platform/upstream/llvm.git
3 years ago[LV] Do not check widening decision for instrs outside of loop.
Florian Hahn [Mon, 3 Aug 2020 08:47:16 +0000 (09:47 +0100)]
[LV] Do not check widening decision for instrs outside of loop.

No widening decisions will be computed for instructions outside the
loop. Do not try to get a widening decision. The load/store will be just
a scalar load, so treating at as normal should be fine I think.

Fixes PR46950.

Reviewed By: dmgreen

Differential Revision: https://reviews.llvm.org/D85087

3 years ago[mlir][Vector] Add transformation + pattern to split vector.transfer_read into full...
Nicolas Vasilache [Mon, 3 Aug 2020 08:39:18 +0000 (04:39 -0400)]
[mlir][Vector] Add transformation + pattern to split vector.transfer_read into full and partial copies.

This revision adds a transformation and a pattern that rewrites a "maybe masked" `vector.transfer_read %view[...], %pad `into a pattern resembling:

```
   %1:3 = scf.if (%inBounds) {
      scf.yield %view : memref<A...>, index, index
    } else {
      %2 = vector.transfer_read %view[...], %pad : memref<A...>, vector<...>
      %3 = vector.type_cast %extra_alloc : memref<...> to
      memref<vector<...>> store %2, %3[] : memref<vector<...>> %4 =
      memref_cast %extra_alloc: memref<B...> to memref<A...> scf.yield %4 :
      memref<A...>, index, index
   }
   %res= vector.transfer_read %1#0[%1#1, %1#2] {masked = [false ... false]}
```
where `extra_alloc` is a top of the function alloca'ed buffer of one vector.

This rewrite makes it possible to realize the "always full tile" abstraction where vector.transfer_read operations are guaranteed to read from a padded full buffer.
The extra work only occurs on the boundary tiles.

Differential Revision: https://reviews.llvm.org/D84631

3 years ago[debugserver] Fix that is_dot_app is producing unused warnings
Raphael Isemann [Mon, 3 Aug 2020 08:23:48 +0000 (10:23 +0200)]
[debugserver] Fix that is_dot_app is producing unused warnings

Some build configurations don't use this static function.

3 years ago[MLIR][Shape] Lower `shape.broadcast` to `scf`
Frederik Gossen [Mon, 3 Aug 2020 08:18:48 +0000 (08:18 +0000)]
[MLIR][Shape] Lower `shape.broadcast` to `scf`

Differential Revision: https://reviews.llvm.org/D85027

3 years ago[MachOYAML] Remove redundant variable initialization. NFC.
Xing GUO [Mon, 3 Aug 2020 08:14:47 +0000 (16:14 +0800)]
[MachOYAML] Remove redundant variable initialization. NFC.

The value of `is64Bit` is initialized in the constructor body.

3 years ago[Attributor] Check nonnull attribute violation in AAUndefinedBehavior
Shinji Okumura [Mon, 3 Aug 2020 08:02:49 +0000 (17:02 +0900)]
[Attributor] Check nonnull attribute violation in AAUndefinedBehavior

This patch makes it possible to handle nonnull attribute violation at callsites in AAUndefinedBehavior.
If null pointer is passed to callee at a callsite and the corresponding argument of callee has nonnull attribute, the behavior of the callee is undefined.
In this patch, violations of argument nonnull attributes is only handled.
But violations of returned nonnull attributes can be handled and I will implement that in a follow-up patch.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D84733

3 years ago[DebugInfo] Make DIEDelta::SizeOf() more explicit. NFCI.
Igor Kudrin [Mon, 3 Aug 2020 08:04:15 +0000 (15:04 +0700)]
[DebugInfo] Make DIEDelta::SizeOf() more explicit. NFCI.

The patch restricts DIEDelta::SizeOf() to accept only DWARF forms that
are actually used in the LLVM codebase. This should make the use of the
class more explicit and help to avoid issues similar to fixed in D83958
and D84094.

Differential Revision: https://reviews.llvm.org/D84095

3 years ago[DebugInfo] Fix misleading using of DWARF forms with DIELabel. NFCI.
Igor Kudrin [Mon, 3 Aug 2020 08:04:08 +0000 (15:04 +0700)]
[DebugInfo] Fix misleading using of DWARF forms with DIELabel. NFCI.

DIELabel can emit only 32- or 64-bit values, while it was created in
some places with DW_FORM_udata, which implies emitting uleb128.
Nevertheless, these places also expected to emit U32 or U64, but just
used a misleading DWARF form. The patch updates those places to use more
appropriate DWARF forms and restricts DIELabel::SizeOf() to accept only
forms that are actually used in the LLVM codebase.

Differential Revision: https://reviews.llvm.org/D84094

3 years ago[DebugInfo] Fix a comment and a variable name. NFC.
Igor Kudrin [Mon, 3 Aug 2020 08:04:00 +0000 (15:04 +0700)]
[DebugInfo] Fix a comment and a variable name. NFC.

DebugLocListIndex keeps the index of an entry list, not the offset.

Differential Revision: https://reviews.llvm.org/D84093

3 years ago[DebugInfo] Make DIELocList::SizeOf() more explicit. NFCI.
Igor Kudrin [Mon, 3 Aug 2020 08:03:37 +0000 (15:03 +0700)]
[DebugInfo] Make DIELocList::SizeOf() more explicit. NFCI.

DIELocList is used with a limited number of DWARF forms, see the only
place where it is instantiated, DwarfCompileUnit::addLocationList().

The patch marks the unexpected execution path in DIELocList::SizeOf()
as unreachable, to reduce ambiguity.

Differential Revision: https://reviews.llvm.org/D84092

3 years ago[libunwind] Make the test depend on the libunwind explicitly.
Daniel Kiss [Mon, 3 Aug 2020 07:40:35 +0000 (09:40 +0200)]
[libunwind] Make the test depend on the libunwind explicitly.

Before this patch the `ninja check-unwind` won't rebuild the unwind library.

Reviewed By: jroelofs

Differential Revision: https://reviews.llvm.org/D85004

3 years ago[NFC] [MIR] Document the reg state flags
Djordje Todorovic [Mon, 3 Aug 2020 07:01:37 +0000 (09:01 +0200)]
[NFC] [MIR] Document the reg state flags

This patch adds documentation for the RegState enumeration.

Differential Revision: https://reviews.llvm.org/D84634

3 years ago[MLIR][SPIRV] Control attributes support for loop and selection
George Mitenkov [Mon, 3 Aug 2020 06:31:08 +0000 (09:31 +0300)]
[MLIR][SPIRV] Control attributes support for loop and selection

This patch handles loopControl and selectionControl in parsing and
printing. In order to reuse the functionality, and avoid handling cases when
`{` of the region is parsed as a dictionary attribute, `control` keyword was
introduced.`None` is a default control attribute. This functionality can be
later extended to `spv.func`.
Also, loopControl and selectionControl can now be (de)serialized.

Reviewed By: antiagainst

Differential Revision: https://reviews.llvm.org/D84175

3 years ago[CMake] Default ENABLE_X86_RELAX_RELOCATIONS to ON
Fangrui Song [Mon, 3 Aug 2020 06:05:50 +0000 (23:05 -0700)]
[CMake] Default ENABLE_X86_RELAX_RELOCATIONS to ON

This makes clang default to -Wa,-mrelax-relocations=yes, which enables
R_386_GOT32X (GNU as enables it regardless of -mrelax-relocations=) and
R_X86_64_[REX_]GOTPCRELX in MC. The produced object files require GNU ld>=2.26
to link. binutils 2.26 is considered a very old release today.

3 years ago[gn build] Port 160ff83765a
LLVM GN Syncbot [Mon, 3 Aug 2020 05:55:14 +0000 (05:55 +0000)]
[gn build] Port 160ff83765a

3 years ago[OpenMP][AMDGCN] Support OpenMP offloading for AMDGCN architecture - Part 3
Saiyedul Islam [Mon, 3 Aug 2020 05:29:48 +0000 (05:29 +0000)]
[OpenMP][AMDGCN] Support OpenMP offloading for AMDGCN architecture - Part 3

Provides AMDGCN and NVPTX specific specialization of getGPUWarpSize,
getGPUThreadID, and getGPUNumThreads methods. Adds tests for AMDGCN
codegen for these methods in generic and simd modes. Also changes the
precondition in InitTempAlloca to be slightly more permissive. Useful for
AMDGCN OpenMP codegen where allocas are created with a cast to an
address space.

Reviewed By: ABataev

Differential Revision: https://reviews.llvm.org/D84260

3 years ago[MC] Default MCAsmBackend::mayNeedRelaxation() to false
Fangrui Song [Mon, 3 Aug 2020 05:13:59 +0000 (22:13 -0700)]
[MC] Default MCAsmBackend::mayNeedRelaxation() to false

3 years ago[FLANG] Fix issues in SELECT TYPE construct when intrinsic type specification is...
compinder [Mon, 3 Aug 2020 03:27:17 +0000 (08:57 +0530)]
[FLANG] Fix issues in SELECT TYPE construct when intrinsic type specification is specified in TYPE GUARD statement.

Fix of PR46789 and PR46830.

Differential Revision: https://reviews.llvm.org/D84290

3 years ago[NFC][PowerPC] Add a multiclass for fsetcc to define them in a uniform way
QingShan Zhang [Mon, 3 Aug 2020 03:25:20 +0000 (03:25 +0000)]
[NFC][PowerPC] Add a multiclass for fsetcc to define them in a uniform way

This is a refactor patch to prepare for adding the support for strict-fsetcc
in PowerPC backend. We want to move their definition into a uniform way so that,
we could add the strict node easier.

Reviewed By: shchenz

Differential Revision: https://reviews.llvm.org/D81712

3 years ago[RISCV] eliminate the repetition declare of SDLoc DL
StephenFan [Fri, 31 Jul 2020 05:31:48 +0000 (13:31 +0800)]
[RISCV] eliminate the repetition declare of SDLoc DL

Differential revision: https://reviews.llvm.org/D85002

3 years agoReland D64327 [MC][ELF] Allow STT_SECTION referencing SHF_MERGE on REL targets
Fangrui Song [Mon, 3 Aug 2020 01:05:17 +0000 (18:05 -0700)]
Reland D64327 [MC][ELF] Allow STT_SECTION referencing SHF_MERGE on REL targets

This drops a GNU gold workaround and reverts the revert commit rL366708.

  Before binutils 2.34, gold -O2 and above did not correctly handle R_386_GOTOFF to
  SHF_MERGE|SHF_STRINGS sections: https://sourceware.org/bugzilla/show_bug.cgi?id=16794

From the original review:

  ... it reduced the size of a big ARM-32 debug image by 33%. It contained ~68M
  of relocations symbols out of total ~71M symbols (96% of symbols table was
  generated for relocations with symbol).

-Wl,-O2 (and -Wl,-O3) is so rare that we should just lower the
optimization level for LLVM_LINKER_IS_GOLD rather than pessimizing all users.

3 years ago[Bindings] Remove ipc_propagation.
Florian Hahn [Sun, 2 Aug 2020 21:36:53 +0000 (22:36 +0100)]
[Bindings] Remove ipc_propagation.

IPConstantPropagation has been removed, also remove the bindings.

3 years agoRecommit "[IPConstProp] Remove and move tests to SCCP."
Florian Hahn [Sun, 2 Aug 2020 21:04:44 +0000 (22:04 +0100)]
Recommit "[IPConstProp] Remove and move tests to SCCP."

This reverts commit 59d6e814ce0e7b40b7cc3ab136b9af2ffab9c6f8.

The cause for the revert (3 clang tests running opt -ipconstprop) was
fixed by removing those lines.

3 years ago[StackSafety, NFC] Don't insert empty objects into the map
Vitaly Buka [Sun, 2 Aug 2020 20:56:38 +0000 (13:56 -0700)]
[StackSafety, NFC] Don't insert empty objects into the map

Result should be the same but it makes generateParamAccessSummary 5x
faster.

3 years ago[Clang] Remove run-lines which use opt to run -ipconstprop.
Florian Hahn [Sun, 2 Aug 2020 20:45:42 +0000 (21:45 +0100)]
[Clang] Remove run-lines which use opt to run -ipconstprop.

ipconstprop is going to get removed and checking opt with specific
passes makes the tests more fragile.

The tests retain the important checks that !callback metadata is created
correctly.

3 years ago[lldb] [test] Fix DW_TAG_GNU_call_site-DW_AT_low_pc.s relocation
Jan Kratochvil [Sun, 2 Aug 2020 20:41:02 +0000 (22:41 +0200)]
[lldb] [test] Fix DW_TAG_GNU_call_site-DW_AT_low_pc.s relocation

I have made the DW_FORM_ref4 relative. One could also use relocated
DW_FORM_ref_addr instead.

Tested with:
        echo 'void f(){}'|clang -o 1.o -c -Wall -g -x c -;./bin/clang -o 1 1.o ../llvm-monorepo/lldb/test/Shell/SymbolFile/DWARF/DW_TAG_GNU_call_site-DW_AT_low_pc.s;./bin/lldb --no-lldbinit ./1 -o r -o 'p p' -o exit

3 years ago[X86] Use parity flag from byte test/cmp instruction for __builtin_parity when input...
Craig Topper [Sun, 2 Aug 2020 16:58:55 +0000 (09:58 -0700)]
[X86] Use parity flag from byte test/cmp instruction for __builtin_parity when input fits in 8 bits.

If the upper bits of the __builtin_parity idiom are known to be
0 we were previously emitting an xor with 0 to get the parity flag.
But we can use cmp/test instead which may expose opportunities for
load folding or combining an AND.

3 years ago[X86] Add test cases for missed opportunity to use a byte test instruction instead...
Craig Topper [Sun, 2 Aug 2020 07:20:11 +0000 (00:20 -0700)]
[X86] Add test cases for missed opportunity to use a byte test instruction instead of an xor with 0 in parity patterns.

If the input to the ctpop fits in 8 bits, we can use the parity
flag from a TEST instruction, but we're currently XORing with 0.

3 years ago[AMDGPU] Regenerate tests to fix whitespace indentations
Simon Pilgrim [Sun, 2 Aug 2020 17:11:01 +0000 (18:11 +0100)]
[AMDGPU] Regenerate tests to fix whitespace indentations

Noticed while updating D77804

3 years agoRemove debug flags from test (NFC)
Mehdi Amini [Sun, 2 Aug 2020 16:57:02 +0000 (16:57 +0000)]
Remove debug flags from test (NFC)

3 years ago[IR] Add IRBuilderBase::CreateVectorSplat(ElementCount EC) variant
Simon Pilgrim [Sun, 2 Aug 2020 15:55:16 +0000 (16:55 +0100)]
[IR] Add IRBuilderBase::CreateVectorSplat(ElementCount EC) variant

As discussed on D81500, this adds a more general ElementCount variant of the build helper and converts the (non-scalable) unsigned NumElts variant to use it internally.

3 years ago[InstSimplify] fold max (max X, Y), X --> max X, Y
Sanjay Patel [Sun, 2 Aug 2020 15:11:05 +0000 (11:11 -0400)]
[InstSimplify] fold max (max X, Y), X --> max X, Y

https://alive2.llvm.org/ce/z/VGgG3M

3 years ago[InstSimplify] add tests for max(max x,y), x) and variants; NFC
Sanjay Patel [Sun, 2 Aug 2020 14:53:51 +0000 (10:53 -0400)]
[InstSimplify] add tests for max(max x,y), x) and variants; NFC

3 years agoGlobalISel: Implement bitcast action for G_EXTRACT_VECTOR_ELEMENT
Matt Arsenault [Mon, 15 Jun 2020 15:54:49 +0000 (11:54 -0400)]
GlobalISel: Implement bitcast action for G_EXTRACT_VECTOR_ELEMENT

For AMDGPU, vectors with elements < 32 bits should be indexed in
32-bit elements and the desired bits extracted from there. For
elements > 64-bits, these should be reduce to 64/32 elements to enable
the normal dynamic indexing paths.

In the dynamic index cases, this produces shorter code most of the
time. This does immediately regress the constant index cases, but this
should be fixed once we have the most basic of shift combines.

The element size > 64 case is pretty much ported from the exisiting
DAG implementation for extract element promote. The increasing element
size case is new.

3 years agoX86InstrInfo.cpp - fix include ordering. NFCI.
Simon Pilgrim [Sun, 2 Aug 2020 14:21:52 +0000 (15:21 +0100)]
X86InstrInfo.cpp - fix include ordering. NFCI.

3 years agoUse merge null and isa<> tests into isa_and_nonnull<>. NFCI.
Simon Pilgrim [Sun, 2 Aug 2020 14:16:00 +0000 (15:16 +0100)]
Use merge null and isa<> tests into isa_and_nonnull<>. NFCI.

3 years ago[DAG] TargetLowering::expandMUL_LOHI - pass SDLoc as const&
Simon Pilgrim [Sun, 2 Aug 2020 14:31:36 +0000 (15:31 +0100)]
[DAG] TargetLowering::expandMUL_LOHI - pass SDLoc as const&

Try to be more consistent with the SDLoc param in the TargetLowering methods.

This also exposes an issue where we were passing a SDNode as a SDLoc, relying on the implicit SDLoc(SDNode) constructor.

3 years ago[DAG] TargetLowering::LowerAsmOutputForConstraint - pass SDLoc as const&
Simon Pilgrim [Sun, 2 Aug 2020 14:03:04 +0000 (15:03 +0100)]
[DAG] TargetLowering::LowerAsmOutputForConstraint - pass SDLoc as const&

Try to be more consistent with the SDLoc param in the TargetLowering methods.

3 years agoRemove unused param tag to fix Wdocumentation warning. NFC.
Simon Pilgrim [Sun, 2 Aug 2020 13:57:39 +0000 (14:57 +0100)]
Remove unused param tag to fix Wdocumentation warning. NFC.

3 years agoRevert "[Attributor] AAPotentialValues Interface"
Shinji Okumura [Sun, 2 Aug 2020 13:48:28 +0000 (22:48 +0900)]
Revert "[Attributor] AAPotentialValues Interface"

The commit cause build failure.

3 years ago[InstSimplify] Reduce code duplication in icmp of binop folds (NFC)
Nikita Popov [Sun, 2 Aug 2020 13:45:24 +0000 (15:45 +0200)]
[InstSimplify] Reduce code duplication in icmp of binop folds (NFC)

For folds where we check for the binop on both the LHS and RHS,
extract a function that expects it on the LHS and call it with
swapped order.

3 years ago[DWARFYAML][debug_aranges] Make the 'Descriptors' field optional.
Xing GUO [Sun, 2 Aug 2020 13:37:11 +0000 (21:37 +0800)]
[DWARFYAML][debug_aranges] Make the 'Descriptors' field optional.

3 years ago[X86] Use const APInt& in for-range loop to avoid unnecessary copies. NFCI.
Simon Pilgrim [Sun, 2 Aug 2020 13:32:05 +0000 (14:32 +0100)]
[X86] Use const APInt& in for-range loop to avoid unnecessary copies. NFCI.

Fixes clang-tidy warning.

3 years ago[X86] Pass SDLoc by const reference. NFCI.
Simon Pilgrim [Sun, 2 Aug 2020 13:23:00 +0000 (14:23 +0100)]
[X86] Pass SDLoc by const reference. NFCI.

3 years ago[X86] Use const APInt& in for-range loop to avoid unnecessary copies. NFCI.
Simon Pilgrim [Sun, 2 Aug 2020 13:15:15 +0000 (14:15 +0100)]
[X86] Use const APInt& in for-range loop to avoid unnecessary copies. NFCI.

Fixes clang-tidy warning.

3 years ago[X86] combineX86ShuffleChain - pull out repeated RootVT.getSizeInBits() calls. NFCI.
Simon Pilgrim [Sun, 2 Aug 2020 13:13:13 +0000 (14:13 +0100)]
[X86] combineX86ShuffleChain - pull out repeated RootVT.getSizeInBits() calls. NFCI.

3 years ago[Attributor] AAPotentialValues Interface
Shinji Okumura [Sun, 2 Aug 2020 09:37:07 +0000 (18:37 +0900)]
[Attributor] AAPotentialValues Interface

This is a split patch of D80991.
This patch introduces AAPotentialValues and its interface only.
For more detail of AAPotentialValues abstract attribute, see the original patch.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D83283

3 years ago[SCEV] Precommit tests with signed counting down loop.
Florian Hahn [Fri, 31 Jul 2020 21:10:39 +0000 (22:10 +0100)]
[SCEV] Precommit tests with signed counting down loop.

From PR46939.

3 years ago[CMake] Pass bugreport URL to standalone clang build
Michał Górny [Thu, 30 Jul 2020 22:11:40 +0000 (00:11 +0200)]
[CMake] Pass bugreport URL to standalone clang build

BUG_REPORT_URL is currently used both in LLVM and in Clang but declared
only in the latter.  This means that it's missing in standalone clang
builds and the driver ends up outputting:

  PLEASE submit a bug report to  and include [...]

(note the missing URL)

To fix this, include LLVM_PACKAGE_BUGREPORT in LLVMConfig.cmake
(similarly to how we pass PACKAGE_VERSION) and use it to fill
BUG_REPORT_URL when building clang standalone.

Differential Revision: https://reviews.llvm.org/D84987

3 years ago[X86] Improve parity idiom recognition to handle (and (truncate (ctpop X)), 1).
Craig Topper [Sun, 2 Aug 2020 05:05:10 +0000 (22:05 -0700)]
[X86] Improve parity idiom recognition to handle (and (truncate (ctpop X)), 1).

Fixes part of PR46954

3 years ago[X86] Add parity test cases for PR46954.
Craig Topper [Sun, 2 Aug 2020 04:57:24 +0000 (21:57 -0700)]
[X86] Add parity test cases for PR46954.

This adds tests cases where the parity idiom of (and (ctpop X, 1))
has a truncate in the middle.

3 years ago[HotColdSplit] Add test case for unlikely attribute in outlined function
AK [Sun, 2 Aug 2020 05:34:44 +0000 (22:34 -0700)]
[HotColdSplit] Add test case for unlikely attribute in outlined function

Authored by: rjf (Ruijie Fang)
Reviewed by: hiraditya,tejohnson,fhahn

Differential Revision: https://reviews.llvm.org/D69384

3 years agoOutline non returning functions unless a longjmp
AK [Sun, 2 Aug 2020 02:15:05 +0000 (19:15 -0700)]
Outline non returning functions unless a longjmp

__assert_fail, abort, exit etc. are cold.
TODO: outline throw

Authored by: rjf (Ruijie Fang)
Reviewed by: hiraditya,tejohnson,fhahn

Differential Revision: https://reviews.llvm.org/D69257

3 years agoUse llvm::is_contained where appropriate (NFC)
Kazu Hirata [Sun, 2 Aug 2020 04:49:38 +0000 (21:49 -0700)]
Use llvm::is_contained where appropriate (NFC)

Use llvm::is_contained where appropriate (NFC)

Reviewed By: kazu

Differential Revision: https://reviews.llvm.org/D85083

3 years ago[llvm-jitlink] Add -phony-externals option to suppress unresolved externals.
Lang Hames [Sun, 2 Aug 2020 00:44:34 +0000 (17:44 -0700)]
[llvm-jitlink] Add -phony-externals option to suppress unresolved externals.

The -phony-externals option adds a generator which explicitly defines any
otherwise unresolved externals as null. This transforms link-time
unresolved-symbol errors into potential runtime null pointer accesses
(if an unresolved external is actually accessed during execution).

This option can be useful in -harness mode to avoid having to mock a
large number of symbols that are not reachable at runtime (e.g. unused
methods referenced by a class vtable).

3 years ago[llvm-jitlink] Support promotion of ODR weak symbols in -harness mode.
Lang Hames [Sat, 1 Aug 2020 04:32:27 +0000 (21:32 -0700)]
[llvm-jitlink] Support promotion of ODR weak symbols in -harness mode.

This prevents weak symbols from being immediately dead-stripped when not
directly referenced from the test harneess, enabling use of weak symbols
from the code under test.

3 years ago[ELF] --wrap: set isUsedInRegularObj of __wrap_ only if it is defined
Fangrui Song [Sun, 2 Aug 2020 01:19:14 +0000 (18:19 -0700)]
[ELF] --wrap: set isUsedInRegularObj of __wrap_ only if it is defined

Fixes PR46169

3 years ago[msan] Respect no_huge_pages_for_shadow.
Evgenii Stepanov [Sat, 1 Aug 2020 00:18:06 +0000 (17:18 -0700)]
[msan] Respect no_huge_pages_for_shadow.

Disable huge pages in the MSan shadow region when
no_huge_pages_for_shadow == true (default).

Differential Revision: https://reviews.llvm.org/D85061

3 years agoUpdated the -I option description.
Andrei Lebedev [Sat, 1 Aug 2020 22:54:11 +0000 (15:54 -0700)]
Updated the -I option description.

3 years ago[NewPM][LVI] Abandon LVI after CVP
Nikita Popov [Thu, 30 Jul 2020 17:14:02 +0000 (19:14 +0200)]
[NewPM][LVI] Abandon LVI after CVP

As mentioned on D70376, LVI can currently cause performance issues
when running under NewPM. The problem is that, unlike the legacy
pass manager, NewPM will not immediately discard the LVI analysis
if the following pass does not need it. This is a problem, because
LVI has a high memory requirement, and mass invalidation of LVI
values is very inefficient. LVI should only be alive during passes
that actively interact with it.

This patch addresses the issue by explicitly abandoning LVI after CVP,
which gets us back to the LegacyPM behavior.

Differential Revision: https://reviews.llvm.org/D84959

3 years ago[X86] Add assembler support for {disp8} and {disp32} to control the size of displacem...
Craig Topper [Sat, 1 Aug 2020 20:25:18 +0000 (13:25 -0700)]
[X86] Add assembler support for {disp8} and {disp32} to control the size of displacement used for memory operands.

These prefixes should override the default behavior and force a larger immediate size. I don't believe gas issues any warning if you use {disp8} when a 32-bit displacement is already required. And this patch doesn't either.

This completes the {disp8} and {disp32} support from PR46650.

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D84793

3 years ago[InstSimplify] Fold abs(abs(x)) -> abs(x)
Craig Topper [Sat, 1 Aug 2020 17:01:58 +0000 (10:01 -0700)]
[InstSimplify] Fold abs(abs(x)) -> abs(x)

It's always safe to pick the earlier abs regardless of the nsw flag. We'll just lose it if it is on the outer abs but not the inner abs.

Differential Revision: https://reviews.llvm.org/D85053

3 years ago[InstCombine] Fold abs(-x) -> abs(x)
Craig Topper [Sat, 1 Aug 2020 16:59:09 +0000 (09:59 -0700)]
[InstCombine] Fold abs(-x) -> abs(x)

Negating the input doesn't matter. I left a FIXME to copy the nsw flag if its present on the neg but not on the abs.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D85055

3 years ago[PPC] Adjust run line for hardware-loops-crash.ll
Florian Hahn [Sat, 1 Aug 2020 19:58:05 +0000 (20:58 +0100)]
[PPC] Adjust run line for hardware-loops-crash.ll

Looks like %s was accidentally dropped.

3 years ago[LCSSA] Provide option for caller to clean up unused PHIs.
Florian Hahn [Sat, 1 Aug 2020 19:38:38 +0000 (20:38 +0100)]
[LCSSA] Provide option for caller to clean up unused PHIs.

formLCSSAForInstructions is used by SCEVExpander, which tracks all
inserted instructions including LCSSA phis using asserting value
handles. This means cleanup needs to happen in the caller.

Extend formLCSSAForInstructions  to take an optional pointer to a
vector. If this argument is non-nullptr, instead of directly deleting
the phis, add them to the vector, so the caller can process them.

This should address various PPC buildbot failures, including
http://lab.llvm.org:8011/builders/clang-ppc64be-linux-lnt/builds/40567

3 years ago[X86][AVX512] Fold concat(and(x,y),and(z,w)) -> and(concat(x,z),concat(y,w)) for...
Simon Pilgrim [Sat, 1 Aug 2020 19:34:39 +0000 (20:34 +0100)]
[X86][AVX512] Fold concat(and(x,y),and(z,w)) -> and(concat(x,z),concat(y,w)) for 512-bit vectors

Helps vpternlog folding on non-AVX512BW targets

3 years ago[libcxx] Add compatible with constraint tests for some shared_ptr constructors.
zoecarver [Sat, 1 Aug 2020 19:06:31 +0000 (12:06 -0700)]
[libcxx] Add compatible with constraint tests for some shared_ptr constructors.

Add shared_ptr tests where the element type and pointer type aren't 'convertible' but are 'compatible'.

Responding to a comment from D81414.

Differential Revision: https://reviews.llvm.org/D81532

3 years ago[X86][AVX] Ensure we only combine to PSHUFLW/PSHUFHW on supporting targets
Simon Pilgrim [Sat, 1 Aug 2020 13:54:21 +0000 (14:54 +0100)]
[X86][AVX] Ensure we only combine to PSHUFLW/PSHUFHW on supporting targets

Noticed while investigating combining from concatenated shuffle vectors, we weren't checking that PSHUFLW/PSHUFHW was legal - we were depending on lowering splitting to subvectors.

3 years ago[LCSSA] Use IRBuilder for PHI creation.
Florian Hahn [Sat, 1 Aug 2020 16:54:23 +0000 (17:54 +0100)]
[LCSSA] Use IRBuilder for PHI creation.

Use IRBuilder instead PHINode::Create. This should not impact the
generated code, but IRBuilder provides a way to register callbacks for
inserted instructions, which is convenient for some users.

Reviewed By: lebedev.ri

Differential Revision: https://reviews.llvm.org/D85037

3 years ago[VectorCombine] add tests for non-zero gep offsets; NFC
Sanjay Patel [Sat, 1 Aug 2020 12:38:28 +0000 (08:38 -0400)]
[VectorCombine] add tests for non-zero gep offsets; NFC

3 years ago[Attributor][NFC] Update description for the dependency graph
Luofan Chen [Sat, 1 Aug 2020 13:27:16 +0000 (21:27 +0800)]
[Attributor][NFC] Update description for the dependency graph

The word "dependency graph" is a bit misleading. When there is an
edge from node A to B (A -> B), it actually mean that B depends on
A and when the state of A is updated, B should also be updated. So
I update the comment to make the description clearer.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D85065

3 years ago[ARM] Distribute post-inc for Thumb2 sign/zero extending loads/stores
David Green [Sat, 1 Aug 2020 13:01:18 +0000 (14:01 +0100)]
[ARM] Distribute post-inc for Thumb2 sign/zero extending loads/stores

This adds sign/zero extending scalar loads/stores to the MVE
instructions added in D77813, allowing us to create up more post-inc
instructions. These are comparatively simple, compared to LDR/STR (which
may be better turned into an LDRD/LDM), but still require some additions
over MVE instructions. Because there are i12 and i8 variants of the
offset loads/stores dealing with different signs, we may need to convert
an i12 address to a i8 negative instruction. t2LDRBi12 can also be
shrunk to a tLDRi under the right conditions, so we need to be careful
with codesize too.

Differential Revision: https://reviews.llvm.org/D78625

3 years ago[mlir][Vector] Simplify code a bit. NFCI.
Benjamin Kramer [Sat, 1 Aug 2020 12:48:42 +0000 (14:48 +0200)]
[mlir][Vector] Simplify code a bit. NFCI.

3 years ago[InstSimplify] simplify abs if operand is known non-negative
Sanjay Patel [Sat, 1 Aug 2020 11:46:23 +0000 (07:46 -0400)]
[InstSimplify] simplify abs if operand is known non-negative

abs() should be rare enough that using value tracking is not going
to be a compile-time cost burden, so use it to reduce a variety of
potential patterns. We do this in DAGCombiner too.

Differential Revision: https://reviews.llvm.org/D85043

3 years ago[InstSimplify] add abs test with assume; NFC
Sanjay Patel [Sat, 1 Aug 2020 11:44:33 +0000 (07:44 -0400)]
[InstSimplify] add abs test with assume; NFC

3 years ago[X86][AVX] Extend v2f64 BROADCAST(LOAD) -> BROADCAST_LOAD to v2i64/v4f32/v4i32
Simon Pilgrim [Sat, 1 Aug 2020 11:28:10 +0000 (12:28 +0100)]
[X86][AVX] Extend v2f64 BROADCAST(LOAD) -> BROADCAST_LOAD to v2i64/v4f32/v4i32

Minor precursor fix for D66004, but helps the SSE41 tests as well as they run with -disable-peephole

3 years ago[clang-tidy][NFC] Small refactor
Nathan James [Sat, 1 Aug 2020 10:04:29 +0000 (11:04 +0100)]
[clang-tidy][NFC] Small refactor

3 years ago[MachineVerifier] Refactor calcRegsPassed. NFC
Evgeny Leviant [Sat, 1 Aug 2020 09:58:52 +0000 (12:58 +0300)]
[MachineVerifier] Refactor calcRegsPassed. NFC

Patch improves performance of verify-machineinstrs pass up to 10x.
Differential revision: https://reviews.llvm.org/D84105

3 years ago[clang-tidy] readability-identifier-naming checks configs for included files
Nathan James [Sat, 1 Aug 2020 09:35:13 +0000 (10:35 +0100)]
[clang-tidy] readability-identifier-naming checks configs for included files

When checking for the style of a decl that isn't in the main file, the check will now search for the configuration that the included files uses to gather the style for its decls.

This can be useful to silence warnings in header files that follow a different naming convention without using header-filter to silence all warnings(even from other checks) in the header file.

Reviewed By: aaron.ballman, gribozavr2

Differential Revision: https://reviews.llvm.org/D84814

3 years ago[X86] Refactor the broadcast and load folding in tryVPTESTM to reduce some code.
Craig Topper [Sat, 1 Aug 2020 06:10:47 +0000 (23:10 -0700)]
[X86] Refactor the broadcast and load folding in tryVPTESTM to reduce some code.

Now we try to load and broadcast together for operand 1. Followed
by load and broadcast for operand 1. Previously we tried load
operand 1, load operand 1, broadcast operand 0, broadcast operand 1.

Now we have a single helper that tries load and broadcast for
one operand that we can just call twice.

3 years ago[SCEV] don't query getSCEV() for incomplete phis
Chen Zheng [Sat, 1 Aug 2020 05:28:15 +0000 (01:28 -0400)]
[SCEV] don't query getSCEV() for incomplete phis

querying getSCEV() for incomplete phis leads to wrong cache value in `ExprToIVMap`,
because incomplete phis may be simplified to same value before get SCEV expression.

Reviewed By: lebedev.ri, mkazantsev

Differential Revision: https://reviews.llvm.org/D77560

3 years ago[X86] Use TargetLowering::getRegClassFor to simplify some code in tryVPTESTM. NFCI
Craig Topper [Sat, 1 Aug 2020 04:38:59 +0000 (21:38 -0700)]
[X86] Use TargetLowering::getRegClassFor to simplify some code in tryVPTESTM. NFCI

3 years agoPowerPC: Don't lower SELECT_CC to PPCISD::FSEL on SPE
Justin Hibbits [Thu, 9 Apr 2020 02:29:40 +0000 (21:29 -0500)]
PowerPC: Don't lower SELECT_CC to PPCISD::FSEL on SPE

SPE doesn't have a fsel instruction, so don't try to lower to it.

This fixes a "Cannot select: tN: f64 = PPCISD::FSEL tX, tY, tZ" error.

Reviewed By: #powerpc, lkail
Differential Revision: https://reviews.llvm.org/D77773

3 years agoPowerPC: Fix SPE extloadf32 handling.
Justin Hibbits [Thu, 16 Apr 2020 14:42:39 +0000 (09:42 -0500)]
PowerPC: Fix SPE extloadf32 handling.

The patterns were incorrect copies from the FPU code, and are
unnecessary, since there's no extended load for SPE.  Just let LLVM
itself do the work by marking it expand.

Reviewed By: #powerpc, lkail
Differential Revision: https://reviews.llvm.org/D78670

3 years ago[VE] Change calling convention to follow ABI
Kazushi (Jam) Marukawa [Sat, 11 Jul 2020 07:01:13 +0000 (16:01 +0900)]
[VE] Change calling convention to follow ABI

Change to expand all arguments and return values to i64 to follow ABI.
Update regression tests also.

Reviewed By: simoll

Differential Revision: https://reviews.llvm.org/D84581

3 years ago[clang-tidy][NFC] Added convienence methods for getting optional options
Nathan James [Sat, 1 Aug 2020 00:45:33 +0000 (01:45 +0100)]
[clang-tidy][NFC] Added convienence methods for getting optional options

These methods abstract away Error handling when trying to read options that can't be parsed by logging the error automatically and returning None.

Reviewed By: gribozavr2

Differential Revision: https://reviews.llvm.org/D84812

3 years ago[AArch64][SVE] Allow vector of pointers as legal type for masked load/store.
Huihui Zhang [Sat, 1 Aug 2020 00:08:17 +0000 (17:08 -0700)]
[AArch64][SVE] Allow vector of pointers as legal type for masked load/store.

Refer to LangRef http://llvm.org/docs/LangRef.html#llvm-masked-load-intrinsics
'llvm.masked.load/store.*’ intrinsics are overloaded intrinsic, which allow the
load/store data to be a vector of any integer, floating-point or pointer data type.

Therefore, allow pointer data type when checking 'isLegalMaskedLoadStore()'.

Reviewed By: paulwalker-arm

Differential Revision: https://reviews.llvm.org/D85045

3 years agoDon't crash if we deserialize a pack expansion type whose pattern
Richard Smith [Sat, 1 Aug 2020 00:07:36 +0000 (17:07 -0700)]
Don't crash if we deserialize a pack expansion type whose pattern
contains no packs.

Fixes a regression from 740a164dec483225cbd02ab6c82199e2747ffacb.

3 years ago[X86] Simplify vpternlog immediate selection.
Craig Topper [Sat, 1 Aug 2020 00:07:20 +0000 (17:07 -0700)]
[X86] Simplify vpternlog immediate selection.

Rather than hardcoding immediate values for 12 different combinations
in a nested pair of switches, we can perform the matched logic
operation on 3 magic constants to calculate the immediate.

Special thanks to this tweet https://twitter.com/rygorous/status/1187034321992871936
for making me realize I could do this.

3 years ago[libc] [obvious] In strrchr, remove cast to unsigned char before
cgyurgyik [Sat, 1 Aug 2020 00:14:34 +0000 (20:14 -0400)]
[libc] [obvious] In strrchr, remove cast to unsigned char before
comparison.

3 years ago[libc] [obvious] Fix strchr and strrchr tests so that constness is
cgyurgyik [Fri, 31 Jul 2020 23:59:29 +0000 (19:59 -0400)]
[libc] [obvious] Fix strchr and strrchr tests so that constness is
actually verified.

3 years ago[compiler-rt][Darwin] Fix linker errors for check-asan
Julian Lettner [Fri, 31 Jul 2020 23:09:13 +0000 (16:09 -0700)]
[compiler-rt][Darwin] Fix linker errors for check-asan

A recent change broke `ninja check-asan` on Darwin by causing an error
during linking of ASan unit tests [1].

Move the addition of `-ObjC` compiler flag outside of the new
`if(COMPILER_RT_STANDALONE_BUILD)` block.  It doesn't add any global
flags (e.g, `${CMAKE_CXX_FLAGS}`) and the decision to add is based
solely on source paths (`${source_rpath}`).

[1] 8b2fcc42b895, https://reviews.llvm.org/D84466

Differential Revision: https://reviews.llvm.org/D85057

3 years agoSupport experimental v extension v0.9.
Hsiangkai Wang [Thu, 4 Jun 2020 19:34:01 +0000 (03:34 +0800)]
Support experimental v extension v0.9.

Differential revision: https://reviews.llvm.org/D81213

3 years agoUpgrade MC to v0.9.
Hsiangkai Wang [Thu, 21 May 2020 09:30:20 +0000 (17:30 +0800)]
Upgrade MC to v0.9.

Differential revision: https://reviews.llvm.org/D80802

3 years ago[ValueTracking] Improve llvm.abs handling in computeKnownBits.
Craig Topper [Fri, 31 Jul 2020 22:46:12 +0000 (15:46 -0700)]
[ValueTracking] Improve llvm.abs handling in computeKnownBits.

Add the optimizations we have in the SelectionDAG version.
Known non-negative copies all known bits. Any known one other than
the sign bit makes result non-negative.

Differential Revision: https://reviews.llvm.org/D85000

3 years ago[OpenMP] Fixed the issue that target memory deallocation might be called when they...
Shilei Tian [Fri, 31 Jul 2020 22:54:09 +0000 (18:54 -0400)]
[OpenMP] Fixed the issue that target memory deallocation might be called when they're being used

This patch fixed the issue that target memory might be deallocated when
they're still being used or before they're used.

Reviewed By: ye-luo

Differential Revision: https://reviews.llvm.org/D84996

3 years ago[InstSimplify] add tests for abs intrinsic; NFC
Sanjay Patel [Fri, 31 Jul 2020 20:35:41 +0000 (16:35 -0400)]
[InstSimplify] add tests for abs intrinsic; NFC

3 years agoUpdated the -I option description.
Andrei Lebedev [Fri, 31 Jul 2020 22:39:12 +0000 (15:39 -0700)]
Updated the -I option description.

3 years agoConvert to early exit (NFC)
Adrian Prantl [Fri, 31 Jul 2020 20:30:59 +0000 (13:30 -0700)]
Convert to early exit (NFC)

3 years ago[mlir] Add shape.with_shape op
Jacques Pienaar [Fri, 31 Jul 2020 21:46:48 +0000 (14:46 -0700)]
[mlir] Add shape.with_shape op

This is an operation that can returns a new ValueShape with a different shape. Useful for composing shape function calls and reusing existing shape transfer functions.

Just adding the op in this change.

Differential Revision: https://reviews.llvm.org/D84217