17 months ago[NFC] Fix pattern name.
Sean Silva [Thu, 26 Nov 2020 00:10:34 +0000 (16:10 -0800)]
[NFC] Fix pattern name.

It still had the old name from before ElementwiseMappable was added.

17 months ago[RISCV] Add an implementation of isFMAFasterThanFMulAndFAdd
Craig Topper [Wed, 25 Nov 2020 23:07:34 +0000 (15:07 -0800)]
[RISCV] Add an implementation of isFMAFasterThanFMulAndFAdd

Start with an assumption that FMA is faster than Fmul+FAdd. If thats not true
on some particular implementation we can add a tuning parameter in the future.

I've update the fmuladd test cases and added new test cases for fast math flag
based contraction.

Differential Revision: https://reviews.llvm.org/D91987

17 months ago[SelectionDAGBuilder] Add SPF_NABS support to visitSelect
Craig Topper [Wed, 25 Nov 2020 22:54:26 +0000 (14:54 -0800)]
[SelectionDAGBuilder] Add SPF_NABS support to visitSelect

We currently don't match this which limits the effectiveness of D91120 until
InstCombine starts canonicalizing to llvm.abs. This should be easy to remove
if/when we remove the SPF_ABS handling.

Differential Revision: https://reviews.llvm.org/D92118

17 months agogithub actions: Use llvmbot token for main branch sync
Tom Stellard [Wed, 25 Nov 2020 22:43:18 +0000 (14:43 -0800)]
github actions: Use llvmbot token for main branch sync

The default github actions token cannot push to restricted branches, so we need to use a token from the llvmbot user.

17 months agoDebugInfo: remove unnecessary mtriple from test/DebugInfo/X86/abstract_origin.ll
David Blaikie [Wed, 25 Nov 2020 21:34:15 +0000 (13:34 -0800)]
DebugInfo: remove unnecessary mtriple from test/DebugInfo/X86/abstract_origin.ll

The test already specifies a triple in the IR itself.

Based on post-commit feedback from Luís Marques.

17 months ago[libc++] NFC: Reindent non-lockfree-atomics feature
Louis Dionne [Wed, 25 Nov 2020 21:14:28 +0000 (16:14 -0500)]
[libc++] NFC: Reindent non-lockfree-atomics feature

17 months ago[ms] [llvm-ml] Implement the expression expansion operator
Eric Astor [Wed, 25 Nov 2020 20:37:00 +0000 (15:37 -0500)]
[ms] [llvm-ml] Implement the expression expansion operator

In text-item contexts, %expr expands to a string containing the results of evaluating `expr`.

Reviewed By: thakis

Differential Revision: https://reviews.llvm.org/D89736

17 months ago[mlir] Add build configuration for Quant unittest
Marius Brehler [Thu, 19 Nov 2020 22:21:00 +0000 (22:21 +0000)]
[mlir] Add build configuration for Quant unittest

17 months ago[RISCV] Add test cases to check that we use (smax X, (neg X)) for abs with Zbb extension.
Craig Topper [Wed, 25 Nov 2020 20:40:25 +0000 (12:40 -0800)]
[RISCV] Add test cases to check that we use (smax X, (neg X)) for abs with Zbb extension.

17 months ago[RISCV] Make SMIN/SMAX/UMIN/UMAX legal with Zbb extension.
Craig Topper [Wed, 25 Nov 2020 20:10:40 +0000 (12:10 -0800)]
[RISCV] Make SMIN/SMAX/UMIN/UMAX legal with Zbb extension.

This is the logically correct thing to do. But it generates worse
code for i32 umin/umax on the rv64 due to type legalize requesting
zext even though the arguments are sext. Maybe we can teach type
legalizer to use sext for umin/umax for RISCV.

It's also producing possibly worse code on i64 on RV32 since we
still end up with selects that become branches. But this seems
like something we could improve in type legalization or DAG combine.

Hopefully this makes D92095 work for RISCV with Zbb.

17 months ago[libc++] Factor out common logic for calling aligned allocation
Louis Dionne [Thu, 12 Nov 2020 20:14:33 +0000 (15:14 -0500)]
[libc++] Factor out common logic for calling aligned allocation

There were a couple of places where we needed to call the underlying
platform's aligned allocation/deallocation function. Instead of having
the same logic all over the place, extract the logic into a pair of
helper functions __libcpp_aligned_alloc and __libcpp_aligned_free.

The code in libcxxabi/src/fallback_malloc.cpp looks like it could be
simplified after this change -- I purposefully did not simplify it
further to keep this change as straightforward as possible, since it
is touching very important parts of the library.

Also, the changes in libcxx/src/new.cpp and libcxxabi/src/stdlib_new_delete.cpp
are basically the same -- I just kept both source files in sync.

The underlying reason for this refactoring is to make it easier to support
platforms that provide aligned allocation through C11's aligned_alloc
function instead of posix_memalign. After this change, we'll only have
to add support for that in a single place.

Differential Revision: https://reviews.llvm.org/D91379

17 months ago[clangd] Track deprecation of 'member' semantic token type in LSP.
Sam McCall [Wed, 25 Nov 2020 20:31:18 +0000 (21:31 +0100)]
[clangd] Track deprecation of 'member' semantic token type in LSP.

17 months ago[MLIR][Affine] Add custom builders for AffineVectorLoadOp/AffineVectorStoreOp
Frank Laub [Wed, 25 Nov 2020 20:22:01 +0000 (20:22 +0000)]
[MLIR][Affine] Add custom builders for AffineVectorLoadOp/AffineVectorStoreOp

Adding missing custom builders for AffineVectorLoadOp & AffineVectorStoreOp. In practice, it is difficult to correctly construct these ops without these builders (because the AffineMap is not included at construction time).

Differential Revision: https://reviews.llvm.org/D86380

17 months ago[MS] Add more 128bit cmpxchg intrinsics for AArch64
Reid Kleckner [Tue, 24 Nov 2020 19:50:33 +0000 (11:50 -0800)]
[MS] Add more 128bit cmpxchg intrinsics for AArch64

The MSVC STL for requires this on ARM64.
Requested in https://llvm.org/pr47099

Depends on D92061

Differential Revision: https://reviews.llvm.org/D92062

17 months ago[MS] Fix double evaluation of MSVC builtin arguments
Reid Kleckner [Tue, 24 Nov 2020 22:48:05 +0000 (14:48 -0800)]
[MS] Fix double evaluation of MSVC builtin arguments

This code got quite twisted because we consider some MSVC builtins to be
target agnostic, and some to be target specific. Target specific
intrinsics have a pattern of doing up-front argument evaluation, while
general intrinsics do not evaluate their arguments up front. As we tried
to share codepaths between the target-specific and target-agnostic
handling, we ended up doing double evaluation.

Instead, have each target handle MSVC intrinsics consistently before up
front argument evaluation. This requires passing less data around and is
more consistent with target independent intrinsic handling.

See D50979 for past examples of this bug. I noticed this while looking
into adding some more intrinsics.

Differential Revision: https://reviews.llvm.org/D92061

17 months ago[Hexagon] Add support for ISD::SMAX/SMIN/UMAX/UMIN instead of custom dag patterns
Simon Pilgrim [Wed, 25 Nov 2020 19:00:33 +0000 (19:00 +0000)]
[Hexagon] Add support for ISD::SMAX/SMIN/UMAX/UMIN instead of custom dag patterns

This should handle the basic integer min/max handling - the HVX ops are still TODO.

This is some necessary cleanup work for min/max ops to eventually help us move the add/sub sat patterns into DAGCombine - D91876.

Differential Revision: https://reviews.llvm.org/D92112

17 months ago[RISCV] Add test cases showing that we don't recognize the select form of NABS in...
Craig Topper [Wed, 25 Nov 2020 18:42:08 +0000 (10:42 -0800)]
[RISCV] Add test cases showing that we don't recognize the select form of NABS in SelectionDAGBuilder so we end up with branches. NFC

There's a FIXME that it should produce (sub 0, (abs)).

17 months agoRemove static function unused after cf1c774.
Paul Robinson [Wed, 25 Nov 2020 18:39:09 +0000 (13:39 -0500)]
Remove static function unused after cf1c774.

Caused some -Werror bot failures.

17 months ago[CostModel][X86] Refresh ISD::ABS costs
Simon Pilgrim [Wed, 25 Nov 2020 18:39:56 +0000 (18:39 +0000)]
[CostModel][X86] Refresh ISD::ABS costs

Update costs now that D92095 and D92102 have tweaked the SSE2 implementation

The SSE42 BLENDVPD cost can actually be used on SSE41 as we don't attempt to generate PCMPGT anymore

Add scalar i16/i32/i64 costs as we can do this cheaply with CMOV

17 months ago[gn build] Port 73fdd998701
LLVM GN Syncbot [Wed, 25 Nov 2020 18:35:53 +0000 (18:35 +0000)]
[gn build] Port 73fdd998701

17 months ago[clangd] Implement clang-tidy options from config
Nathan James [Wed, 25 Nov 2020 18:35:34 +0000 (18:35 +0000)]
[clangd] Implement clang-tidy options from config

Added some new ClangTidyOptionsProvider like classes designed for clangd work flow.
These providers are designed to source the options on the worker thread but in a thread safe manner.
This is done through making the options getter take a pointer to the filesystem used by the worker thread which natuarally is from a ThreadsafeFS.
Internal caching in the providers is also guarded.

The providers don't inherit from `ClangTidyOptionsProvider` instead they share a base class which is able to create a provider for the `ClangTidyContext` using a specific FileSystem.
This approach means one provider can be used for multiple contexts even though `ClangTidyContext` owns its provider.

Depends on D90531

Reviewed By: sammccall

Differential Revision: https://reviews.llvm.org/D91029

17 months ago[DAG] Legalize abs(x) -> umin(x,sub(0,x)) iff umin/sub are legal
Simon Pilgrim [Wed, 25 Nov 2020 18:05:41 +0000 (18:05 +0000)]
[DAG] Legalize abs(x) -> umin(x,sub(0,x)) iff umin/sub are legal

If umin() is legal, this is likely to result in smaller codegen expansion for abs(x) than the xor(add,ashr) method.

Followup to D92095

Alive2: https://alive2.llvm.org/ce/z/8nuX6s  https://alive2.llvm.org/ce/z/q2hB9w

17 months ago[FastISel] Flush local value map on ever instruction
Paul Robinson [Wed, 18 Nov 2020 21:27:14 +0000 (16:27 -0500)]
[FastISel] Flush local value map on ever instruction

Local values are constants or addresses that can't be folded into
the instruction that uses them. FastISel materializes these in a
"local value" area that always dominates the current insertion
point, to try to avoid materializing these values more than once
(per block).

https://reviews.llvm.org/D43093 added code to sink these local
value instructions to their first use, which has two beneficial
effects. One, it is likely to avoid some unnecessary spills and
reloads; two, it allows us to attach the debug location of the
user to the local value instruction. The latter effect can
improve the debugging experience for debuggers with a "set next
statement" feature, such as the Visual Studio debugger and PS4
debugger, because instructions to set up constants for a given
statement will be associated with the appropriate source line.

There are also some constants (primarily addresses) that could be
produced by no-op casts or GEP instructions; the main difference
from "local value" instructions is that these are values from
separate IR instructions, and therefore could have multiple users
across multiple basic blocks. D43093 avoided sinking these, even
though they were emitted to the same "local value" area as the
other instructions. The patch comment for D43093 states:

  Local values may also be used by no-op casts, which adds the
  register to the RegFixups table. Without reversing the RegFixups
  map direction, we don't have enough information to sink these

This patch undoes most of D43093, and instead flushes the local
value map after(*) every IR instruction, using that instruction's
debug location. This avoids sometimes incorrect locations used
previously, and emits instructions in a more natural order.

This does mean materialized values are not re-used across IR
instruction boundaries; however, only about 5% of those values
were reused in an experimental self-build of clang.

(*) Actually, just prior to the next instruction. It seems like
it would be cleaner the other way, but I was having trouble
getting that to work.

Differential Revision: https://reviews.llvm.org/D91734

17 months ago[RISCV] Custom type legalize i32 fshl/fshr on RV64 with Zbt.
Craig Topper [Wed, 25 Nov 2020 17:43:16 +0000 (09:43 -0800)]
[RISCV] Custom type legalize i32 fshl/fshr on RV64 with Zbt.

This adds custom opcodes for FSLW/FSRW so we can type legalize
fshl/fshr without needing to match a sign_extend_inreg.

I've used the operand order from fshl/fshr to make the isel
pattern similar to the non-W form. It was also hard to decide
another order since the register instruction has the shift amount
as the second operand, but the immediate instruction has it as
the third operand.

Differential Revision: https://reviews.llvm.org/D91479

17 months ago[OpenMP][Docs] Add more content, call coordinates, FAQ entries, links
Johannes Doerfert [Wed, 25 Nov 2020 17:49:30 +0000 (11:49 -0600)]
[OpenMP][Docs] Add more content, call coordinates, FAQ entries, links

17 months ago[Flang][Docs] Update call information and add two more calls
Johannes Doerfert [Wed, 25 Nov 2020 15:59:50 +0000 (09:59 -0600)]
[Flang][Docs] Update call information and add two more calls

Call information have been updated and the OpenMP in LLVM as well as
Classic Flang call have been added.

17 months ago[OpenMP] libomp: fix non-X86, non-AARCH64 builds
AndreyChurbanov [Wed, 25 Nov 2020 17:40:23 +0000 (20:40 +0300)]
[OpenMP] libomp: fix non-X86, non-AARCH64 builds

Commit https://reviews.llvm.org/rG7b5254223acbf2ef9cd278070c5a84ab278d7e5f
broke the build for some architectures, because macro KMP_PREFIX_UNDERSCORE
was defined only for x86, x86_64 and aarch64. This patch defines it for other
architectures (as a no-op).

Differential Revision: https://reviews.llvm.org/D92027

17 months ago[OpenMP][OMPT] Introduce a guard to handle OMPT return address
Joachim Protze [Wed, 18 Nov 2020 11:49:19 +0000 (12:49 +0100)]
[OpenMP][OMPT] Introduce a guard to handle OMPT return address

This is an alternative approach to address inconsistencies pointed out in: D90078
This patch makes sure that the return address is reset, when leaving the scope.
In some cases, I had to move the macro out of an if-statement to have it in the
right scope, in some cases I added an additional block to restrict the scope.

This patch does not handle inconsistencies, which might occur if the return
address is still set when we call into the application.

Test case (repeated_calls.c) provided by @hbae

Differential Revision: https://reviews.llvm.org/D91692

17 months ago[OpenMP][OMPT] Implement verbose tool loading
Isabel Thärigen [Tue, 27 Oct 2020 13:05:28 +0000 (14:05 +0100)]
[OpenMP][OMPT] Implement verbose tool loading

OpenMP 5.1 introduces the new env variable
OMP_TOOL_VERBOSE_INIT=(disabled|stdout|stderr|<filename>) to enable verbose
loading and initialization of OMPT tools.
This env variable helps to understand the cause when loading of a tool fails
(e.g., undefined symbols or dependency not in LD_LIBRARY_PATH)

Tests for this patch are integrated into the different existing tool loading
tests, making these tests more verbose. An Archer specific verbose test is
integrated into an existing Archer test.

Patch prepared by: Isabel Thärigen

Differential Revision: https://reviews.llvm.org/D91464

17 months ago[gn build] Download prebuilt mac-arm64 binary now that it exists
Nico Weber [Wed, 25 Nov 2020 17:08:10 +0000 (12:08 -0500)]
[gn build] Download prebuilt mac-arm64 binary now that it exists

Sadly requires an ugly workaround for an ugly bug, but still nicer than
building locally.

17 months ago[clang][test] Fix prefix operator++ signature in iterators
Endre Fülöp [Fri, 16 Oct 2020 07:37:42 +0000 (09:37 +0200)]
[clang][test] Fix prefix operator++ signature in iterators

Prefix operator++ should return the iterator incremented by reference.

Differential Revision: https://reviews.llvm.org/D89528

17 months ago[ELF] Rename adjustRelaxExpr to adjustTlsExpr and delete the unused `data` parameter...
Fangrui Song [Wed, 25 Nov 2020 17:00:55 +0000 (09:00 -0800)]
[ELF] Rename adjustRelaxExpr to adjustTlsExpr and delete the unused `data` parameter. NFC

Reviewed By: psmith

Differential Revision: https://reviews.llvm.org/D91995

17 months ago[SVE] Fix TypeSize warning in RuntimePointerChecking::insert
Joe Ellis [Tue, 24 Nov 2020 10:08:04 +0000 (10:08 +0000)]
[SVE] Fix TypeSize warning in RuntimePointerChecking::insert

The TypeSize warning would occur because RuntimePointerChecking::insert
was not scalable vector aware. The fix is to use
ScalarEvolution::getSizeOfExpr to grab the size of types.

Differential Revision: https://reviews.llvm.org/D90171

17 months ago[ELF] Add TargetInfo::adjustGotPcExpr for `R_GOT_PC` relaxations. NFC
Fangrui Song [Wed, 25 Nov 2020 16:43:26 +0000 (08:43 -0800)]
[ELF] Add TargetInfo::adjustGotPcExpr for `R_GOT_PC` relaxations. NFC

With this change, `TargetInfo::adjustRelaxExpr` is only related to TLS
relaxations and a subsequent clean-up can delete the `data` parameter.

Differential Revision: https://reviews.llvm.org/D92079

17 months ago[AMDGPU] Actually fully update opt-pipeline.ll test to account for -loop-idiom vs...
Roman Lebedev [Wed, 25 Nov 2020 16:38:51 +0000 (19:38 +0300)]
[AMDGPU] Actually fully update opt-pipeline.ll test to account for -loop-idiom vs -indvars switch

17 months ago[MC][WebAssembly] Only emit indirect function table import if needed
Andy Wingo [Wed, 25 Nov 2020 16:31:05 +0000 (08:31 -0800)]
[MC][WebAssembly] Only emit indirect function table import if needed

The indirect function table, synthesized by the linker, is needed if and
only if there are TABLE_INDEX relocs.

Differential Revision: https://reviews.llvm.org/D91637

17 months ago[PassManager] Run Induction Variable Simplification pass *after* Recognize loop idiom...
Roman Lebedev [Wed, 25 Nov 2020 16:17:25 +0000 (19:17 +0300)]
[PassManager] Run Induction Variable Simplification pass *after* Recognize loop idioms pass, not before

Currently, `-indvars` runs first, and then immediately after `-loop-idiom` does.
I'm not really sure if `-loop-idiom` requires `-indvars` to run beforehand,
but i'm *very* sure that `-indvars` requires `-loop-idiom` to run afterwards,
as it can be seen in the phase-ordering test.

LoopIdiom runs on two types of loops: countable ones, and uncountable ones.
For uncountable ones, IndVars obviously didn't make any change to them,
since they are uncountable, so for them the order should be irrelevant.
For countable ones, well, they should have been countable before IndVars
for IndVars to make any change to them, and since SCEV is used on them,
it shouldn't matter if IndVars have already canonicalized them.
So i don't really see why we'd want the current ordering.

Should this cause issues, it will give us a reproducer test case
that shows flaws in this logic, and we then could adjust accordingly.

While this is quite likely beneficial in-the-wild already,
it's a required part for the full motivational pattern
behind `left-shift-until-bittest` loop idiom (D91038).

Reviewed By: dmgreen

Differential Revision: https://reviews.llvm.org/D91800

17 months ago[WebAssembly] Factor out WasmTableType in binary format
Andy Wingo [Wed, 25 Nov 2020 15:54:31 +0000 (07:54 -0800)]
[WebAssembly] Factor out WasmTableType in binary format

This commit factors out a WasmTableType definition from WasmTable, as is
the case for WasmGlobal and other data types.  Also add support for
extracting the SymbolName for a table from the linking section's symbol

Differential Revision: https://reviews.llvm.org/D91849

17 months ago[AMDGPU] Emit stack frame size in metadata
Sebastian Neubauer [Fri, 23 Oct 2020 09:20:20 +0000 (11:20 +0200)]
[AMDGPU] Emit stack frame size in metadata

Add .shader_functions to pal metadata, which contains the stack frame
size for all non-entry-point functions.

Differential Revision: https://reviews.llvm.org/D90036

17 months ago[flang][openacc] Semantic check for cache directive
Valentin Clement [Wed, 25 Nov 2020 15:28:02 +0000 (10:28 -0500)]
[flang][openacc] Semantic check for cache directive

Add semantic check for the cache directive. According to section 2.10 from the specification:
A var in a cache directive must be a single array element or a simple subarray.

Reviewed By: kiranchandramohan

Differential Revision: https://reviews.llvm.org/D90184

17 months ago[DAG] Legalize abs(x) -> smax(x,sub(0,x)) iff smax/sub are legal
Simon Pilgrim [Wed, 25 Nov 2020 15:03:03 +0000 (15:03 +0000)]
[DAG] Legalize abs(x) -> smax(x,sub(0,x)) iff smax/sub are legal

If smax() is legal, this is likely to result in smaller codegen expansion for abs(x) than the xor(add,ashr) method.

This is also what PowerPC has been doing for its abs implementation, so it lets us get rid of a load of custom lowering code there (and which was never updated when they added smax lowering).

Alive2: https://alive2.llvm.org/ce/z/xRk3cD

Differential Revision: https://reviews.llvm.org/D92095

17 months ago[PowerPC] Regenerate vec_select.ll tests and add <1 x i128> test case
Simon Pilgrim [Wed, 25 Nov 2020 11:50:16 +0000 (11:50 +0000)]
[PowerPC] Regenerate vec_select.ll tests and add <1 x i128> test case

17 months agoFix case mismatch between definition and declaration
Guillaume Chatelet [Wed, 25 Nov 2020 14:23:31 +0000 (14:23 +0000)]
Fix case mismatch between definition and declaration

17 months ago[libc++] ADL-proof <variant> by adding _VSTD:: qualification on calls.
Arthur O'Dwyer [Tue, 24 Nov 2020 14:59:26 +0000 (09:59 -0500)]
[libc++] ADL-proof <variant> by adding _VSTD:: qualification on calls.

Differential Revision: https://reviews.llvm.org/D92036

17 months ago[clangd] PopulateSwitch: disable on dependent enums.
Adam Czachorowski [Tue, 24 Nov 2020 19:47:37 +0000 (20:47 +0100)]
[clangd] PopulateSwitch: disable on dependent enums.

If the enum is a dependent type, we would crash somewhere in
getIntWidth(). -Wswitch diagnostic doesn't work on dependent enums

Differential Revision: https://reviews.llvm.org/D92051

17 months ago[LAA] NFC: Rename [get]MaxSafeRegisterWidth -> [get]MaxSafeVectorWidthInBits
Cullen Rhodes [Wed, 18 Nov 2020 18:13:08 +0000 (18:13 +0000)]
[LAA] NFC: Rename [get]MaxSafeRegisterWidth -> [get]MaxSafeVectorWidthInBits

MaxSafeRegisterWidth is a misnomer since it actually returns the maximum
safe vector width. Register suggests it relates directly to a physical
register where it could be a vector spanning one or more physical

Reviewed By: sdesmalen

Differential Revision: https://reviews.llvm.org/D91727

17 months agoRevert "[libc++] P1645 constexpr for <numeric>"
Mark de Wever [Wed, 25 Nov 2020 12:46:08 +0000 (13:46 +0100)]
Revert "[libc++] P1645 constexpr for <numeric>"

This reverts commit eb9b063539c34d0d4dd14e8516eeb77bb8b9e4bd.

The commit fails to build on build bots using LLVM 8.

17 months agoFix a typo in the documentation to unbreak the sphinx builder.
Aaron Ballman [Wed, 25 Nov 2020 12:33:38 +0000 (07:33 -0500)]
Fix a typo in the documentation to unbreak the sphinx builder.

17 months ago[llvm-readelf/obj] - Stop using `reportError` when dumping notes.
Georgii Rymar [Tue, 24 Nov 2020 11:54:48 +0000 (14:54 +0300)]
[llvm-readelf/obj] - Stop using `reportError` when dumping notes.

This starts using `reportUniqueWarnings` instead of `reportError`
in the code that is responsible for dumping notes.

Differential revision: https://reviews.llvm.org/D92021

17 months ago[libc++] P1645 constexpr for <numeric>
Mark de Wever [Tue, 24 Nov 2020 13:55:55 +0000 (14:55 +0100)]
[libc++] P1645 constexpr for <numeric>

Implements P1645: constexpr for <numeric> algorithms

Reviewed By: ldionne, #libc

Differential Revision: https://reviews.llvm.org/D90569

17 months ago[clangd] Use TimePoint<> instead of system_clock::time_point, it does matter after...
Sam McCall [Wed, 25 Nov 2020 11:49:18 +0000 (12:49 +0100)]
[clangd] Use TimePoint<> instead of system_clock::time_point, it does matter after all.

17 months ago[ARM][AArch64] Adding Neoverse N2 CPU support
Mark Murray [Mon, 16 Nov 2020 13:11:35 +0000 (13:11 +0000)]
[ARM][AArch64] Adding Neoverse N2 CPU support

Add support for the Neoverse N2 CPU to the ARM and AArch64 backends.

Differential Revision: https://reviews.llvm.org/D91695

17 months agoSemaExpr.cpp - use castAs<> instead of getAs<> as we dereference the pointer directly...
Simon Pilgrim [Tue, 24 Nov 2020 17:36:58 +0000 (17:36 +0000)]
SemaExpr.cpp - use castAs<> instead of getAs<> as we dereference the pointer directly. NFCI.

castAs<> will assert the correct cast type instead of just returning null, which we then try to dereference immediately.

17 months agoTargetInfo.cpp - use castAs<> instead of getAs<> as we dereference the pointer direct...
Simon Pilgrim [Tue, 24 Nov 2020 16:56:24 +0000 (16:56 +0000)]
TargetInfo.cpp - use castAs<> instead of getAs<> as we dereference the pointer directly. NFCI.

castAs<> will assert the correct cast type instead of just returning null, which we then try to dereference immediately.

17 months agoCGCall.cpp - use castAs<> instead of getAs<> as we dereference the pointer directly...
Simon Pilgrim [Tue, 24 Nov 2020 16:55:26 +0000 (16:55 +0000)]
CGCall.cpp - use castAs<> instead of getAs<> as we dereference the pointer directly. NFCI.

castAs<> will assert the correct cast type instead of just returning null, which we then try to dereference immediately in the setUsedBits call.

17 months agoSemaExpr.cpp - use castAs<> instead of getAs<> as we dereference the pointer directly...
Simon Pilgrim [Tue, 24 Nov 2020 16:51:55 +0000 (16:51 +0000)]
SemaExpr.cpp - use castAs<> instead of getAs<> as we dereference the pointer directly. NFCI.

castAs<> will assert the correct cast type instead of just returning null, which we then try to dereference immediately.

17 months agoDetectDeadLanes.cpp - remove unused headers. NFCI.
Simon Pilgrim [Tue, 24 Nov 2020 16:20:03 +0000 (16:20 +0000)]
DetectDeadLanes.cpp - remove unused headers. NFCI.

17 months ago[SVE][CodeGen] Add a DAG combine to extend mscatter indices
Kerry McLaughlin [Wed, 25 Nov 2020 10:54:31 +0000 (10:54 +0000)]
[SVE][CodeGen] Add a DAG combine to extend mscatter indices

This patch adds a target-specific DAG combine for mscatter to promote indices
with element types i8 or i16 before legalisation, plus various tests with illegal types.

Reviewed By: sdesmalen

Differential Revision: https://reviews.llvm.org/D90945

17 months ago[gn build] Port d95db1693cb
LLVM GN Syncbot [Wed, 25 Nov 2020 11:13:15 +0000 (11:13 +0000)]
[gn build] Port d95db1693cb

17 months ago[clangd] Extract common file-caching logic from ConfigProvider.
Sam McCall [Wed, 23 Sep 2020 17:58:32 +0000 (19:58 +0200)]
[clangd] Extract common file-caching logic from ConfigProvider.

The plan is to use this to use this for .clang-format, .clang-tidy, and
compile_commands.json. (Currently the former two are reparsed every
time, and the latter is cached forever and changes are never seen).

Differential Revision: https://reviews.llvm.org/D88172

17 months ago[VPlan] Add VPReductionSC to VPUser::classof, unify VPValue IDs.
Florian Hahn [Wed, 25 Nov 2020 10:04:22 +0000 (10:04 +0000)]
[VPlan] Add VPReductionSC to VPUser::classof, unify VPValue IDs.

This is a follow-up to 00a66011366c7b037d6680e6015524a41b761c34 to make
isa<VPReductionRecipe> work and unifies the VPValue ID names, by making
sure they all consistently start with VPV*.

17 months ago[clangd] Fix a tsan failure.
Haojian Wu [Wed, 25 Nov 2020 10:46:41 +0000 (11:46 +0100)]
[clangd] Fix a tsan failure.

Tracer must be set up before calling any clangd-specific functions.

17 months ago[OpenCL] Move kernel arg type tests into one file
Sven van Haastregt [Wed, 25 Nov 2020 10:20:30 +0000 (10:20 +0000)]
[OpenCL] Move kernel arg type tests into one file

Keep all kernel parameter type diagnostic tests in
invalid-kernel-parameters.cl .

Differential Revision: https://reviews.llvm.org/D92033

17 months ago[clang][SVE] Activate macro `__ARM_FEATURE_SVE_VECTOR_OPERATORS`.
Francesco Petrogalli [Wed, 25 Nov 2020 09:37:00 +0000 (09:37 +0000)]
[clang][SVE] Activate macro `__ARM_FEATURE_SVE_VECTOR_OPERATORS`.

The macro is emitted when wargeting SVE code generation with the additional command line option `-msve-vector-bits=<N>`.

The behavior implied by the macro is described in sections " Behavior specific to SVE vectors" of the SVE ACLE (Version 00bet6) that can be found at https://developer.arm.com/documentation/100987/latest

Reviewed By: rengolin, rsandifo-arm

Differential Revision: https://reviews.llvm.org/D90956

17 months ago[ORC] Cast to const void* to silence a GCC warning. NFC.
Martin Storsjö [Wed, 25 Nov 2020 10:06:26 +0000 (12:06 +0200)]
[ORC] Cast to const void* to silence a GCC warning. NFC.

17 months ago[ORC] Remove a superfluous semicolon, silencing GCC warnings. NFC.
Martin Storsjö [Wed, 25 Nov 2020 10:05:53 +0000 (12:05 +0200)]
[ORC] Remove a superfluous semicolon, silencing GCC warnings. NFC.

17 months ago[libObject,llvm-readelf] - Stop describing a section/segment in `notes_begin()`.
Georgii Rymar [Tue, 24 Nov 2020 14:34:12 +0000 (17:34 +0300)]
[libObject,llvm-readelf] - Stop describing a section/segment in `notes_begin()`.

`notes_begin()` is used for iterating over notes. This API in some cases might print
section type and index. At the same time during iterating, the `Elf_Note_Iterator`
might omit it as it doesn't have this info.

Because of above we might have the redundant duplication of information in warnings:
(See D92021).

warning: '[[FILE]]': unable to read notes from the SHT_NOTE section with index 1: SHT_NOTE section [index 1] has invalid offset (0x40) or size (0xffff0000)

This change stops reporting section index/type in Object/ELF.h/notes_begin().
(FTR, this was introduced by me for llvm-readobj in D64470).
Instead we can describe sections/program headers on the caller side.

Differential revision: https://reviews.llvm.org/D92081

17 months ago[clangd] Add metrics for invalid name.
Haojian Wu [Wed, 25 Nov 2020 09:49:54 +0000 (10:49 +0100)]
[clangd] Add metrics for invalid name.

Differential Revision: https://reviews.llvm.org/D92082

17 months ago[obj2yaml] - Dump section offsets in some cases.
Georgii Rymar [Mon, 9 Nov 2020 11:52:46 +0000 (14:52 +0300)]
[obj2yaml] - Dump section offsets in some cases.

Currently we never dump the `sh_offset` key.
Though it sometimes an important information.

To reduce the noise this patch implements the following logic:
1) The "Offset" key for the first section is always emitted.
2) If we can derive the offset for a next section naturally,
   then the "Offset" key is omitted.

By "naturally" I mean that section[X] offset is expected to be:
offsetOf(section[X]) == alignTo(section[X - 1].sh_offset + section[X - 1].sh_size, section[X].sh_addralign)

So, when it has the expected value, we omit it from the output.

Differential revision: https://reviews.llvm.org/D91152

17 months ago[SchedModels] Return earlier removed checks
Evgeny Leviant [Wed, 25 Nov 2020 09:07:35 +0000 (12:07 +0300)]
[SchedModels] Return earlier removed checks

It is possible that some write resource is variant in model A
and sequence in model B. Such case will trigger assertion in
getAllPredicates function.

17 months ago[clangd] Avoid type hierarchy crash on incomplete type
Nathan Ridge [Wed, 25 Nov 2020 08:11:54 +0000 (03:11 -0500)]
[clangd] Avoid type hierarchy crash on incomplete type

Fixes https://github.com/clangd/clangd/issues/597

Differential Revision: https://reviews.llvm.org/D92077

17 months ago[llvm-readobj] - An attempt to fix BB after D92018.
Georgii Rymar [Wed, 25 Nov 2020 08:37:40 +0000 (11:37 +0300)]
[llvm-readobj] - An attempt to fix BB after D92018.

AVR and PPC64 bots reports link errors:

/tmp/cclOvLx0.s: Assembler messages:
/tmp/cclOvLx0.s:9223: Error: symbol `_ZN4llvm12function_refIFvvEE11callback_fnIUlvE2_EEvl' is already defined
/tmp/cclOvLx0.s:9227: Error: symbol `.L._ZN4llvm12function_refIFvvEE11callback_fnIUlvE2_EEvl' is already defined
/tmp/cclOvLx0.s:10272: Error: symbol `_ZN4llvm12function_refIFvvEE11callback_fnIUlvE2_EEvl' is already defined
/tmp/cclOvLx0.s:10276: Error: symbol `.L._ZN4llvm12function_refIFvvEE11callback_fnIUlvE2_EEvl' is already defined
/tmp/cclOvLx0.s:10285: Error: symbol `_ZN4llvm12function_refIFvvEE11callback_fnIUlvE2_EEvl' is already defined
/tmp/cclOvLx0.s:10289: Error: symbol `.L._ZN4llvm12function_refIFvvEE11callback_fnIUlvE2_EEvl' is already defined

/tmp/ccFJYr6I.s: Assembler messages:
/tmp/ccFJYr6I.s:6284: Error: symbol `_ZN4llvm12function_refIFvvEE11callback_fnIUlvE2_EEvl' is already defined
/tmp/ccFJYr6I.s:7053: Error: symbol `_ZN4llvm12function_refIFvvEE11callback_fnIUlvE2_EEvl' is already defined
/tmp/ccFJYr6I.s:7093: Error: symbol `_ZN4llvm12function_refIFvvEE11callback_fnIUlvE2_EEvl' is already defined

I *guess* the reason might be the default lambda argument. I've removed it.

17 months ago[VPlan] Switch VPWidenRecipe to be a VPValue
David Green [Wed, 25 Nov 2020 08:25:06 +0000 (08:25 +0000)]
[VPlan] Switch VPWidenRecipe to be a VPValue

Similar to other patches, this makes VPWidenRecipe a VPValue. Because of
the way it interacts with the reduction code it also slightly alters the
way that VPValues are registered, removing the up front NeedDef and
using getOrAddVPValue to create them on-demand if needed instead.

Differential Revision: https://reviews.llvm.org/D88447

17 months ago[VPlan] Turn VPReductionRecipe into a VPValue
David Green [Tue, 24 Nov 2020 10:08:59 +0000 (10:08 +0000)]
[VPlan] Turn VPReductionRecipe into a VPValue

This converts the VPReductionRecipe into a VPValue, like other
VPRecipe's in preparation for traversing def-use chains. It also makes
it a VPUser, now storing the used VPValues as operands.

It doesn't yet change how the VPReductionRecipes are created. It will
need to call replaceAllUsesWith from the original recipe they replace,
but that is not done yet as VPWidenRecipe need to be created first.

Differential Revision: https://reviews.llvm.org/D88382

17 months ago[llvm-readelf/obj] - Deduplicate the logic that prints notes. NFCI.
Georgii Rymar [Tue, 24 Nov 2020 10:49:40 +0000 (13:49 +0300)]
[llvm-readelf/obj] - Deduplicate the logic that prints notes. NFCI.

We have a similar logic for LLVM/GNU styles that can be deduplicated.
This will allow to replace `reportError` calls with `reportUniqueWarning`
calls in a single place.

Differential revision: https://reviews.llvm.org/D92018

17 months ago[NFC][ARM][PhaseOrdering] Add one more test for D91800: LoopIdiom should be before...
Roman Lebedev [Wed, 25 Nov 2020 07:51:54 +0000 (10:51 +0300)]
[NFC][ARM][PhaseOrdering] Add one more test for D91800: LoopIdiom should be before IndVars

17 months ago[PowerPC] Probe the gap between stackptr and realigned stackptr
Kai Luo [Wed, 25 Nov 2020 01:38:19 +0000 (01:38 +0000)]
[PowerPC] Probe the gap between stackptr and realigned stackptr

During reviewing https://reviews.llvm.org/D84419, @efriedma mentioned the gap between realigned stack pointer and origin stack pointer should be probed too whatever the alignment is. This patch fixes the issue for PPC64.

Reviewed By: jsji

Differential Revision: https://reviews.llvm.org/D88078

17 months ago[CHR] Use pred_size (NFC)
Kazu Hirata [Wed, 25 Nov 2020 06:52:29 +0000 (22:52 -0800)]
[CHR] Use pred_size (NFC)

17 months agoRevert "[SCEV] Generalize no-self-wrap check in isLoopInvariantExitCondDuringFirstIte...
Max Kazantsev [Wed, 25 Nov 2020 06:20:15 +0000 (13:20 +0700)]
Revert "[SCEV] Generalize no-self-wrap check in isLoopInvariantExitCondDuringFirstIterations"

This reverts commit 7dcc8899174f44b7447bc48a9f2ff27f5458f8b7.

This patch introduced a logical error that breaks whole logic of this analysis.
All checks we are making are supposed to be loop-independent, so that we could
safely remove the range check. The 'nw' fact is loop-dependent, so we can remove
the check basing on facts from this very check.

Motivating examples will follow-up.

17 months ago[DAGCombine] Add hook to allow target specific test for sqrt input
QingShan Zhang [Wed, 25 Nov 2020 05:37:15 +0000 (05:37 +0000)]
[DAGCombine] Add hook to allow target specific test for sqrt input

PowerPC has instruction ftsqrt/xstsqrtdp etc to do the input test for software square root.
LLVM now tests it with smallest normalized value using abs + setcc. We should add hook to
target that has test instructions.

Reviewed By: Spatel, Chen Zheng, Qiu Chao Fang

Differential Revision: https://reviews.llvm.org/D80706

17 months ago[NewPM] Add pipeline EP callback after initial frontend cleanup
Arthur Eubanks [Thu, 19 Nov 2020 17:38:14 +0000 (09:38 -0800)]
[NewPM] Add pipeline EP callback after initial frontend cleanup

This matches the legacy PM's EP_ModuleOptimizerEarly. Some backends use
this extension point and adding the pass somewhere else like
PipelineStartEPCallback doesn't work.

Reviewed By: ychen

Differential Revision: https://reviews.llvm.org/D91804

17 months ago[IndVars] Use more precise context when eliminating narrowing
Max Kazantsev [Wed, 25 Nov 2020 04:46:22 +0000 (11:46 +0700)]
[IndVars] Use more precise context when eliminating narrowing

When deciding to widen narrow use, we may need to prove some facts
about it. For proof, the context is used. Currently we take the instruction
being widened as the context.

However, we may be more precise here if we take as context the point that
dominates all users of instruction being widened.

Differential Revision: https://reviews.llvm.org/D90456
Reviewed By: skatkov

17 months ago[DAG][PowerPC] Fix dropped `nsw` flag in `SimplifySetCC` by adding `doesNodeExist...
Kai Luo [Wed, 25 Nov 2020 03:28:48 +0000 (03:28 +0000)]
[DAG][PowerPC] Fix dropped `nsw` flag in `SimplifySetCC` by adding `doesNodeExist` helper

`SimplifySetCC` invokes `getNodeIfExists` without passing `Flags` argument and `getNodeIfExists` uses a default `SDNodeFlags` to intersect the original flags, as a consequence, flags like `nsw` is dropped. Added a new helper function `doesNodeExist` to check if a node exists without modifying its flags.

Reviewed By: #powerpc, nemanjai

Differential Revision: https://reviews.llvm.org/D89938

17 months ago[PPC][AIX] Add vector callee saved registers for AIX extended vector ABI
Zarko Todorovski [Wed, 25 Nov 2020 03:37:03 +0000 (22:37 -0500)]
[PPC][AIX] Add vector callee saved registers for AIX extended vector ABI

This patch is the initial patch for support of the AIX extended vector ABI.  The extended ABI treats vector registers V20-V31 as non-volatile and we add them as callee saved registers in this patch.

Reviewed By: sfertile

Differential Revision: https://reviews.llvm.org/D88676

17 months agoTry to fix tests after e16c0a9a68971 with CLANG_DEFAULT_LINKER=lld
Nico Weber [Wed, 25 Nov 2020 03:34:10 +0000 (22:34 -0500)]
Try to fix tests after e16c0a9a68971 with CLANG_DEFAULT_LINKER=lld

Tests that pass -mlinker-version=old version and that then don't
expect new flags to be passed need to explicitly request the system
linker now.

17 months ago[NFC][Test] Format the test for IEEE Long double
QingShan Zhang [Wed, 25 Nov 2020 03:00:24 +0000 (03:00 +0000)]
[NFC][Test] Format the test for IEEE Long double

17 months agoDon't assume the clang binary name contains the string "clang".
Richard Smith [Wed, 25 Nov 2020 02:51:08 +0000 (18:51 -0800)]
Don't assume the clang binary name contains the string "clang".

Also ensure the -cc1 argument is actually part of the clang -cc1 command
line rather than some unrelated command line.

17 months ago[SCEV] Use isa<> pattern for testing for CouldNotCompute [NFC]
Philip Reames [Wed, 25 Nov 2020 02:45:37 +0000 (18:45 -0800)]
[SCEV] Use isa<> pattern for testing for CouldNotCompute [NFC]

Some older code - and code copied from older code - still directly tested against the singelton result of SE::getCouldNotCompute.  Using the isa<SCEVCouldNotCompute> form is both shorter, and more readable.

17 months agoMake CallInst::updateProfWeight emit i32 weights instead of i64
Arthur Eubanks [Sat, 31 Oct 2020 19:38:03 +0000 (12:38 -0700)]
Make CallInst::updateProfWeight emit i32 weights instead of i64

Typically branch_weights are i32, not i64.
This fixes entry_counts_cold.ll under NPM.

Reviewed By: asbirlea

Differential Revision: https://reviews.llvm.org/D90539

17 months agoDebugInfo: Remove llc_dwarf usage from tests already relying on a target triple in...
David Blaikie [Wed, 25 Nov 2020 01:45:47 +0000 (17:45 -0800)]
DebugInfo: Remove llc_dwarf usage from tests already relying on a target triple in the IR

17 months ago[PowerPC][FP128] Fix the incorrect calling convention for IEEE long double on Power8
QingShan Zhang [Wed, 25 Nov 2020 01:38:57 +0000 (01:38 +0000)]
[PowerPC][FP128] Fix the incorrect calling convention for IEEE long double on Power8

For now, we are using the GPR to pass the arguments/return value for fp128 on Power8,
which is incorrect. It should be VSR. The reason why we do it this way is that,
we are setting the fp128 as illegal which make LLVM try to emulate it with i128 on
Power8. So, we need to correct it as legal.

Reviewed By: Nemanjai

Differential Revision: https://reviews.llvm.org/D91527

17 months agoDebugInfo: Add some missing explicit target triples.
David Blaikie [Wed, 25 Nov 2020 01:32:34 +0000 (17:32 -0800)]
DebugInfo: Add some missing explicit target triples.

Based on D91043 by Luís Marques. Thanks Luís!

Differential Revision: https://reviews.llvm.org/D91043

17 months agoFix compilation issue reported by MSVC user on cfe-dev
Reid Kleckner [Wed, 25 Nov 2020 01:26:11 +0000 (17:26 -0800)]
Fix compilation issue reported by MSVC user on cfe-dev

MSVC seems to think this `friend class TrailingObjects;` declaration is
declaring a TrailingObjects class instead of naming the injected base
class. Remove `class` so it does the right thing.

17 months ago[clang-tidy] Support IgnoredRegexp configuration to selectively suppress identifier...
smhc [Wed, 25 Nov 2020 01:18:38 +0000 (01:18 +0000)]
[clang-tidy] Support IgnoredRegexp configuration to selectively suppress identifier naming checks

The idea of suppressing naming checks for variables is to support code bases that allow short variables named e.g 'x' and 'i' without prefix/suffixes or casing styles. This was originally proposed as a 'ShortSizeThreshold' however has been made more generic with a regex to suppress identifier naming checks for those that match.

Reviewed By: njames93, aaron.ballman

Differential Revision: https://reviews.llvm.org/D90282

17 months ago[mlir][sparse] add parallelization strategies to sparse compiler
Aart Bik [Tue, 24 Nov 2020 23:36:10 +0000 (15:36 -0800)]
[mlir][sparse] add parallelization strategies to sparse compiler

This CL adds the ability to request different parallelization strategies
for the generate code. Every "parallel" loop is a candidate, and converted
to a parallel op if it is an actual for-loop (not a while) and the strategy
allows dense/sparse outer/inner parallelization.

This will connect directly with the work of @ezhulenev on parallel loops.

Still TBD: vectorization strategy

Reviewed By: penpornk

Differential Revision: https://reviews.llvm.org/D91978

17 months agoTreat a placeholder type for class template argument deduction as
Richard Smith [Wed, 25 Nov 2020 00:53:58 +0000 (16:53 -0800)]
Treat a placeholder type for class template argument deduction as
substitutable for the deduced template.

As agreed in https://github.com/itanium-cxx-abi/cxx-abi/issues/109.

17 months ago[lld] Add --no-lto-whole-program-visibility
Teresa Johnson [Tue, 24 Nov 2020 22:53:02 +0000 (14:53 -0800)]
[lld] Add --no-lto-whole-program-visibility

Enables overriding earlier --lto-whole-program-visibility.

Variant of D91583 while discussing alternate ways to identify and
handle the --export-dynamic case.

Differential Revision: https://reviews.llvm.org/D92060

17 months agoFix mangling of substitutions for template-prefixes.
Richard Smith [Tue, 24 Nov 2020 23:20:06 +0000 (15:20 -0800)]
Fix mangling of substitutions for template-prefixes.

Previously we only considered using a substitution for a template-name
after already having mangled its prefix, so we'd produce nonsense
manglings like NS3_S4_IiEE where we should simply produce NS4_IiEE.

This is not ABI-compatible with previous Clang versions, and the old
behavior is restored by -fclang-abi-compat=11.0 or earlier.

17 months ago[Sanitizer][RISCV] Fix redefinition of REG_SP
Luís Marques [Wed, 25 Nov 2020 00:03:34 +0000 (00:03 +0000)]
[Sanitizer][RISCV] Fix redefinition of REG_SP

The include header sys/ucontext.h already defines REG_SP as 2, causing
redefinition warnings during compilation. This patch fixes that issue.
(We also can't just use the numerical definition provided by the header,
as REG_SP is used in this file this refers to a struct field.)

Differential Revision: https://reviews.llvm.org/D90934

17 months ago[LAA] Minor code style tweaks [NFC]
Philip Reames [Tue, 24 Nov 2020 23:49:16 +0000 (15:49 -0800)]
[LAA] Minor code style tweaks [NFC]