading a string
Fangrui Song [Sun, 10 May 2020 17:12:26 +0000 (10:12 -0700)]
[gcov] Don't skip leading zeros when reading a string

Even a 2003 version of gcov_read_string does not have the behavior
described by rL204881. The wrong impression might come from
libclang_rt.profile (GCDAProfiling.c)'s corrupted output (since the
initial import rL144865). Note, the corrupted output crashes gcov.

4 years ago[gcov] Fix .gcda decoding and support GCC 8, 9 and 10
Fangrui Song [Sat, 9 May 2020 17:36:34 +0000 (10:36 -0700)]
[gcov] Fix .gcda decoding and support GCC 8, 9 and 10

GCDAProfiling.c unnecessarily writes function names to .gcda files.
GCC 4.2 gcc/libgcov.c (now renamed to libgcc/libgcov*) did not write function
names. gcov-7 (compatible) crashes on .gcda produced by libclang_rt.profile
rL176173 realized the problem and introduced a mode to remove function
names.

llvm-cov code apparently takes GCDAProfiling.c output format as truth
and tries to decode function names.  Additionally, llvm-cov tries to
decode tags in certain order which does not match libgcov emitted .gcda
files.

This patch fixes the .gcda decoder and makes it work with GCC 8 and 9
(10 is compatible with 9). Note, line statistics are broken and not
fixed by this patch.

Add test/tools/llvm-cov/gcov-{4.7,8,9}.c to test compatibility.

4 years ago[LAA] Move runtime-check generation to Transforms/Utils/loopUtils (NFC)
Florian Hahn [Sun, 10 May 2020 15:48:06 +0000 (16:48 +0100)]
[LAA] Move runtime-check generation to Transforms/Utils/loopUtils (NFC)

Currently LAA's uses of ScalarEvolutionExpander blocks moving the
expander from Analysis to Transforms. Conceptually the expander does not
fit into Analysis (it is only used for code generation) and
runtime-check generation also seems to be better suited as a
transformation utility.

Reviewers: Ayal, anemet

Reviewed By: Ayal

Differential Revision: https://reviews.llvm.org/D78460

4 years agoCodeMetrics.cpp - remove unused includes. NFC.
Simon Pilgrim [Sun, 10 May 2020 15:59:39 +0000 (16:59 +0100)]
CodeMetrics.cpp - remove unused includes. NFC.

4 years agoLoopUtils.h - remove unused forward declarations. NFC.
Simon Pilgrim [Sun, 10 May 2020 15:58:34 +0000 (16:58 +0100)]
LoopUtils.h - remove unused forward declarations. NFC.

4 years agoLocal.h - remove unused forward declaration. NFC.
Simon Pilgrim [Sun, 10 May 2020 15:57:46 +0000 (16:57 +0100)]
Local.h - remove unused forward declaration. NFC.

4 years ago[InstCombine] canonicalize bitcast after insertelement into undef
Sanjay Patel [Sun, 10 May 2020 15:37:47 +0000 (11:37 -0400)]
[InstCombine] canonicalize bitcast after insertelement into undef

We have a transform in the opposite direction only for the x86 MMX type,
Other types are not handled either way before this patch.

The motivating case from PR45748:
https://bugs.llvm.org/show_bug.cgi?id=45748
...is the last test diff. In that example, we are triggering an existing
bitcast transform, so we reduce the number of casts, and that should give
us the ideal x86 codegen.

Differential Revision: https://reviews.llvm.org/D79171

4 years ago[CodeGen] fix test to be (mostly) independent of LLVM optimizer; NFC
Sanjay Patel [Sun, 10 May 2020 15:25:37 +0000 (11:25 -0400)]
[CodeGen] fix test to be (mostly) independent of LLVM optimizer; NFC

This test would break with the proposed change to IR canonicalization
in D79171.

The test tried to do the right thing by only using -mem2reg with opt,
but it was using -O3 before that step, so the opt part was meaningless.

4 years ago[CodeGen] fix test to be (mostly) independent of LLVM optimizer; NFC
Sanjay Patel [Sun, 10 May 2020 15:19:43 +0000 (11:19 -0400)]
[CodeGen] fix test to be (mostly) independent of LLVM optimizer; NFC

This test would break with the proposed change to IR canonicalization
in D79171. The raw unoptimized IR from clang is massive, so I've
replaced -instcombine with -mem2reg to make it more manageable,
but still be unlikely to break with unrelated changed to optimization.

4 years ago[InstCombine] matchOrConcat - match BITREVERSE
Simon Pilgrim [Sun, 10 May 2020 15:00:13 +0000 (16:00 +0100)]
[InstCombine] matchOrConcat - match BITREVERSE

Fold or(zext(bitreverse(x)),shl(zext(bitreverse(y)),bw/2) -> bitreverse(or(zext(x),shl(zext(y),bw/2))

Practically this is the same as the BSWAP pattern so we might as well handle it.

4 years agoRecommit "[LAA] Remove one addRuntimeChecks function (NFC)."
Florian Hahn [Sun, 10 May 2020 14:04:49 +0000 (15:04 +0100)]
Recommit "[LAA] Remove one addRuntimeChecks function (NFC)."

The failing assertion has been fixed and the problematic test case has
been added.

This reverts the revert commit fc44617f28847417e55836193bbe8e9c3f09eca9.

4 years agoPassInstrumentation.h - reduce TypeName.h include to StringRef forward declaration...
Simon Pilgrim [Sun, 10 May 2020 13:22:02 +0000 (14:22 +0100)]
PassInstrumentation.h - reduce TypeName.h include to StringRef forward declaration. NFC.

4 years agoCoverageMapping.h - remove unused StringSet.h include. NFC.
Simon Pilgrim [Sun, 10 May 2020 13:19:54 +0000 (14:19 +0100)]
CoverageMapping.h - remove unused StringSet.h include. NFC.

4 years agoPassManager.h - remove unused raw_ostream.h include. NFC.
Simon Pilgrim [Sun, 10 May 2020 13:17:52 +0000 (14:17 +0100)]
PassManager.h - remove unused raw_ostream.h include. NFC.

4 years agoMangler.h - reduce GlobalValue.h include to forward declaration. NFC.
Simon Pilgrim [Sun, 10 May 2020 13:03:13 +0000 (14:03 +0100)]
Mangler.h - reduce GlobalValue.h include to forward declaration. NFC.

Exposes implicit dependency in ModuleSymbolTable.h - added missing Module forward declaration.

4 years agoLoopPass.h - remove unnecessary PMStack forward declaration. NFC.
Simon Pilgrim [Sun, 10 May 2020 13:01:26 +0000 (14:01 +0100)]
LoopPass.h - remove unnecessary PMStack forward declaration. NFC.

We have to include LegacyPassManagers.h where its defined.

4 years agoRevert "[LAA] Remove one addRuntimeChecks function (NFC)."
Florian Hahn [Sun, 10 May 2020 12:27:10 +0000 (13:27 +0100)]
Revert "[LAA] Remove one addRuntimeChecks function (NFC)."

This reverts commit c28114c8ffde705d7e16cd4c065fd23269661c81.

This causes some bots to fail:

http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-android/builds/30596/steps/build%20android%2Faarch64/logs/stdio

4 years ago[mlir] Affine dim and symbol checkers: support detached regions
Alex Zinenko [Sun, 10 May 2020 12:14:10 +0000 (14:14 +0200)]
[mlir] Affine dim and symbol checkers: support detached regions

Functions checking whether an SSA value is a valid dimension or symbol for
affine operations can be called on values defined in a detached region (a
region that is not yet attached to an operation), for example, during parsing
or operation construction. These functions will attempt to uncondtionally
dereference a pointer to the parent operation of a region, which may be null
(as fixed by the previous commit, uninitialized before that). Since one cannot
know to which operation a region will be attached, conservatively this
operation would not be a valid affine scope and act accordingly, instead of
crashing.

4 years ago[mlir] Initialize the Region::container field to nullptr by default
Alex Zinenko [Sun, 10 May 2020 12:14:04 +0000 (14:14 +0200)]
[mlir] Initialize the Region::container field to nullptr by default

Region has a default constructor that is called when a region is constructed
while an operation is being created, and therefore before the region can be
attached to this operation. The `container` field is uninitialized, which makes
it impossible to check programmatically if a Region is attached to an operation
or not, leading to sly memory errors when this field is read. Initialize it to
nullptr by default and thus make sure one can check if a region is attached to
an operation or not.

4 years ago[LAA] Remove one addRuntimeChecks function (NFC).
Florian Hahn [Sun, 10 May 2020 11:26:33 +0000 (12:26 +0100)]
[LAA] Remove one addRuntimeChecks function (NFC).

In order to reduce the API surface area (preparation for D78460), remove
a addRuntimeChecks() function and do the additional check in the single
caller.

Reviewers: Ayal, anemet

Reviewed By: Ayal

Differential Revision: https://reviews.llvm.org/D79679

4 years ago[InstCombine] fold fpext into exact integer-to-FP cast
Sanjay Patel [Sun, 10 May 2020 10:59:30 +0000 (06:59 -0400)]
[InstCombine] fold fpext into exact integer-to-FP cast

We can combine a floating-point extension cast with a conversion
from integer if we know the earlier cast is exact.

This is an optimization suggested in PR36617:
https://bugs.llvm.org/show_bug.cgi?id=36617#c19

However, this patch does not change the example suggested there.
This patch only uses the existing analysis to handle cases where
the integer source value magnitude is narrower than the
intermediate FP mantissa (guarantees that the conversion to FP is
exact). Follow-up patches to the analysis function can enable
more cases.

Differential Revision: https://reviews.llvm.org/D79116

4 years ago[gn build] Port 35d867a790c
LLVM GN Syncbot [Sun, 10 May 2020 11:04:15 +0000 (11:04 +0000)]
[gn build] Port 35d867a790c

4 years ago[clangd] Filter pch related flags coming from the user
Kadir Cetinkaya [Sat, 9 May 2020 10:34:06 +0000 (12:34 +0200)]
[clangd] Filter pch related flags coming from the user

Summary:
PCH format is unstable, hence using a preamble built with a different
version of clang (or even worse, a different compiler) might result in
unexpected behaviour.

PCH creation on the other hand is something clangd wouldn't want to perform, as
it doesn't generate any output files.

This patch makes sure clangd drops any PCH related compile commands after
parsing the command line args.

Fixes https://github.com/clangd/clangd/issues/248

Reviewers: sammccall

Subscribers: mgorny, ilya-biryukov, MaskRay, jkorous, arphaman, usaxena95, cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D79669

4 years ago[libcxx testing] Remove ALLOW_RETRIES from sleep_until.pass.cpp
David Zarzycki [Sun, 10 May 2020 09:51:14 +0000 (05:51 -0400)]
[libcxx testing] Remove ALLOW_RETRIES from sleep_until.pass.cpp

Operating systems are best effort by default, so we cannot assume that
sleep-like APIs return as soon as we'd like.

Even if a sleep-like API returns when we want it to, the potential for
preemption means that attempts to measure time are subject to delays.

4 years agoRevert two patches, not ready to be shared
Sylvestre Ledru [Sun, 10 May 2020 09:26:53 +0000 (11:26 +0200)]
Revert two patches, not ready to be shared

Revert "clang-tidy doc: add a note for checkers with an autofix"

This reverts commit dc0f79ea5b99b9609a21c29089c294ab82d88094.

Revert "add_new_check.py: Update of the template to add an autofix section"

This reverts commit f97f92e5b0fd14240c2ec27c44e1b3cfddfdf668.

4 years agoclang-tidy doc: Fix the syntax to use rst list type
Sylvestre Ledru [Sun, 10 May 2020 09:22:53 +0000 (11:22 +0200)]
clang-tidy doc: Fix the syntax to use rst list type

4 years agoclang-tidy doc: add a note for checkers with an autofix
Sylvestre Ledru [Mon, 13 Apr 2020 20:33:11 +0000 (22:33 +0200)]
clang-tidy doc: add a note for checkers with an autofix

Summary:
Currently, when looking at a checker documentation, we have to go back
to the whole list or look at the sources to figure out if an autofix
is available or not.

Reviewers: alexfh, aaron.ballman

Subscribers: wuzish, Eugene.Zelenko, nemanjai, kbarton, arphaman, Charusso, cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D77983

4 years agoadd_new_check.py: Update of the template to add an autofix section
Sylvestre Ledru [Mon, 13 Apr 2020 20:22:19 +0000 (22:22 +0200)]
add_new_check.py: Update of the template to add an autofix section

Summary: As discussed here: https://reviews.llvm.org/D77983

Reviewers: alexfh

Subscribers: cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D78052

4 years ago[lld-macho] Delete dylink-lazy test for now
Jez Ng [Sun, 10 May 2020 09:22:33 +0000 (02:22 -0700)]
[lld-macho] Delete dylink-lazy test for now

Summary: It's breaking Windows builds

4 years ago[lld-macho] Try to unbreak dylink-lazy test on Windows
Jez Ng [Sun, 10 May 2020 07:22:36 +0000 (00:22 -0700)]
[lld-macho] Try to unbreak dylink-lazy test on Windows

4 years agoAdd missing pass initialization
Arthur Eubanks [Sat, 9 May 2020 03:19:05 +0000 (20:19 -0700)]
Add missing pass initialization

Summary: This was preventing MemorySanitizerLegacyPass from appearing in --print-after-all.

Reviewers: vitalybuka

Subscribers: hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D79661

4 years ago[lld-macho] Support pc-relative section relocations
Jez Ng [Mon, 27 Apr 2020 19:59:20 +0000 (12:59 -0700)]
[lld-macho] Support pc-relative section relocations

Summary: So far we've only supported symbol relocations.

Reviewers: ruiu, pcc, MaskRay, smeenai, alexshap, gkm, Ktwu, christylee

Subscribers: llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D79211

4 years ago[lld-macho] Use export trie instead of symtab when linking against dylibs
Jez Ng [Thu, 23 Apr 2020 03:00:57 +0000 (20:00 -0700)]
[lld-macho] Use export trie instead of symtab when linking against dylibs

Summary:
This allows us to link against stripped dylibs. Moreover, it's simply
more correct: The symbol table includes symbols that the dylib uses but
doesn't export.

This temporarily regresses our ability to do lazy symbol binding because
dyld_stub_binder isn't in libSystem's export trie. Rather, it is in one
of the sub-libraries libSystem re-exports. (This doesn't affect our
tests since we are mocking out dyld_stub_binder there.) A follow-up diff
will address this by adding support for sub-libraries.

Depends on D79114.

Reviewers: ruiu, pcc, MaskRay, smeenai, alexshap, gkm, Ktwu, christylee

Subscribers: mgorny, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D79226

4 years ago[lld-macho] Dylib symbols should always replace undefined symbols
Jez Ng [Wed, 29 Apr 2020 12:30:47 +0000 (05:30 -0700)]
[lld-macho] Dylib symbols should always replace undefined symbols

Summary:
Otherwise we get undefined symbol errors depending on the order of
arguments on the command line.

Depends on D78270.

Reviewers: ruiu, pcc, MaskRay, smeenai, alexshap, gkm, Ktwu, christylee

Subscribers: llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D79114

4 years ago[lld-macho] Support calls to functions in dylibs
Jez Ng [Wed, 6 May 2020 00:38:10 +0000 (17:38 -0700)]
[lld-macho] Support calls to functions in dylibs

Summary:
This diff implements lazy symbol binding -- very similar to the PLT
mechanism in ELF.

ELF's .plt section is broken up into two sections in Mach-O:
StubsSection and StubHelperSection. Calls to functions in dylibs will
end up calling into StubsSection, which contains indirect jumps to
addresses stored in the LazyPointerSection (the counterpart to ELF's
.plt.got).

Initially, the LazyPointerSection contains addresses that point into one
of the entry points in the middle of the StubHelperSection. The code in
StubHelperSection will push on the stack an offset into the
LazyBindingSection. The push is followed by a jump to the beginning of
the StubHelperSection (similar to PLT0), which then calls into
dyld_stub_binder. dyld_stub_binder is a non-lazily bound symbol, so this
call looks it up in the GOT.

The stub binder will look up the bind opcodes in the LazyBindingSection
at the given offset. The bind opcodes will tell the binder to update the
address in the LazyPointerSection to point to the symbol, so that
subsequent calls don't have to redo the symbol resolution. The binder
will then jump to the resolved symbol.

Depends on D78269.

Reviewers: ruiu, pcc, MaskRay, smeenai, alexshap, gkm, Ktwu, christylee

Subscribers: llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D78270

4 years ago[lld-macho] Follow-up to D77893
Jez Ng [Wed, 6 May 2020 00:25:58 +0000 (17:25 -0700)]
[lld-macho] Follow-up to D77893

Summary:
1. Don't have isHidden() depend on isNeeded(). Whether a section is
  hidden is orthogonal from whether it is needed: hidden sections will
  never have a header regardless of whether they have a body. (I know we
  override this method with return false for synthetic sections, but
  regardless I think it's confusing to write it this way for non-synthetic
  sections.)

2. Don't call writeTo() on unneeded sections. D78270 assumes that this
  is true when implementing the stub helper section.

3. Filter out the unneeded sections early on to avoid having to deal
   with them in multiple places.

4. Remove assumption in test that the referenced file has no other symbols.
  (We should create separate input files for future tests to avoid such
  issues.)

Reviewers: ruiu, pcc, MaskRay, smeenai, alexshap, gkm, Ktwu, christylee

Subscribers: llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D79460

4 years ago[sanitizer] Enable whitelist/blacklist in new PM
Jinsong Ji [Sun, 10 May 2020 02:33:57 +0000 (02:33 +0000)]
[sanitizer] Enable whitelist/blacklist in new PM

https://reviews.llvm.org/D63616 added `-fsanitize-coverage-whitelist`
and `-fsanitize-coverage-blacklist` for clang.

However, it was done only for legacy pass manager.
This patch enable it for new pass manager as well.

Reviewed By: vitalybuka

Differential Revision: https://reviews.llvm.org/D79653

4 years ago[X86] Add test cases for vXi16 PMULH opportunities that don't end in truncate.
Craig Topper [Sun, 10 May 2020 00:04:34 +0000 (17:04 -0700)]
[X86] Add test cases for vXi16 PMULH opportunities that don't end in truncate.

We already have matching for extend+mul+shift+trunc. But we could
also match up to the shift without the truncate and just extend the
result. That would still be a savings.

4 years agoGlobalISel: Handle more cases in lowerUnmergeValues
Matt Arsenault [Sun, 29 Mar 2020 22:04:53 +0000 (18:04 -0400)]
GlobalISel: Handle more cases in lowerUnmergeValues

Handle scalar sources, as well as vectors.

4 years ago[BreakFalseDeps] Harden pickBestRegisterForUndef against changing tied operands or...
Craig Topper [Sat, 9 May 2020 22:33:08 +0000 (15:33 -0700)]
[BreakFalseDeps] Harden pickBestRegisterForUndef against changing tied operands or physical registers that aren't renamable.

I don't have any test cases since X86 doesn't return any tied
operands from getUndefRegClearance today. But conceivably we could
want BreakFalseDeps to insert a dependency breaking XOR for
a tied operand in the future.

4 years ago[LAA] Remove unneeded PtrRtChecking argument (NFC).
Florian Hahn [Sat, 9 May 2020 21:07:51 +0000 (22:07 +0100)]
[LAA] Remove unneeded PtrRtChecking argument (NFC).

The argument is not required and simplifies D78460 a bit.

4 years ago[X86] isVectorShiftByScalarCheap - don't limit fast XOP vector shifts to 128-bit...
Simon Pilgrim [Sat, 9 May 2020 21:09:01 +0000 (22:09 +0100)]
[X86] isVectorShiftByScalarCheap - don't limit fast XOP vector shifts to 128-bit vectors

XOP targets have fast per-element vector shifts and we're better off splitting to 128-bit shifts where necessary (which is what we already do in LowerShift).

4 years ago[Clang] Pass -z max-page-size to linker for Fuchsia
Petr Hosek [Sat, 9 May 2020 20:44:20 +0000 (13:44 -0700)]
[Clang] Pass -z max-page-size to linker for Fuchsia

Currently all Fuchsia ABIs use a 4k page size, departing from
the recommended page sizes in the respective psABI documents.

Differential Revision: https://reviews.llvm.org/D79667

4 years ago[Clang] Pass --pack-dyn-relocs=relr to lld for Fuchsia
Petr Hosek [Sat, 9 May 2020 20:42:19 +0000 (13:42 -0700)]
[Clang] Pass --pack-dyn-relocs=relr to lld for Fuchsia

The compact format is fully supported on Fuchsia and is the
preferred default.

Patch By: mcgrathr

Differential Revision: https://reviews.llvm.org/D79665

4 years agoGlobalISel: Move code into lowering for G_MERGE_VALUES
Matt Arsenault [Sun, 29 Mar 2020 19:51:54 +0000 (15:51 -0400)]
GlobalISel: Move code into lowering for G_MERGE_VALUES

Currently this code exists in widenScalar for G_MERGE_VALUE
sources. I'm not sure if the existing expansion in widenScalar should
be removed or not. The widenScalar variant tries to extend to the
requested size, but this just uses the original bitwidth.

4 years agoGlobalISel: Combine G_UNMERGE_VALUES with G_TRUNC
Matt Arsenault [Sat, 4 Apr 2020 19:07:28 +0000 (15:07 -0400)]
GlobalISel: Combine G_UNMERGE_VALUES with G_TRUNC

G_BITCAST can be lowered with a pair of G_UNMERGE_VALUES and
G_MERGE_VALUES with different types, but G_UNMERGE_VALUES of a vector
can also be implemented with a bitcast to a scalar, which introduces
the possibility for infinite loops. Try to eliminate an illegal source
register type in the artifact combiner to avoid this from happening.

Avoids infinite looping in the legalizer in a future patch which
allows lowering G_UNMERGE_VALUES of a vector source with a G_BITCAST.

4 years agoInstCombine: Broaden copy-constant-to-alloca optimization
Matt Arsenault [Fri, 8 May 2020 19:35:41 +0000 (15:35 -0400)]
InstCombine: Broaden copy-constant-to-alloca optimization

Consider any constant memory type, not just global constants. AMDGPU
kernel parameters are effectively global constants, but appear as
either reads from an intrinsic derived pointer or function argument.

4 years agoFix typo
Matt Arsenault [Wed, 6 May 2020 00:25:03 +0000 (20:25 -0400)]
Fix typo

4 years agoAMDGPU: Skip GetUnderlyingObject check in pointsToConstantMemory
Matt Arsenault [Fri, 8 May 2020 19:53:41 +0000 (15:53 -0400)]
AMDGPU: Skip GetUnderlyingObject check in pointsToConstantMemory

Check the address space first before searching for the object
definition to save compile time. As an added bonus, this will now
treat casts to constant addrspace as constant.

We also seemed to be missing targeted tests for this, so add a few
missing other cases too.

4 years ago[CodeGenPrepare][X86] Add x16i16, v32i8 and XOP vector shift by scalar amount tests
Simon Pilgrim [Sat, 9 May 2020 14:54:53 +0000 (15:54 +0100)]
[CodeGenPrepare][X86] Add x16i16, v32i8 and XOP vector shift by scalar amount tests

Helps improve test coverage of the XOP modes in X86TargetLowering::isVectorShiftByScalarCheap (and where we always return false for vXi8 vector shifts).

4 years ago[X86] Add XOP vector shift by scalar amount tests
Simon Pilgrim [Sat, 9 May 2020 14:53:06 +0000 (15:53 +0100)]
[X86] Add XOP vector shift by scalar amount tests

Helps improve test coverage of the XOP modes in X86TargetLowering::isVectorShiftByScalarCheap

4 years ago[X86] Teach getUndefRegClearance that we use undef for inputs to PUNPCK in some cases.
Craig Topper [Sat, 9 May 2020 18:25:36 +0000 (11:25 -0700)]
[X86] Teach getUndefRegClearance that we use undef for inputs to PUNPCK in some cases.

This enables the register to be changed from XMM/YMM/ZMM0 to
instead match the other source. This prevents a false
dependency.

I added all the integer unpck instructions, but the tests
only show changes for BW and WD.

Unfortunately, we can have undef on operand 1 or 2 of the AVX
instructions. This breaks the interface with hasUndefRegUpdate
which used to tell which operand to check.

Now we scan the input operands looking for an undef register and
then ask hasUndefRegUpdate if its an instruction we care about
and which operands of that instruction we care about.

I also had to make some changes to the load folding code to
always pass operand 1 to hasUndefRegUpdate. I've updated
hasUndefRegUpdate to return false when ForLoadFold is set for
instructions that are not explicitly blocked for load folding in
isel patterns.

Differential Revision: https://reviews.llvm.org/D79615

4 years ago[X86] Add an assert that v32i16/v64i8 splitting in LowerVSETCC should only occur...
Craig Topper [Sat, 9 May 2020 18:23:27 +0000 (11:23 -0700)]
[X86] Add an assert that v32i16/v64i8 splitting in LowerVSETCC should only occur when AVX512BW is disabled. NFC

With BWI we should only get a v32i1/v64i1 result type.

4 years ago[ARM] Convert VDUPLANE to VDUP under MVE
David Green [Sat, 9 May 2020 14:17:50 +0000 (15:17 +0100)]
[ARM] Convert VDUPLANE to VDUP under MVE

Unlike Neon, MVE does not have a way of duplicating from a vector lane,
so a VDUPLANE currently selects to a VDUP(move_from_lane(..)). This
forces that to be done earlier as a dag combine to allow other folds to
happen.

It converts to a VDUP(EXTRACT). On FP16 this is then folded to a
VGETLANEu to prevent it from creating a vmovx;vmovhr pair, using a
single move_from_reg instead.

Differential Revision: https://reviews.llvm.org/D79606

4 years ago[NFC] Small rework to RenamerClangTidyCheck addUsage
Nathan James [Sat, 9 May 2020 17:57:05 +0000 (18:57 +0100)]
[NFC] Small rework to RenamerClangTidyCheck addUsage

4 years ago[lldb/debugserver] Include TargetConditionals.h where needed
Fred Riss [Sat, 9 May 2020 17:11:28 +0000 (10:11 -0700)]
[lldb/debugserver] Include TargetConditionals.h where needed

MachProcess.mm uses a TARGET_OS_ macro without directly including
TargetConditionals.h. This currently works as we get the header
as an indirect dependency, but might not in the future.

I just spent some time investigating an internal regression
caused by a similar issue, so I audited the codebase for such
cases.

4 years ago[clangd] Fix data race in BackgroundIndex test
Kadir Cetinkaya [Sat, 9 May 2020 16:14:48 +0000 (18:14 +0200)]
[clangd] Fix data race in BackgroundIndex test

MockFSProvider is not thread-safe. Make sure we don't modify it while
background index is working.

4 years ago[flang] Make implicit conversion explicit in assignment
Tim Keith [Sat, 9 May 2020 16:10:08 +0000 (09:10 -0700)]
[flang] Make implicit conversion explicit in assignment

When intrinsic types are assigned there are some implicit conversions
that take place. This change make them explicit in the types
representation of assignments.

Differential Revision: https://reviews.llvm.org/D79637

4 years ago[clang-tidy] RenamerClangTidy now renames dependent member expr when the member can...
Nathan James [Fri, 8 May 2020 18:30:18 +0000 (19:30 +0100)]
[clang-tidy] RenamerClangTidy now renames dependent member expr when the member can be resolved

Summary:
Sometimes in templated code Member references are reported as `DependentScopeMemberExpr` because that's what the standard dictates, however in many trivial cases it is easy to resolve the reference to its actual Member.
Take this code:
```
template<typename T>
class A{
  int value;
  A& operator=(const A& Other){
    value = Other.value;
    this->value = Other.value;
    return *this;
  }
};
```
When ran with `clang-tidy file.cpp -checks=readability-identifier-naming --config="{CheckOptions: [{key: readability-identifier-naming.MemberPrefix, value: m_}]}" -fix`
Current behaviour:
```
template<typename T>
class A{
  int m_value;
  A& operator=(const A& Other){
    m_value = Other.value;
    this->value = Other.value;
    return *this;
  }
};
```
As `this->value` and `Other.value` are Dependent they are ignored when creating the fix-its, however this can easily be resolved.
Proposed behaviour:
```
template<typename T>
class A{
  int m_value;
  A& operator=(const A& Other){
    m_value = Other.m_value;
    this->m_value = Other.m_value;
    return *this;
  }
};
```

Reviewers: aaron.ballman, JonasToth, alexfh, hokein, gribozavr2

Reviewed By: aaron.ballman

Subscribers: merge_guards_bot, xazax.hun, cfe-commits

Tags: #clang, #clang-tools-extra

Differential Revision: https://reviews.llvm.org/D73052

4 years ago[libcxx testing] Make three locking tests more reliable
David Zarzycki [Sat, 9 May 2020 15:10:41 +0000 (11:10 -0400)]
[libcxx testing] Make three locking tests more reliable

The challenge with measuring time in tests is that slow and/or busy
machines can cause tests to fail in unexpected ways. After this change,
three tests should be much more robust. The only remaining and tiny race
that I can think of is preemption after `--countDown`. That being said,
the race isn't fixable because the standard library doesn't provide a
way to count threads that are waiting to acquire a lock.

Reviewers: ldionne, EricWF, howard.hinnant, mclow.lists, #libc

Reviewed By: ldionne, #libc

Subscribers: dexonsmith, jfb, broadwaylamb, libcxx-commits

Tags: #libc

Differential Revision: https://reviews.llvm.org/D79406

4 years agoLTO.h - reduce includes to forward declarations. NFC.
Simon Pilgrim [Sat, 9 May 2020 14:10:51 +0000 (15:10 +0100)]
LTO.h - reduce includes to forward declarations. NFC.

Add missing ToolOutputFile.h dependency to BackendUtil.cpp

4 years agoLLParser.h - remove unused ValueHandle.h include. NFC.
Simon Pilgrim [Sat, 9 May 2020 14:08:48 +0000 (15:08 +0100)]
LLParser.h - remove unused ValueHandle.h include. NFC.

4 years ago[X86] Remove mul(abs(x),abs(x)) -> mul(x,x) tests
Simon Pilgrim [Sat, 9 May 2020 14:07:15 +0000 (15:07 +0100)]
[X86] Remove mul(abs(x),abs(x)) -> mul(x,x) tests

This is handled in InstCombine (D79319) and its unlikely that these can occur in DAG (see D79304).

4 years ago[X86] Allow combineVectorCompareAndMaskUnaryOp to handle 'all-bits' general case
Simon Pilgrim [Sat, 9 May 2020 13:53:07 +0000 (14:53 +0100)]
[X86] Allow combineVectorCompareAndMaskUnaryOp to handle 'all-bits' general case

For the sint_to_fp(and(X,C)) -> and(X,sint_to_fp(C)) fold, allow combineVectorCompareAndMaskUnaryOp to match any X that ComputeNumSignBits says is all-bits, not just SETCC.

Noticed while investigating mask promotion issues in PR45808

4 years ago[X86] Add test cases for 'abs from mul patterns' (PR45691)
Simon Pilgrim [Sat, 9 May 2020 13:41:27 +0000 (14:41 +0100)]
[X86] Add test cases for 'abs from mul patterns' (PR45691)

4 years ago[clangd] Fix a data race in RecordsLatencies test
Kadir Cetinkaya [Sat, 9 May 2020 13:42:15 +0000 (15:42 +0200)]
[clangd] Fix a data race in RecordsLatencies test

4 years ago[X86] Add tests showing failure of combineVectorCompareAndMaskUnaryOp to handle ...
Simon Pilgrim [Sat, 9 May 2020 13:24:38 +0000 (14:24 +0100)]
[X86] Add tests showing failure of combineVectorCompareAndMaskUnaryOp to handle 'all-bits' general case

For the sint_to_fp(and(X,C)) -> and(X,sint_to_fp(C)) fold, combineVectorCompareAndMaskUnaryOp only matches X against SETCC (with an all-bits result) when really it could accept anything that ComputeNumSignBits says is all-bits.

Noticed while investigating mask promotion issues in PR45808

4 years agoNativeFormatting.h - reduce raw_ostream.h include to forward declaration. NFC.
Simon Pilgrim [Sat, 9 May 2020 12:31:39 +0000 (13:31 +0100)]
NativeFormatting.h - reduce raw_ostream.h include to forward declaration. NFC.

4 years ago[clang-format] [PR34574] Handle [[nodiscard]] attribute in class declaration
mydeveloperday [Sat, 9 May 2020 10:26:38 +0000 (11:26 +0100)]
[clang-format] [PR34574] Handle [[nodiscard]] attribute in class declaration

Summary:
https://bugs.llvm.org/show_bug.cgi?id=34574
https://bugs.llvm.org/show_bug.cgi?id=38401

```
template <typename T>
class [[nodiscard]] result
{
  public:
    result(T&&)
    {
    }
};
```

formats incorrectly to

```
template <typename T>
class [[nodiscard]] result{public : result(T &&){}};
```

Reviewed By: krasimir

Subscribers: cfe-commits

Tags: #clang, #clang-format

Differential Revision: https://reviews.llvm.org/D79354

4 years ago[MLIR] Register JIT event listeners with RTDyldObjectLinkingLayer
Eugene Zhulenev [Sat, 9 May 2020 09:13:50 +0000 (11:13 +0200)]
[MLIR] Register JIT event listeners with RTDyldObjectLinkingLayer

Use a new API to register JIT event listeners.

Differential Revision: https://reviews.llvm.org/D78435

4 years ago[lldb] [testsuite] TestReproducerAttach.py: Fix dependency on external symbol files
Jan Kratochvil [Sat, 9 May 2020 07:06:37 +0000 (09:06 +0200)]
[lldb] [testsuite] TestReproducerAttach.py: Fix dependency on external symbol files

D55859 and D63339 prevented needless dependencies on system symbol
files. This testcase was checked-in afterwards and it brings back one
such unwanted dependency. Under some circumstances it may cause false
FAILs and/or excessive resource usage to run the testcase.

clang-format does not support .py so I have formatted it as I found most
compatible.

Also this is not a full testcase-style initialization, for example
--no-lldbinit ignores env("NO_LLDBINIT") setting which lldbtest.py does
implement:
  # If we spawn an lldb process for test (via pexpect), do not load the
  # init file unless told otherwise.
  if os.environ.get("NO_LLDBINIT") != "NO":
      self.lldbOption += " --no-lldbinit"

But this is what lldbpexpect.py does - it also ignores
env("NO_LLDBINIT"). Sure one could also fix lldbpexpect.py to unify the
initialization more with lldbtest.py but I find that outside of the
scope of this patch.

Differential Revision: https://reviews.llvm.org/D79649

4 years ago[Driver] Don't pass -u__llvm_profile_runtime for clang -fprofile-arcs a.o
Fangrui Song [Sat, 9 May 2020 06:13:19 +0000 (23:13 -0700)]
[Driver] Don't pass -u__llvm_profile_runtime for clang -fprofile-arcs a.o

clang --coverage a.o       # InstrProfilingRuntime.cpp.o not linked in
clang --fprofile-arcs a.o  # InstrProfilingRuntime.cpp.o unexpectedly linked in

Fix --fprofile-arcs.

4 years ago[NFC] Clean up in MCObjectStreamer and X86AsmBackend
Shengchen Kan [Sat, 9 May 2020 04:47:44 +0000 (12:47 +0800)]
[NFC] Clean up in MCObjectStreamer and X86AsmBackend

4 years agoRevert "Relands "[YAMLVFSWriter][Test][NFC] Add couple tests""
Jan Korous [Sat, 9 May 2020 04:36:29 +0000 (21:36 -0700)]
Revert "Relands "[YAMLVFSWriter][Test][NFC] Add couple tests""

This reverts commit 49b32d80416288b6eb8e26f76c40a8e32c20a361.

4 years ago[DebugInfo] Dump raw data in a case of decoding error of an expression.
Igor Kudrin [Sat, 9 May 2020 03:03:38 +0000 (10:03 +0700)]
[DebugInfo] Dump raw data in a case of decoding error of an expression.

It looks like that was an initial intention, but some code paths in
`DWARFExpression::Operation::extract()` did not initialize `EndOffset`
properly.

Differential Revision: https://reviews.llvm.org/D79622

4 years agoFix parsing of enum-base to follow C++11 rules.
Richard Smith [Sat, 9 May 2020 02:24:18 +0000 (19:24 -0700)]
Fix parsing of enum-base to follow C++11 rules.

Previously we implemented non-standard disambiguation rules to
distinguish an enum-base from a bit-field but otherwise treated a :
after an elaborated-enum-specifier as introducing an enum-base. That
misparses various examples (anywhere an elaborated-type-specifier can
appear followed by a colon, such as within a ternary operator or
_Generic).

We now implement the C++11 rules, with the old cases accepted as
extensions where that seemed reasonable. These amount to:
 * an enum-base must always be accompanied by an enum definition (except
   in a standalone declaration of the form 'enum E : T;')
 * in a member-declaration, 'enum E :' always introduces an enum-base,
   never a bit-field
 * in a type-specifier (or similar context), 'enum E :' is not
   permitted; the colon means whatever else it would mean in that
   context.

Fixed underlying types for enums are also permitted in Objective-C and
under MS extensions, plus as a language extension in all other modes.
The behavior in ObjC and MS extensions modes is unchanged (but the
bit-field disambiguation is a bit better); remaining language modes
follow the C++11 rules.

Fixes PR45726, PR39979, PR19810, PR44941, and most of PR24297, plus C++
core issues 1514 and 1966.

4 years agoRelands "[YAMLVFSWriter][Test][NFC] Add couple tests"
Jan Korous [Fri, 8 May 2020 02:11:51 +0000 (19:11 -0700)]
Relands "[YAMLVFSWriter][Test][NFC] Add couple tests"

Fixed test for Windows.

Differential Revision: https://reviews.llvm.org/D79552

4 years agoclang: Cleanup usage of CreateMemCpy
Matt Arsenault [Fri, 8 May 2020 18:39:07 +0000 (14:39 -0400)]
clang: Cleanup usage of CreateMemCpy

It handles the the pointee type casts in preparation for opaque
pointers.

4 years ago[Driver] Add -fno-test-coverage
Fangrui Song [Fri, 8 May 2020 23:50:34 +0000 (16:50 -0700)]
[Driver] Add -fno-test-coverage

4 years ago[hwasan] Allow -hwasan-globals flag to appear more than once.
Evgenii Stepanov [Fri, 8 May 2020 23:30:45 +0000 (16:30 -0700)]
[hwasan] Allow -hwasan-globals flag to appear more than once.

4 years ago[hwasan] Untag destination address in hwasan_posix_memalign.
Evgenii Stepanov [Fri, 8 May 2020 23:29:46 +0000 (16:29 -0700)]
[hwasan] Untag destination address in hwasan_posix_memalign.

Required on X86 because no TBI.

4 years ago[Driver] Don't warn -Wunused-command-line-argument for --coverage -ftest-coverage...
Fangrui Song [Fri, 8 May 2020 23:14:41 +0000 (16:14 -0700)]
[Driver] Don't warn -Wunused-command-line-argument for --coverage -ftest-coverage -fprofile-arcs

4 years ago[LangRef] Describe linkage types, allocation size of declarations for global variables
Matthias Schiffer [Fri, 8 May 2020 23:20:43 +0000 (16:20 -0700)]
[LangRef] Describe linkage types, allocation size of declarations for global variables

Linkage type was only referenced for functions, not for global
variables.

Clarify that LLVM doesn't make assumption about the allocation size when
no definitive initializer for a global variable is known.

Differential Revision: https://reviews.llvm.org/D78952

4 years ago[Driver] Reorganize --coverage -ftest-coverage -fprofile-arcs related tests
Fangrui Song [Fri, 8 May 2020 22:04:03 +0000 (15:04 -0700)]
[Driver] Reorganize --coverage -ftest-coverage -fprofile-arcs related tests

And fix a comment about __llvm_profile_runtime

4 years ago[SelectionDAG] Remove ConstantPoolSDNode::getAlignment.
Craig Topper [Fri, 8 May 2020 22:06:37 +0000 (15:06 -0700)]
[SelectionDAG] Remove ConstantPoolSDNode::getAlignment.

Use getAlign instead.

Differential Revision: https://reviews.llvm.org/D79459

4 years ago[SelectionDAG] Use Align/MaybeAlign for ConstantPoolSDNode.
Craig Topper [Fri, 8 May 2020 22:06:15 +0000 (15:06 -0700)]
[SelectionDAG] Use Align/MaybeAlign for ConstantPoolSDNode.

This patch stores the alignment for ConstantPoolSDNode as an
Align and updates the getConstantPool interface to take a MaybeAlign.

Removing getAlignment() will be done as a follow up.

Differential Revision: https://reviews.llvm.org/D79436

4 years agoAdd Operation::moveAfter
Geoffrey Martin-Noble [Fri, 8 May 2020 22:34:13 +0000 (22:34 +0000)]
Add Operation::moveAfter

This revision introduces an Operation::moveAfter mirroring
Operation::moveBefore to move an operation after another
existing operation or an iterator in a specified block.

Resolves https://bugs.llvm.org/show_bug.cgi?id=45799

Differential Revision: https://reviews.llvm.org/D79640

4 years ago[AMDGPU] Vectorize alloca thru bitcast
Stanislav Mekhanoshin [Tue, 5 May 2020 23:17:53 +0000 (16:17 -0700)]
[AMDGPU] Vectorize alloca thru bitcast

This is mostly useful if alloca element type is not integer
and then casted to an integer for load or store. We now can
vectorize an [i32] alloca but cannot do so for [float].

There also a separate patch needed to properly lower 64 bit
types after they vectorized. At the moment these are lowered
via scratch anyway.

Differential Revision: https://reviews.llvm.org/D79641

4 years ago[TRE][NFC] Refactor shared state into member variables.
Layton Kifer [Fri, 8 May 2020 21:18:53 +0000 (14:18 -0700)]
[TRE][NFC] Refactor shared state into member variables.

Separate functions that require shared state into a class to avoid
needing to pass them though multiple functions just to be available
where needed.

The main motivation for this is that we would like to remove the
limitation that accumulator values be dynamic constant, which would
require additional shared state between call eliminations in the same
function, compounding this issue.

Differential Revision: https://reviews.llvm.org/D79299

4 years ago[COFF] Use Expected in COFFObjectFile creation
Reid Kleckner [Fri, 8 May 2020 20:58:56 +0000 (13:58 -0700)]
[COFF] Use Expected in COFFObjectFile creation

The constructor error out parameter was a bit awkward. Wrap it in a
factory method which can return an error. Make the constructor private.

4 years ago[COFF] Migrate COFFObjectFile to Expected<T>
Reid Kleckner [Fri, 8 May 2020 17:41:05 +0000 (10:41 -0700)]
[COFF] Migrate COFFObjectFile to Expected<T>

I noticed that std::error_code() does one-time initialization. Avoid
that overhead with Expected<T> and llvm::Error. Also, it is consistent
with the virtual interface and ELF, and generally cleaner.

Reviewed By: MaskRay

Differential Revision: https://reviews.llvm.org/D79643

4 years ago[clang][WebAssembly] Only expose wait and notify builtins with atomics
Thomas Lively [Thu, 7 May 2020 00:17:48 +0000 (17:17 -0700)]
[clang][WebAssembly] Only expose wait and notify builtins with atomics

Summary:
Since the underlying wait and notify instructions are only available
when the atomics feature is enabled, it only makes sense to expose
their builtin functions when atomics are enabled.

Reviewers: aheejin, sunfish

Subscribers: dschuff, sbc100, jgravelle-google, jfb, cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D79534

4 years ago[WebAssembly] Disallow 'shared-mem' rather than 'atomics'
Thomas Lively [Thu, 7 May 2020 02:33:24 +0000 (19:33 -0700)]
[WebAssembly] Disallow 'shared-mem' rather than 'atomics'

Summary:
The WebAssembly backend automatically lowers atomic operations and TLS
to nonatomic operations and non-TLS data when either are present and
the atomics or bulk-memory features are not present, respectively. The
resulting object is no longer thread-safe, so the linker has to be
told not to allow it to be linked into a module with shared
memory. This was previously done by disallowing the 'atomics' feature,
which prevented any objct with its atomic operations or TLS removed
from being linked with any object containing atomics or TLS, and
therefore preventing it from being linked into a module with shared
memory since shared memory requires atomics.

However, as of https://github.com/WebAssembly/threads/issues/144, the
validation rules are relaxed to allow atomic operations to validate
with unshared memories, which makes it perfectly safe to link an
object with stripped atomics and TLS with another object that still
contains TLS and atomics as long as the resulting module has an
unshared memory. To allow this kind of link, this patch disallows a
pseudo-feature 'shared-mem' rather than 'atomics' to communicate to
the linker that the object is not thread-safe. This means that the
'atomics' feature is available to accurately reflect whether or not an
object has atomics enabled.

As a drive-by tweak, this change also requires that bulk-memory be
enabled in addition to atomics in order to use shared memory. This is
because initializing shared memories requires bulk-memory operations.

Reviewers: aheejin, sbc100

Subscribers: dschuff, jgravelle-google, hiraditya, sunfish, jfb, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D79542

4 years ago[Target][XCOFF] Correctly halt when mixing AIX or XCOFF with ppc64le
Hubert Tong [Fri, 8 May 2020 20:40:28 +0000 (16:40 -0400)]
[Target][XCOFF] Correctly halt when mixing AIX or XCOFF with ppc64le

The code to prevent using `PPCXCOFFMCAsmInfo` with little-endian targets
used an incorrect check. Also, there does not appear to be sufficient
earlier checking to prevent failing this check, so the check here is
upgraded to be a `report_fatal_error`.

`PPCAIXAsmPrinter` was also missing a check against use with
little-endian targets. This patch adds such a check in.

4 years ago[XCOFF] XCOFF constants, MCObjectFileInfo placeholder code for DWARF; NFC
Hubert Tong [Fri, 8 May 2020 20:32:12 +0000 (16:32 -0400)]
[XCOFF] XCOFF constants, MCObjectFileInfo placeholder code for DWARF; NFC

Summary:
This patch introduces the constants defined to identify DWARF sections
in XCOFF into `llvm/BinaryFormat/XCOFF.h` and adds (NFC) placeholder
code to `llvm/lib/MC/MCObjectFileInfo.cpp` where the DWARF sections for
XCOFF are to be set up.

Reviewers: jasonliu, sfertile, daltenty, DiggerLin, Xiangling_L

Reviewed By: jasonliu, sfertile, DiggerLin

Differential Revision: https://reviews.llvm.org/D79220

4 years ago[AIX] Avoid structor alias; die before bad alias codegen
Hubert Tong [Fri, 8 May 2020 20:29:58 +0000 (16:29 -0400)]
[AIX] Avoid structor alias; die before bad alias codegen

Summary:
`AsmPrinter::emitGlobalIndirectSymbol` is dependent on
`MCStreamer::emitAssignment` to produce `.set` directives for alias
symbols; however, the `.set` pseudo-op on AIX is documented as not
usable with external relocatable terms or expressions, which limits its
applicability in generating alias symbols.

Disable generating aliases on AIX until a different implementation
strategy is available.

Reviewers: cebowleratibm, jasonliu, sfertile, daltenty, DiggerLin

Reviewed By: jasonliu

Differential Revision: https://reviews.llvm.org/D79044

4 years ago[GlobalISel] Don't add duplicate successors to MBBs when translating indirectbr
Jessica Paquette [Thu, 7 May 2020 23:25:34 +0000 (16:25 -0700)]
[GlobalISel] Don't add duplicate successors to MBBs when translating indirectbr

This fixes a verifier failure on a bot:

http://green.lab.llvm.org/green/job/test-suite-verify-machineinstrs-aarch64-O0-g/

```
*** Bad machine code: MBB has duplicate entries in its successor list. ***
- function:    foo
- basic block: %bb.5 indirectgoto (0x7fe3d687ca08)
```

One of the GCC torture suite tests (pr70460.c) has an indirectbr instruction
which has duplicate blocks in its destination list.

According to the langref this is allowed:

> Blocks are allowed to occur multiple times in the destination list, though
> this isn’t particularly useful.
(https://www.llvm.org/docs/LangRef.html#indirectbr-instruction)

We don't allow this in MIR. So, when we translate such an instruction, the
verifier screams.

This patch makes `translateIndirectBr` check if a successor has already been
added to a block. If the successor is present, it is skipped rather than added
twice.

Differential Revision: https://reviews.llvm.org/D79609

4 years ago[VectorCombine] scalarize binop of inserted elements into vector constants
Sanjay Patel [Fri, 8 May 2020 20:29:07 +0000 (16:29 -0400)]
[VectorCombine] scalarize binop of inserted elements into vector constants

As with the extractelement patterns that are currently in vector-combine,
there are going to be several possible variations on this theme. This
should be the clearest, simplest example.

Scalarization is the right direction for target-independent canonicalization,
and InstCombine has some of those folds already, but it doesn't do this.
I proposed a similar transform in D50992. Here in vector-combine, we can
check the cost model to be sure it's profitable, so there should be less risk.

Differential Revision: https://reviews.llvm.org/D79452