platform/upstream/llvm.git
20 months ago[InstSimplify] Add additional insertvalue into undef tests (NFC)
Nikita Popov [Wed, 15 Feb 2023 15:32:46 +0000 (16:32 +0100)]
[InstSimplify] Add additional insertvalue into undef tests (NFC)

20 months ago[llvm][AArch64] Fix BTI after returns_twice when call has no attributes
David Spickett [Wed, 15 Feb 2023 08:55:22 +0000 (08:55 +0000)]
[llvm][AArch64] Fix BTI after returns_twice when call has no attributes

Previously we were looking for the returns twice attribute by manually
getting the function attributes from the call. This meant that we only
found attributes on the call itself, not what it was calling.

So if you had:
%call1 = call i32 @setjmp(ptr noundef null)

We would not BTI protect that even though setjmp clearly needs it.

Clang happens to produce:
%call = call i32 @setjmp(ptr noundef null) #0 ; returns_twice

So all valid calls were protected. This is not guaranteed,
the frontend may choose not to put attributes on the call.

It is undefined behaviour to call setjmp indirectly
(https://pubs.opengroup.org/onlinepubs/9699919799/functions/setjmp.html)
but as I misused the APIs here I think it's worth fixing up regardless.

Added comments to the test file where the IR differs from what
clang would output.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D144082

20 months agoRecommit "[ConstraintElimination] Change debug output to display variable names."
Zain Jaffal [Wed, 15 Feb 2023 15:22:42 +0000 (15:22 +0000)]
Recommit "[ConstraintElimination] Change debug output to display variable names."

This reverts commit 62d0e1a8541f93dfbf66d982f66da32676df2df7.

remove `dumpWithNames` function

20 months agoRevert "[ConstraintElimination] Change debug output to display variable names."
Zain Jaffal [Wed, 15 Feb 2023 15:21:46 +0000 (15:21 +0000)]
Revert "[ConstraintElimination] Change debug output to display variable names."

This reverts commit 869c87ad10e87db7c032c3464338ab9d50916510.

`dumpWithNames` function should be removed

20 months ago[ConstraintElimination] Change debug output to display variable names.
Zain Jaffal [Wed, 15 Feb 2023 15:07:14 +0000 (15:07 +0000)]
[ConstraintElimination] Change debug output to display variable names.

Previously when constraint system outputs the rows in the system the variables used are x1,2...n making it hard to infer which ones they relate to in the IR

Reviewed By: fhahn

Differential Revision: https://reviews.llvm.org/D142618

20 months ago[reland][libc] Separate memcpy implementations per arch
Guillaume Chatelet [Wed, 15 Feb 2023 10:47:10 +0000 (10:47 +0000)]
[reland][libc] Separate memcpy implementations per arch

As x86_64 implementations is likely to grow up to a point where it's no more manageable to have all implementations in the same file.

20 months ago[mlir][memref] Bufferize memref.tensor_store op
Matthias Springer [Wed, 15 Feb 2023 14:26:14 +0000 (15:26 +0100)]
[mlir][memref] Bufferize memref.tensor_store op

This change adds the BufferizableOpInterface implementation for memref.tensor_store.

Differential Revision: https://reviews.llvm.org/D144080

20 months ago[mlir][linalg] Add bufferize_to_allocation transform op
Matthias Springer [Wed, 15 Feb 2023 14:09:34 +0000 (15:09 +0100)]
[mlir][linalg] Add bufferize_to_allocation transform op

This transform materializes a buffer allocation for a given tensor value. All uses of the original value are replaced with the allocation.

Certain non-DPS ops may have an optimized lowering path that bufferizes the entire defining op. Such optimization is added for `tensor.pad` as part of this change.

The resulting IR can be further bufferized with One-Shot Bufferize.

Differential Revision: https://reviews.llvm.org/D144022

20 months ago[ConstantFold] Check for constant global earlier (NFC)
Nikita Popov [Wed, 15 Feb 2023 13:33:59 +0000 (14:33 +0100)]
[ConstantFold] Check for constant global earlier (NFC)

Check that the underlying object is a constant global with
definitive initializer upfront, so we can skip the more expensive
offset calculation logic if we can't perform the fold anyway.

20 months ago[lld][RISCV][test] Expand testing in riscv-attributes.s
Alex Bradbury [Wed, 15 Feb 2023 14:00:54 +0000 (14:00 +0000)]
[lld][RISCV][test] Expand testing in riscv-attributes.s

This patch is related to the discussion in
<https://discourse.llvm.org/t/rfc-resolving-issues-related-to-extension-versioning-in-risc-v/68472/1>.
It slightly expands the coverage of behaviour for unrecognized
extensions (there are slightly different forms of error for zfoo vs e.g.
'y'), and adds coverage for an extension with an
unrecognized/unsupported version number.

Use "unrecognized" terminology because the expectation is the error
checking will be reduced in a follow-up patch posted for review.

20 months ago[LangRef] improve documentation of SNaN in the default FP environment
Sanjay Patel [Wed, 15 Feb 2023 13:56:24 +0000 (08:56 -0500)]
[LangRef] improve documentation of SNaN in the default FP environment

Make it explicit that SNaN is not handled differently than
QNaN in the LLVM default floating-point environment.

Note that an IEEE-754-compliant model disallows transforms
like "X * 1.0 -> X". That is because math operations are
expected to convert SNaN to QNaN (set the signaling bit).

But LLVM has had those kinds of transforms from the beginning:
https://alive2.llvm.org/ce/z/igb55y

We should be IEEE-754-compliant under strict-FP (the logic is
implemented with a helper named canIgnoreSNaN()), but I don't
think there is any demand to do that with default optimization.

See issue #43070 for earlier draft/discussion about this change.

Differential Revision: https://reviews.llvm.org/D143074

20 months ago[bazel] Fix missing dependency in clang-tools-extra/clang-tidy:llvmlibc
Guillaume Chatelet [Wed, 15 Feb 2023 13:54:13 +0000 (13:54 +0000)]
[bazel] Fix missing dependency in clang-tools-extra/clang-tidy:llvmlibc

20 months ago[NVPTX] Fix NVPTX output name in the driver with `-save-temps`
Joseph Huber [Wed, 15 Feb 2023 13:37:47 +0000 (07:37 -0600)]
[NVPTX] Fix NVPTX output name in the driver with `-save-temps`

Summary:
Currently, OpenMP and direct compilation uses an NVPTX toolchain to
directly invoke the CUDA tools from Clang to do the assembling and
linking of NVPTX codes. This breaks under `-save-temps` because of a
workaround. The `nvlink` linker does not accept `.o` files, so we need
to be selective when we output these. The previous logic keyed off of
presense in the temp files and wasn't a great solution. Change this to
just query the input args for `-c` to see if we stop at the assembler.

Fixes https://github.com/llvm/llvm-project/issues/60767

20 months agoRevert "[libc] Separate memcpy implementations per arch"
Guillaume Chatelet [Wed, 15 Feb 2023 13:29:07 +0000 (13:29 +0000)]
Revert "[libc] Separate memcpy implementations per arch"

This is patch is still breaking downstream users...
This reverts commit 97e441dc6cfae31bc56b375e43899946ec7f08a8.

20 months ago[libc] Separate memcpy implementations per arch
Guillaume Chatelet [Wed, 15 Feb 2023 10:47:10 +0000 (10:47 +0000)]
[libc] Separate memcpy implementations per arch

As x86_64 implementations is likely to grow up to a point where it's no more manageable to have all implementations in the same file.

20 months ago[mlir][transform] Add transform.get_result op
Matthias Springer [Wed, 15 Feb 2023 13:15:41 +0000 (14:15 +0100)]
[mlir][transform] Add transform.get_result op

This transform op returns a value handle pointing to the specified OpResult of the targeted op.

Differential Revision: https://reviews.llvm.org/D144087

20 months ago[clang-format] assert(false) -> llvm_unreachable
Benjamin Kramer [Wed, 15 Feb 2023 13:10:09 +0000 (14:10 +0100)]
[clang-format] assert(false) -> llvm_unreachable

Avoids warnings in -asserts builds.

FormatTokenSource.h:240:3: error: non-void function does not return a value [-Werror,-Wreturn-type]
  }
  ^

20 months ago[bazel] create a clang-tidy binary target
Jathu Satkunarajah [Wed, 15 Feb 2023 12:51:52 +0000 (13:51 +0100)]
[bazel] create a clang-tidy binary target

Create a binary target for clang-tidy. Tested by running:

```
$ bazel build --config=generic_clang @llvm-project//clang-tools-extra/...
$ bazel test --config=generic_clang @llvm-project//clang-tools-extra/...
$ bazel run --config=generic_clang @llvm-project//clang-tools-extra/clang-tidy -- --help
```

Reviewed By: #bazel_build, aaronmondal

Differential Revision: https://reviews.llvm.org/D143804

20 months ago[MachineTraceMetrics] Add local strategy
Anton Sidorenko [Tue, 14 Feb 2023 10:30:44 +0000 (13:30 +0300)]
[MachineTraceMetrics] Add local strategy

This strategy makes each trace local to the basic block. For in-order cores some
heuristics work better when we do local decisions. For example, MachineCombiner
may expect that instructions outside the current basic block do not lengthen
the critical path when we execute instructions in order or the core has a
small re-order buffer.

This patch only introduce the strategy, real use-case is added in the further
pathes.

Depends on D140539

Reviewed By: spatel

Differential Revision: https://reviews.llvm.org/D140540

20 months ago[X86] Remove abs(sub_nsw()) -> abds fold from adbu test file
Simon Pilgrim [Wed, 15 Feb 2023 12:44:14 +0000 (12:44 +0000)]
[X86] Remove abs(sub_nsw()) -> abds fold from adbu test file

Copy+paste typo - it was correctly removed from 128/512 variants

20 months ago[lldb] Fix a log format warning on Windows, don't assume uint64_t is a long type
Martin Storsjö [Wed, 15 Feb 2023 08:42:20 +0000 (10:42 +0200)]
[lldb] Fix a log format warning on Windows, don't assume uint64_t is a long type

On Windows, long is 32 bit, and uint64_t is long long.

Differential Revision: https://reviews.llvm.org/D144078

20 months ago[clang-format] Enable FormatTokenSource to insert tokens.
Manuel Klimek [Wed, 1 Feb 2023 13:18:44 +0000 (13:18 +0000)]
[clang-format] Enable FormatTokenSource to insert tokens.

In preparation for configured macro replacements in formatting,
add the ability to insert tokens to FormatTokenSource, and implement
token insertion in IndexedTokenSource.

Differential Revision: https://reviews.llvm.org/D143070

20 months agoAMDGPU: Fix not adding to depth in getNegatedExpression
Matt Arsenault [Wed, 15 Feb 2023 12:31:40 +0000 (08:31 -0400)]
AMDGPU: Fix not adding to depth in getNegatedExpression

20 months agoclang: Rename misleading test name
Matt Arsenault [Tue, 31 Jan 2023 17:16:59 +0000 (13:16 -0400)]
clang: Rename misleading test name

This is a test for attribute propagation, not metadata

20 months ago[lld][ARM][NFCI][1/3]Big Endian support - Removing assumptions
Simi Pallipurath [Wed, 15 Feb 2023 11:06:04 +0000 (11:06 +0000)]
[lld][ARM][NFCI][1/3]Big Endian support - Removing assumptions

Change:
 - Replacing the memcpy that assume little endian with the endian-aware write.

Shouldn't affect the output for now, just a prerequisite for the next patches.

Reviewed By: MaskRay

Differential Revision: https://reviews.llvm.org/D140201

20 months ago[AArch64][NFC] Rename AEK_SMEF64 and AEK_SMEI64 feature flags
Caroline Concatto [Wed, 15 Feb 2023 09:40:41 +0000 (09:40 +0000)]
[AArch64][NFC] Rename  AEK_SMEF64 and AEK_SMEI64 feature flags

Update feature flag names from:
AEK_SMEF64  to AEK_SMEF64F64
and
AEK_SMEI64 to AEK_SMEI16I64
These feature flags had their name changed in this previous patch
 https://reviews.llvm.org/D135974

Reviewed By: c-rhodes

Differential Revision: https://reviews.llvm.org/D143989

20 months ago[SimpleLoopUnswitch] Canonicalize conditions for injection of invariant condition
Max Kazantsev [Wed, 15 Feb 2023 11:15:24 +0000 (18:15 +0700)]
[SimpleLoopUnswitch] Canonicalize conditions for injection of invariant condition

When loop condition isn't immediately in the form supported by invariant injection
unswitching, try to canonicalize it to this form.

Differential Revision: https://reviews.llvm.org/D143175
Reviewed By: skatkov

20 months ago[Test][NFC] Rename one of test parameters to avoid confusing associations
Max Kazantsev [Wed, 15 Feb 2023 11:07:02 +0000 (18:07 +0700)]
[Test][NFC] Rename one of test parameters to avoid confusing associations

Name 'len' implicitly hints that it should be non-negative, but in fact in some of
these tests the baseline profitable scenario for the transform to be profitable is
when this thing is negative. Rename to avoid this confusion.

20 months ago[Symbolize][MinGW] Support demangling i386 call-conv-decorated C++ names
Alvin Wong [Tue, 14 Feb 2023 22:38:52 +0000 (06:38 +0800)]
[Symbolize][MinGW] Support demangling i386 call-conv-decorated C++ names

On i386 Windows, after C++ names have been Itanium-mangled, the C name
mangling specific to its call convention may also be applied on top.
This change teaches symbolizer to be able to demangle this type of
mangled names.

As part of this change, `demanglePE32ExternCFunc` has also been modified
to fix unwanted stripping for vectorcall names when the demangled name
is supposed to contain a leading `_`. Notice that the vectorcall
mangling does not add either an `_` or `@` prefix. The old code always
tries to strip the prefix first, which for Itanium mangled names in
vectorcall, the leading underscore of the Itanium name gets stripped
instead and breaks the Itanium demangler.

Differential Revision: https://reviews.llvm.org/D144049

20 months ago[CVP] Add additional ctlz tests (NFC)
Nikita Popov [Wed, 15 Feb 2023 10:43:52 +0000 (11:43 +0100)]
[CVP] Add additional ctlz tests (NFC)

20 months ago[XCore] Adapt threads.ll to opaque pointers.
Nigel Perks [Tue, 14 Feb 2023 16:39:09 +0000 (16:39 +0000)]
[XCore] Adapt threads.ll to opaque pointers.

Differential Revision: https://reviews.llvm.org/D144085

20 months ago[SimpleLoopUnswitch] Fix overflowing frequencies case
Max Kazantsev [Wed, 15 Feb 2023 10:35:45 +0000 (17:35 +0700)]
[SimpleLoopUnswitch] Fix overflowing frequencies case

When branch is so hot that sum of its frequencies overflows, we have an assertion
failure when trying to construct BranchProbability. Such cases are not interesting
as candidate for unswitching anyways (their branches are too hot), so reject them.

20 months ago[clang-tidy][NFC] Remove ModernizeTidyModule::getModuleOptions
Carlos Galvez [Sun, 12 Feb 2023 10:58:37 +0000 (10:58 +0000)]
[clang-tidy][NFC] Remove ModernizeTidyModule::getModuleOptions

Most of the options stated there are duplicated already in
the implementation of each check as a default value for
each option.

The only place where this is not the case is the nullptr
check. Move the default option there instead. Only the
HICPP guidelines alias this modernize check, and there is
nothing in the documentation that suggests it should have
a different default value than the main modernize check.

Differential Revision: https://reviews.llvm.org/D143843

20 months ago[DAGCombine] Fold redundant select
Samuel Parker [Wed, 15 Feb 2023 10:32:16 +0000 (10:32 +0000)]
[DAGCombine] Fold redundant select

Recommit bbdf24357932b064f2aa18ea1356b474e0220dde.

Original commit message:

If a chain of two selects share a true/false value and are controlled
by two setcc nodes, that are never both true, we can fold away one of
the selects. So, the following:
(select (setcc X, const0, eq), Y,
  (select (setcc X, const1, eq), Z, Y))

Can be combined to:
  select (setcc X, const1, eq) Z, Y

Differential Revision: https://reviews.llvm.org/D142535

20 months ago[libc][NFC] Make tuning macros start with LIBC_COPT_
Guillaume Chatelet [Mon, 13 Feb 2023 14:52:26 +0000 (14:52 +0000)]
[libc][NFC] Make tuning macros start with LIBC_COPT_

Rename preprocessor definitions that control tuning of llvm libc.

Differential Revision: https://reviews.llvm.org/D143913

20 months agoAdd InstSimplify tests for comparisons between known constants
Matt Devereau [Wed, 15 Feb 2023 09:39:15 +0000 (09:39 +0000)]
Add InstSimplify tests for comparisons between known constants

20 months ago[JITLink] Drop const qualifier from argument to ELFLinkGraphBuilder's ctors (NFC)
Stefan Gränitz [Tue, 14 Feb 2023 14:08:44 +0000 (15:08 +0100)]
[JITLink] Drop const qualifier from argument to ELFLinkGraphBuilder's ctors (NFC)

These values are moved into the base class constructors, so the `const` doesn't make any sense. Turns out, I accidentally introduced it myself with 2ed91da0f1f3 and since than it spread by copy/paste.

20 months ago[JITLink] Fix whitespace in debug dumps (NFC)
Stefan Gränitz [Tue, 14 Feb 2023 14:01:51 +0000 (15:01 +0100)]
[JITLink] Fix whitespace in debug dumps (NFC)

20 months ago[bazel][mlir][examples] Add missing dependency for 72429a42ac33564fa82449d99dc234da32...
Benjamin Kramer [Wed, 15 Feb 2023 09:33:49 +0000 (10:33 +0100)]
[bazel][mlir][examples] Add missing dependency for 72429a42ac33564fa82449d99dc234da32a05498

20 months agoRevert "InstCombine: Fold is.fpclass(x, fcZero) to fcmp oeq 0"
Matt Arsenault [Wed, 15 Feb 2023 09:27:09 +0000 (05:27 -0400)]
Revert "InstCombine: Fold is.fpclass(x, fcZero) to fcmp oeq 0"

This reverts commit df78976d023a6b7fcf64bc695261b7b402fcede0.

I pushed the wrong patch

20 months ago[flang] Update intrinsic types to unlimited polymorphic form
Valentin Clement [Wed, 15 Feb 2023 09:20:21 +0000 (10:20 +0100)]
[flang] Update intrinsic types to unlimited polymorphic form

This patch updates the code added in D143888 to avoid
overwriting some part of the types when updating it
for unlimited polymorphic types.

Reviewed By: jeanPerier, PeteSteinfeld

Differential Revision: https://reviews.llvm.org/D143995

20 months agoAMDGPU: Override getNegatedExpression constant handling
Matt Arsenault [Mon, 13 Feb 2023 18:27:29 +0000 (14:27 -0400)]
AMDGPU: Override getNegatedExpression constant handling

Ignore the multiple use heuristics of the default
implementation, and report cost based on inline immediates. This
is mostly interesting for -0 vs. 0. Gets a few small improvements.
fneg_fadd_0_f16 is a small regression. We could probably avoid this
if we handled folding fneg into div_fixup.

20 months agoAMDGPU: Refactor isConstantCostlierToNegate
Matt Arsenault [Mon, 13 Feb 2023 11:42:47 +0000 (07:42 -0400)]
AMDGPU: Refactor isConstantCostlierToNegate

20 months agoInstCombine: Fold is.fpclass(x, fcZero) to fcmp oeq 0
Matt Arsenault [Fri, 11 Nov 2022 00:15:34 +0000 (16:15 -0800)]
InstCombine: Fold is.fpclass(x, fcZero) to fcmp oeq 0

This requires the denormal mode to definitively be IEEE handling.

20 months ago[mlir][bufferization] Add restrict and writable attrs to to_tensor
Matthias Springer [Wed, 15 Feb 2023 08:51:42 +0000 (09:51 +0100)]
[mlir][bufferization] Add restrict and writable attrs to to_tensor

`restrict` is similar to the C++ restrict keyword. Results of `to_tensor` that have the `restrict` attribute are guaranteed to not alias any other `to_tensor` result (after bufferization).

Note: Since `to_memref` ops are not supported by One-Shot Bufferize and all bufferizable ops follow DPS rules (i.e., the buffer of the result is the buffer of an operand or an alias thereof), the buffer of a `to_tensor` op that has the `restrict` attribute is always an entirely "new" buffer that is not aliasing with the future buffer of any tensor value in the entire program. This makes such `to_tensor` ops "safe" from a bufferization perspective; they cannot cause RaW conflicts.

Differential Revision: https://reviews.llvm.org/D144021

20 months ago[SROA] NFC: Look at TypeStoreSize scalable property, rather than at type directly.
Sander de Smalen [Tue, 14 Feb 2023 14:08:02 +0000 (14:08 +0000)]
[SROA] NFC: Look at TypeStoreSize scalable property, rather than at type directly.

Some places in the code have checks for isa<ScalableVectorType> and use
that to bail out of the code. It's also possible to look directly at the
allocated type-size and check if the size is scalable. This means it's
possible to also support other scalable types that are not vectors (i.e.
TargetExtType).

This is split out from D136861.

20 months ago[llvm-c] Add C API methods to match 64bit ArrayType C++ API signatures
Meghan Denny [Wed, 15 Feb 2023 08:47:32 +0000 (09:47 +0100)]
[llvm-c] Add C API methods to match 64bit ArrayType C++ API signatures

Fixes https://github.com/llvm/llvm-project/issues/56496.

As mentioned in the issue, new functions LLVMArrayType2 and
LLVMGetArrayLength2 are created so as to not break the old API.
The old methods are then marked as deprecated and callers are
updated.

Differential Revision: https://reviews.llvm.org/D143700

20 months ago[mlir] Add vectorize_nd_extract attribute to the masked_vectorize Op
Andrzej Warzynski [Thu, 26 Jan 2023 10:18:44 +0000 (10:18 +0000)]
[mlir] Add vectorize_nd_extract attribute to the masked_vectorize Op

This patch simply adds `vectorize_nd_extract` (that's currently only
used for the `vectorize` Op) to `masked_vectorize`. A test is added to
verify that it works as expected - it prevents the masked vectorisation
of `tensor.extract`, which is currently not supported.

Differential Revision: https://reviews.llvm.org/D142634

20 months ago[Metarenamer] Use 'inst' as default name for instructions
Max Kazantsev [Wed, 15 Feb 2023 08:34:24 +0000 (15:34 +0700)]
[Metarenamer] Use 'inst' as default name for instructions

Currently we use 'tmp', which is also a keyword for FileCheck. It leads to this
annoying warning whenever a script for auto-generation of checks is used.
It is especially annoying that it happens to every test affected by metarenamer.

Just use another prefix for metarenamed names to avoid this.

Differential Revision: https://reviews.llvm.org/D144001
Reviewed By: nikic

20 months ago[clang][analyzer] Make messages of StdCLibraryFunctionsChecker user-friendly
Balázs Kéri [Wed, 15 Feb 2023 07:54:20 +0000 (08:54 +0100)]
[clang][analyzer] Make messages of StdCLibraryFunctionsChecker user-friendly

Warnings and notes of checker alpha.unix.StdLibraryFunctionArgs are
improved. Previously one warning and one note was emitted for every
finding, now one warning is emitted only that contains a detailed
description of the found issue.

Reviewed By: Szelethus

Differential Revision: https://reviews.llvm.org/D143194

20 months ago[LoopFuse] Remove legacy pass
Fangrui Song [Wed, 15 Feb 2023 07:53:39 +0000 (23:53 -0800)]
[LoopFuse] Remove legacy pass

Following recent changes to remove non-core legacy passes.

20 months ago[MLIR][Tensor] Introduce a pattern to propagate through `tensor.pad`
Lorenzo Chelini [Mon, 13 Feb 2023 13:44:03 +0000 (14:44 +0100)]
[MLIR][Tensor] Introduce a pattern to propagate through `tensor.pad`

Introduce a pattern to 'push down' a `tensor.unpack` through a
`tensor.pad`. The propagation happens if the unpack does not touch the
padded dimensions.

Reviewed By: hanchung

Differential Revision: https://reviews.llvm.org/D143907

20 months ago[InstCombine] Fix InstCombinerImpl::foldICmpMulConstant for nsw and nuw mul with...
Craig Topper [Wed, 15 Feb 2023 07:43:17 +0000 (23:43 -0800)]
[InstCombine] Fix InstCombinerImpl::foldICmpMulConstant for nsw and nuw mul with unsigned compare.

If we have both an nsw and nuw flag, we would see the nsw flag
first and only handle signed comparisons.

This patch ignores the nsw flag if the comparison isn't signed.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D143766

20 months ago[mlir][Bazel] Add missing dependencies for 72429a42ac33564fa82449d99dc234da32a05498
Adrian Kuegel [Wed, 15 Feb 2023 07:30:57 +0000 (08:30 +0100)]
[mlir][Bazel] Add missing dependencies for 72429a42ac33564fa82449d99dc234da32a05498

20 months ago[mlir] Add RewriterBase::replaceAllUsesWith for Blocks.
Ingo Müller [Tue, 24 Jan 2023 15:42:38 +0000 (15:42 +0000)]
[mlir] Add RewriterBase::replaceAllUsesWith for Blocks.

When changing IR in a RewriterPattern, all changes must go through the
rewriter. There are several convenience functions in RewriterBase that
help with high-level modifications, such as replaceAllUsesWith for
Values, but there is currently none to do the same task for Blocks.

Reviewed By: mehdi_amini, ingomueller-net

Differential Revision: https://reviews.llvm.org/D142525

20 months agoAdd a pass that builds a debug info scope for LLVMFuncOp (adding a DISubprogramAttr)
Mehdi Amini [Wed, 15 Feb 2023 05:15:30 +0000 (21:15 -0800)]
Add a pass that builds a debug info scope for LLVMFuncOp (adding a DISubprogramAttr)

This may be seen as a hack, but it allows for any piece of MLIR to be able
to end up with DWARF debug info through LLVM.
Assuming the operations in the function have location such as FileLineCol,
this provides backtraces with line tables and allows to step in a debugger.
That makes this pass a perfect companion to -snapshot-op-locations

It was also the default behavior of MLIR to LLVM IR translation until MLIR
got support for proper debug info attributes.

Differential Revision: https://reviews.llvm.org/D144069

20 months ago[mlir][Bazel] Add dependency needed for e9b82a5c
Adrian Kuegel [Wed, 15 Feb 2023 06:36:00 +0000 (07:36 +0100)]
[mlir][Bazel] Add dependency needed for e9b82a5c

20 months ago[mlir][Vector] Enable masking for static shapes
Diego Caballero [Wed, 15 Feb 2023 06:01:42 +0000 (06:01 +0000)]
[mlir][Vector] Enable masking for static shapes

Support for masking static shapes was already implemented in the past
but not enabled so this patch is just removing a pre-condition check and
adding some tests with static shapes.

Reviewed By: ThomasRaoux

Differential Revision: https://reviews.llvm.org/D143937

20 months ago[mlir][Vector] Add support for masked vector gather ops
Diego Caballero [Wed, 15 Feb 2023 06:00:12 +0000 (06:00 +0000)]
[mlir][Vector] Add support for masked vector gather ops

This patch adds support for masked vector.gather ops using the
vector.mask representation. It includes the implementation of the
MaskableOpInterface, Linalg vectorizer support and lowering to LLVM.

Reviewed By: ThomasRaoux

Differential Revision: https://reviews.llvm.org/D143939

20 months ago[mlir][Vector] Add support for masked vector.contract
Diego Caballero [Wed, 15 Feb 2023 05:46:15 +0000 (05:46 +0000)]
[mlir][Vector] Add support for masked vector.contract

This patch adds support for masking vector.contract ops with the
vector.mask approach. This also includes the lowering of vector.contract
through the vector.outerproduct path to LLVM. For now, this only adds
support for one of the many potential flavors of
vector.contract/vector.outerproduct but unsupported cases will fail
gratefully.

Reviewed By: ThomasRaoux

Differential Revision: https://reviews.llvm.org/D143965

20 months ago[mlir][Vector] Add LLVM lowering for masked reductions
Diego Caballero [Mon, 13 Feb 2023 19:31:23 +0000 (19:31 +0000)]
[mlir][Vector] Add LLVM lowering for masked reductions

This patch adds the conversion patterns to lower masked reduction
operations to the corresponding vp intrinsics in LLVM.

Reviewed By: ftynse

Differential Revision: https://reviews.llvm.org/D142177

20 months ago[LoopVersioningLICM] Remove legacy pass
Fangrui Song [Wed, 15 Feb 2023 05:40:45 +0000 (21:40 -0800)]
[LoopVersioningLICM] Remove legacy pass

Following recent changes to remove non-core features of the legacy PM/optimization pipeline.

20 months ago[mlir] Use std::optional instead of llvm::Optional (NFC)
Kazu Hirata [Wed, 15 Feb 2023 04:44:15 +0000 (20:44 -0800)]
[mlir] Use std::optional instead of llvm::Optional (NFC)

This is part of an effort to migrate from llvm::Optional to
std::optional:

https://discourse.llvm.org/t/deprecating-llvm-optional-x-hasvalue-getvalue-getvalueor/63716

20 months ago[mlir] Use std::optional instead of llvm::Optional (NFC)
Kazu Hirata [Wed, 15 Feb 2023 04:33:36 +0000 (20:33 -0800)]
[mlir] Use std::optional instead of llvm::Optional (NFC)

This is part of an effort to migrate from llvm::Optional to
std::optional:

https://discourse.llvm.org/t/deprecating-llvm-optional-x-hasvalue-getvalue-getvalueor/63716

20 months ago[llvm] Use std::optional instead of llvm::Optional (NFC)
Kazu Hirata [Wed, 15 Feb 2023 04:27:41 +0000 (20:27 -0800)]
[llvm] Use std::optional instead of llvm::Optional (NFC)

This is part of an effort to migrate from llvm::Optional to
std::optional:

https://discourse.llvm.org/t/deprecating-llvm-optional-x-hasvalue-getvalue-getvalueor/63716

20 months ago[mlir] Fix -Wsign-compare in SparseTensorRewriting.cpp and Sparsification.cpp (NFC)
Jie Fu [Wed, 15 Feb 2023 03:56:40 +0000 (11:56 +0800)]
[mlir] Fix -Wsign-compare in SparseTensorRewriting.cpp and Sparsification.cpp (NFC)

/home/jiefu/llvm-project/mlir/lib/Dialect/SparseTensor/Transforms/Sparsification.cpp:279:33: error: comparison of integers of different signs: 'int64_t' (aka 'long') and 'const mlir::sparse_tensor::Level' (aka 'const unsigned long') [-Werror,-Wsign-compare]
    assert(env.op().getRank(&t) == lvlRank);
           ~~~~~~~~~~~~~~~~~~~~ ^  ~~~~~~~
/usr/include/assert.h:93:27: note: expanded from macro 'assert'
     (static_cast <bool> (expr)                                         \
                          ^~~~
1 error generated.

/home/jiefu/llvm-project/mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorRewriting.cpp:788:29: error: comparison of integers of different signs: 'int64_t' (aka 'long') and 'const mlir::sparse_tensor::Dimension' (aka 'const unsigned long') [-Werror,-Wsign-compare]
    assert(srcRTT.getRank() == dimRank);
           ~~~~~~~~~~~~~~~~ ^  ~~~~~~~
/usr/include/assert.h:93:27: note: expanded from macro 'assert'
     (static_cast <bool> (expr)                                         \
                          ^~~~
/home/jiefu/llvm-project/mlir/lib/Dialect/SparseTensor/Transforms/SparseTensorRewriting.cpp:810:31: error: comparison of integers of different signs: 'int64_t' (aka 'long') and 'const mlir::sparse_tensor::Dimension' (aka 'const unsigned long') [-Werror,-Wsign-compare]
      assert(srcRTT.getRank() == dimRank);
             ~~~~~~~~~~~~~~~~ ^  ~~~~~~~
/usr/include/assert.h:93:27: note: expanded from macro 'assert'
     (static_cast <bool> (expr)                                         \
                          ^~~~
2 errors generated.

20 months ago[gn build] Port f1c4241fb6e5
LLVM GN Syncbot [Wed, 15 Feb 2023 03:51:14 +0000 (03:51 +0000)]
[gn build] Port f1c4241fb6e5

20 months ago[RISCV] Add new pass to transform undef to pseudo for vector values.
Piyou Chen [Wed, 15 Feb 2023 03:42:26 +0000 (19:42 -0800)]
[RISCV] Add new pass to transform undef to pseudo for vector values.

RISC-V vector instruction has register overlapping constraint for certain
instructions, and will cause illegal instruction trap if violated, we use
early clobber to model this constraint, but it can't prevent register allocator
allocated same or overlapped if the input register is undef value, so convert
IMPLICIT_DEF to temporary pseudo could prevent that happen, it's not best way
to resolve this. Ideally we should model the constraint right, but before we
model the constraint right, it's the approach to prevent that happen.

See also: https://github.com/llvm/llvm-project/issues/50157

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D129735

20 months ago[RISCV] precommit test for D129735
Piyou Chen [Wed, 15 Feb 2023 03:29:59 +0000 (19:29 -0800)]
[RISCV] precommit test for D129735

Reviewed By: kito-cheng

Differential Revision: https://reviews.llvm.org/D137763

20 months ago[NFC] Refactor ModuleDeclStateTest to make it not dependent on Frontend
Chuanqi Xu [Wed, 15 Feb 2023 03:29:29 +0000 (11:29 +0800)]
[NFC] Refactor ModuleDeclStateTest to make it not dependent on Frontend

Required in https://reviews.llvm.org/D137526. And it is indeed odd that
LexTest depends on ClangFrontend.

20 months ago[ORC] Add ELFNixPlatform::Create overload -- Pass ORC runtime as def generator.
Lang Hames [Wed, 15 Feb 2023 03:13:41 +0000 (19:13 -0800)]
[ORC] Add ELFNixPlatform::Create overload -- Pass ORC runtime as def generator.

The existing Create method took a path to the ORC runtime and created a
StaticLibraryDefinitionGenerator for it. The new overload takes a
std::unique_ptr<DefinitionGenerator> directly instead. This provides more
flexibility when constructing MachOPlatforms. E.g. The runtime archive can be
embedded in a special section in the ORC controller executable or library,
rather than being on-disk.

This is the ELFNixPlatform equivalent of the MachOPlatform change in
be2fc577c38.

20 months ago[mlir][sparse] Factoring out SparseTensorType class
wren romano [Wed, 15 Feb 2023 02:20:45 +0000 (18:20 -0800)]
[mlir][sparse] Factoring out SparseTensorType class

This change adds a new `SparseTensorType` class for making the "dim" vs "lvl" distinction more overt, and for abstracting over the differences between sparse-tensors and dense-tensors.  In addition, this change also adds new type aliases `Dimension`, `Level`, and `FieldIndex` to make code more self-documenting.

Although the diff is very large, the majority of the changes are mechanical in nature (e.g., changing types to use the new aliases, updating variable names to match, etc).  Along the way I also made many variables `const` when they could be; the majority of which required only adding the keyword.  A few places had conditional definitions of these variables, requiring actual code changes; however, that was only done when the overall change was extremely local and easy to extract.  All these changes are included in the current patch only because it would be too onerous to split them off into a separate patch.

Reviewed By: aartbik

Differential Revision: https://reviews.llvm.org/D143800

20 months ago[C++20] [Modules] [ClangScanDeps] Ensure that we can mix the use of and clang modules
Chuanqi Xu [Wed, 15 Feb 2023 02:59:07 +0000 (10:59 +0800)]
[C++20] [Modules] [ClangScanDeps] Ensure that we can mix the use of and clang modules

Add a test to ensure that the clang-scan-deps won't crash due to the
unexpected use of clang modules.

20 months ago[clang-format][NFC] Reformat clang/tools/clang-format/fuzzer/
Owen Pan [Wed, 15 Feb 2023 02:54:36 +0000 (18:54 -0800)]
[clang-format][NFC] Reformat clang/tools/clang-format/fuzzer/

20 months ago[mlir][sparse] fix a bug in UnpackOp converter.
Peiming Liu [Wed, 15 Feb 2023 02:18:54 +0000 (02:18 +0000)]
[mlir][sparse] fix a bug in UnpackOp converter.

UnpackOp Converter used to create reallocOp unconditionally, but it might cause issue when the requested memory size is smaller than the actually storage.

Reviewed By: aartbik

Differential Revision: https://reviews.llvm.org/D144065

20 months ago[InstCombine] fold icmp of the sum of ext bool based on limited range
chenglin.bi [Wed, 15 Feb 2023 02:33:54 +0000 (10:33 +0800)]
[InstCombine] fold icmp of the sum of ext bool based on limited range

For the pattern `(zext i1 X) + (sext i1 Y)`, the constant range is [-1, 1].
We can simplify the pattern by logical operations. Like:

```
    (zext i1 X) + (sext i1 Y) == -1 -->  ~X & Y
    (zext i1 X) + (sext i1 Y) == 0  --> ~(X ^ Y)
    (zext i1 X) + (sext i1 Y) == 1 --> X & ~Y
```
And other predicates can the combination of these results:

```
    (zext i1 X) + (sext i1 Y)) != -1 --> X | ~Y
    (zext i1 X) + (sext i1 Y)) s> -1 --> X | ~Y
    (zext i1 X) + (sext i1 Y)) u< -1 --> X | ~Y
    (zext i1 X) + (sext i1 Y)) s> 0 --> X & ~Y
    (zext i1 X) + (sext i1 Y)) s< 0 --> ~X & Y
    (zext i1 X) + (sext i1 Y)) != 1 --> ~X | Y
    (zext i1 X) + (sext i1 Y)) s< 1 --> ~X | Y
    (zext i1 X) + (sext i1 Y)) u> 1 --> ~X & Y
```

All alive proofs:
https://alive2.llvm.org/ce/z/KmgDpF
https://alive2.llvm.org/ce/z/fLwWa9
https://alive2.llvm.org/ce/z/ZKQn2P

Fix: https://github.com/llvm/llvm-project/issues/59666

Reviewed By: spatel

Differential Revision: https://reviews.llvm.org/D143373

20 months ago[SeparateConstOffsetFromGEP] Fix: `b - a` matched `a - b` during reuniteExts
Liren Peng [Wed, 15 Feb 2023 02:18:41 +0000 (02:18 +0000)]
[SeparateConstOffsetFromGEP] Fix: `b - a` matched `a - b` during reuniteExts

During the SeparateConstOffsetFromGEP pass, a - b and b - a will be
considered equivalent in some instances.

An example- the IR contains:

  BB1:
      %add = add %a, 511
      br label %BB2
  BB2:
      %sub2 = sub %b,  %a
      br label %BB3
  BB3:
      %sub1 = sub %add, %b
      %gep = getelementptr float, ptr %p, %sub1

Step 1 in the SeparateConstOffsetFromGEP pass, after split constant index:

  BB1:
      %add = add %a, 511
      br label %BB2
  BB2:
      %sub2 = sub %b,  %a
      br label %BB3
  BB3:
      %sub.t = sub %a, %b
      %gep.base = getelementptr float, ptr %p, %sub.t
      %gep = getelementptr float, ptr %gep.base, 511

Step 2, after reuniteExts:

  BB1:
      br label %BB2
  BB2:
      %sub2 = sub %b,  %a
      br label %BB3
  BB3:
      %gep.base = getelementptr float, ptr %p, %sub2
      %gep = getelementptr float, ptr %gep.base, 511

Obviously, reuniteExts treated a - b and b - a as equivalent.
This patch fixes that.

Reviewed By: nikic, spatel

Differential Revision: https://reviews.llvm.org/D143542

20 months ago[Instcombine] Precommit tests for icmp range; NFC
chenglin.bi [Wed, 15 Feb 2023 02:31:27 +0000 (10:31 +0800)]
[Instcombine] Precommit tests for icmp range; NFC

20 months ago[NFC][SeparateConstOffsetFromGEP] Added flag `lower-gep`
Liren Peng [Wed, 15 Feb 2023 01:22:31 +0000 (01:22 +0000)]
[NFC][SeparateConstOffsetFromGEP] Added flag `lower-gep`

We need such a flag to check whether the transformation is correct if
LowerGEP was enabled.

Reviewed By: nikic, arsenm, spatel

Differential Revision: https://reviews.llvm.org/D143980

20 months ago[bazel] Port 26662ac010ef50e65e2774eab84f325aa09360fe
Fangrui Song [Wed, 15 Feb 2023 01:42:48 +0000 (17:42 -0800)]
[bazel] Port 26662ac010ef50e65e2774eab84f325aa09360fe

20 months ago[Clang][RISCV] Guard vector int64, float32, float64 with semantic analysis
eopXD [Thu, 9 Feb 2023 18:14:59 +0000 (10:14 -0800)]
[Clang][RISCV] Guard vector int64, float32, float64 with semantic analysis

Depends on D143657

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D143665

20 months ago[mlir] Fix two build warnings in VectorToGPU (NFC)
Jie Fu [Wed, 15 Feb 2023 01:31:16 +0000 (09:31 +0800)]
[mlir] Fix two build warnings in VectorToGPU (NFC)

In file included from /data/llvm-project/mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp:13:
/data/llvm-project/mlir/include/mlir/Conversion/VectorToGPU/VectorToGPU.h:15:1: error: class 'LogicalResult' was previously declared as a struct; this is valid, but may result in linker errors under the Microsoft C++ ABI [-Werror,-Wmismatched-tags]
class LogicalResult;
^
/data/llvm-project/mlir/include/mlir/Support/LogicalResult.h:26:22: note: previous use is here
struct [[nodiscard]] LogicalResult {
                     ^
/data/llvm-project/mlir/include/mlir/Conversion/VectorToGPU/VectorToGPU.h:15:1: note: did you mean struct here?
class LogicalResult;
^~~~~
struct

/data/llvm-project/mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp:724:5: error: ignoring return value of function declared with 'nodiscard' attribute [-Werror,-Wunused-result]
    rewriter.notifyMatchFailure(
    ^~~~~~~~~~~~~~~~~~~~~~~~~~~
2 errors generated.

20 months ago[gn build] Add missing count dependency for check-asan
Arthur Eubanks [Wed, 15 Feb 2023 01:33:15 +0000 (17:33 -0800)]
[gn build] Add missing count dependency for check-asan

20 months ago[scudo] Update ring buffer test to make it accept zero size
Chia-hung Duan [Wed, 15 Feb 2023 01:23:13 +0000 (01:23 +0000)]
[scudo] Update ring buffer test to make it accept zero size

allocation ring buffer is allowed to be zero. Update the logic in the
test so that on the platform that disables it won't fail this case.

Reviewed By: fmayer

Differential Revision: https://reviews.llvm.org/D144055

20 months ago[LoopUnrollAndJam] Remove legacy pass
Fangrui Song [Wed, 15 Feb 2023 01:23:15 +0000 (17:23 -0800)]
[LoopUnrollAndJam] Remove legacy pass

Following recent changes to remove non-core features of the legacy PM/optimization pipeline.

20 months ago[scudo] Add the thread-safety annotations
Chia-hung Duan [Fri, 10 Feb 2023 21:54:03 +0000 (21:54 +0000)]
[scudo] Add the thread-safety annotations

This CL adds the proper thread-safety annotations for most of the
functions and variables. However, given the restriction of the current
architecture, in some cases, we may not be able to use the annotations
easily. The followings are two exceptions,

1. enable()/disable(): Many structures in scudo are enabled/disabled by
   acquiring the lock in each instance. This makes those structure act
   like a `lock`. We can't mark those functions with ACQUIRE()/RELEASE()
   because that makes the entire allocator become another `lock`. In the
   end, that implies we need to *acquire* the `allocator` before each
   malloc et al. request. Therefore, adding a variable to tell the
   status of those structures may be a better way to cooperate with
   thread-safety annotation.

2. TSD/TSD shared/TSD exclusive: These three have simiar restrictions as
   mentioned above. In addition, they don't always need to be released
   if it's a thread local instance. However, thread-safety analysis
   doesn't support conditional branch. Which means we can't mark the
   proper annotations around the uses of TSDs. We may consider to make
   it consistent and which makes the code structure simpler.

This CL is supposed to introduce the annotations with the least code
refactoring. So only trivial thread safety issues will be addressed
here. For example, lacking of acquiring certain lock before accessing
certain variables will have the ScopedLock inserted. Other than that,
they are supposed to be done in the later changes.

Reviewed By: cferris

Differential Revision: https://reviews.llvm.org/D140706

20 months ago[LoongArch] Make use of addu16i.d for adds with suitable immediates
WANG Xuerui [Tue, 14 Feb 2023 00:39:19 +0000 (08:39 +0800)]
[LoongArch] Make use of addu16i.d for adds with suitable immediates

Ideally `addu16i.d` could be paired with `{ld,st}ptr` for faster memory
accesses with 32-bit-aligned offsets (it was designed for this purpose),
but it would require more work and the original use case (GP-relative
accesses) does not exist any more with the current LoongArch psABI.

It could still be used for accelerating additions of certain constants
though, which is what this patch intends to do.

Reviewed By: SixWeining, gonglingqin

Differential Revision: https://reviews.llvm.org/D143710

20 months ago[LoongArch] Add baseline tests for `addu16i.d` codegen. NFC
WANG Xuerui [Tue, 14 Feb 2023 00:39:12 +0000 (08:39 +0800)]
[LoongArch] Add baseline tests for `addu16i.d` codegen. NFC

Reviewed By: gonglingqin

Differential Revision: https://reviews.llvm.org/D143846

20 months ago[Clang][Test] Add llvm-lto, llvm-lto2 and llvm-profdata to the tool substitutions...
Weining Lu [Tue, 14 Feb 2023 00:36:43 +0000 (08:36 +0800)]
[Clang][Test] Add llvm-lto, llvm-lto2 and llvm-profdata to the tool substitutions list

Similar to issue fixed in D107155, some lit tests invoke `llvm-lto`,
`llvm-lto2` and `llvm-profdata` without going through the substitution
system. While the test runner correctly picks up these binaries from
the build directory, it doesn't print its absolute path. When copying
the invocations when reproducing test failures, this can result in
`command not found: llvm-lto` or other errors (caused by using wrong
version of the tool).

This patch adds these tools to the `tools` variable in
`clang/test/lit.cfg.py` which will then call add_tool_substitutions.

Reviewed By: dblaikie

Differential Revision: https://reviews.llvm.org/D143806

20 months agoRemove trailing whitespace in `X86InstrAVX512.td`; NFC
Noah Goldstein [Wed, 15 Feb 2023 01:12:54 +0000 (19:12 -0600)]
Remove trailing whitespace in `X86InstrAVX512.td`; NFC

Summary:

Reviewers:

Subscribers:

20 months agoRemove incorrect comment around `truncateAVX512SetCCNoBWI`; NFC
Noah Goldstein [Wed, 15 Feb 2023 01:12:27 +0000 (19:12 -0600)]
Remove incorrect comment around `truncateAVX512SetCCNoBWI`; NFC

Differential Revision: https://reviews.llvm.org/D144060

20 months agoAdd transform (and/or (icmp eq/ne (A, C)), (icmp eq/ne (A, -C))) -> (icmp eq/ne ...
Noah Goldstein [Tue, 14 Feb 2023 01:39:42 +0000 (19:39 -0600)]
Add transform (and/or (icmp eq/ne (A, C)), (icmp eq/ne (A, -C))) -> (icmp eq/ne (ABS A), ABS(C))

This can be beneficial if there is a fast `ABS` (For example with X86
`vpabs`) or if there is a dominating ABS(A) in the `DAG`.

Note `C` is constant so `ABS(C)` is just a constant.

Alive2 Links:
EQ: https://alive2.llvm.org/ce/z/829F-c
NE: https://alive2.llvm.org/ce/z/tsS8bU

Reviewed By: pengfei

Differential Revision: https://reviews.llvm.org/D142601

20 months agoTests for (and/or (icmp eq/ne A, C), (icmp eq/ne A, -C)) <--> (icmp eq/ne (ABS A...
Noah Goldstein [Tue, 14 Feb 2023 01:40:02 +0000 (19:40 -0600)]
Tests for (and/or (icmp eq/ne A, C), (icmp eq/ne A, -C)) <--> (icmp eq/ne (ABS A), ABS(C)); NFC

Reviewed By: pengfei

Differential Revision: https://reviews.llvm.org/D142600

20 months agoTransform `(icmp eq/ne Abs(A), Pow2)` -> `(and/or (icmp eq/ne A,Pow2), (icmp eq/ne...
Noah Goldstein [Tue, 14 Feb 2023 01:40:18 +0000 (19:40 -0600)]
Transform `(icmp eq/ne Abs(A), Pow2)` -> `(and/or (icmp eq/ne A,Pow2), (icmp eq/ne A,-Pow2))`

Only if Abs(A) has one use, in which case the `(and/or (icmp eq/ne
A,Pow2), (icmp eq/ne A,-Pow2))` can be optimized in
`DAGCombiner::foldAndOrOfSETCC`.

Alive Links:
EQ: https://alive2.llvm.org/ce/z/gTxSgV
NE: https://alive2.llvm.org/ce/z/MUf57Y

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D142345

20 months agoAdd Transform for `(and/or (eq/ne A,Pow2),(eq/ne A,-Pow2))`->`(eq/ne (and (and A...
Noah Goldstein [Tue, 14 Feb 2023 01:40:35 +0000 (19:40 -0600)]
Add Transform for `(and/or (eq/ne A,Pow2),(eq/ne A,-Pow2))`->`(eq/ne (and (and A,Pow2),~(Pow2*2)), 0)`

In many instances this can be preferable if the `icmp` -> `i1` cannot be
done in one instruction (such as X86 for scalars).

At the moment guarded behind `TLI.isDesirableToCombineLogicOpOfSETCC`.

alive2 links:
https://alive2.llvm.org/ce/z/nLm5sN
https://alive2.llvm.org/ce/z/moEcyE

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D142344

20 months agoAdd tests for folding `(and/or (icmp eq/ne A, Pow2), (icmp eq/ne A, -Pow2))`; NFC
Noah Goldstein [Tue, 14 Feb 2023 01:40:50 +0000 (19:40 -0600)]
Add tests for folding `(and/or (icmp eq/ne A, Pow2), (icmp eq/ne A, -Pow2))`; NFC

Differential Revision: https://reviews.llvm.org/D142343

20 months agoTransform vector SET{LE/ULT/ULE} -> SETLT and SET{GE/UGT/UGE} -> SETGT if possible
Noah Goldstein [Tue, 14 Feb 2023 01:39:05 +0000 (19:39 -0600)]
Transform vector SET{LE/ULT/ULE} -> SETLT and SET{GE/UGT/UGE} -> SETGT if possible

SETLT and SETGT can use `{v}pcmpgt` directly whereas the other SETCC
variants require some other instructions as well. On AVX512, which has
vector comparisons for all SETCC variants, this can still be
preferable if the destination is a vector. And if the destination is a
mask, its the same performance.

The transform for unsigned SETCC takes place if we know from
`KnownBits` that LHS/RHS have the same sign.

The transform for LE/GE -> LT/GT takes place if LHS/RHS is constant
and we can inc/dec all elements in the operand without overflowing
(both signed and unsigned).

Alive2 Links (on i8 so they don't timeout):
sge_s:   https://alive2.llvm.org/ce/z/rMPt9_
sge_s_2: https://alive2.llvm.org/ce/z/G74Mhs
sge_u:   https://alive2.llvm.org/ce/z/PTWARM
sge_u_2: https://alive2.llvm.org/ce/z/L9dsNn
sgt_s:   https://alive2.llvm.org/ce/z/q2CHEK
sgt_u:   https://alive2.llvm.org/ce/z/YPLnZ8
sle_s:   https://alive2.llvm.org/ce/z/HyYhQ_
sle_s_2: https://alive2.llvm.org/ce/z/ck6NkT
sle_u:   https://alive2.llvm.org/ce/z/tyF_wN
sle_u_2: https://alive2.llvm.org/ce/z/et8t98
slt_s:   https://alive2.llvm.org/ce/z/oCP43b
slt_u:   https://alive2.llvm.org/ce/z/EpLLPx
uge_s:   https://alive2.llvm.org/ce/z/rqSDwi
uge_s_2: https://alive2.llvm.org/ce/z/67UTXu
uge_u:   https://alive2.llvm.org/ce/z/yBNG9C
uge_u_2: https://alive2.llvm.org/ce/z/UhHYc_
ugt_s:   https://alive2.llvm.org/ce/z/tY9va4
ugt_u:   https://alive2.llvm.org/ce/z/F9zeAT
ule_s:   https://alive2.llvm.org/ce/z/1MNgka
ule_s_2: https://alive2.llvm.org/ce/z/oiS7Ls
ule_u:   https://alive2.llvm.org/ce/z/8DveC3
ule_u_2: https://alive2.llvm.org/ce/z/jGp2M7
ult_s:   https://alive2.llvm.org/ce/z/chzfwP
ult_u:   https://alive2.llvm.org/ce/z/Jj_JYu

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D142254

20 months ago[mlir][Vector] NFC - Fail gracefully on size mismatch instead of assert
Nicolas Vasilache [Wed, 15 Feb 2023 00:04:05 +0000 (16:04 -0800)]
[mlir][Vector] NFC - Fail gracefully on size mismatch instead of assert

20 months ago[mlir][Vector] Significantly improve VectorToGPU.cpp
Nicolas Vasilache [Fri, 10 Feb 2023 15:22:30 +0000 (07:22 -0800)]
[mlir][Vector] Significantly improve VectorToGPU.cpp

This revision performs a bunch of cleanups and tracks free-flowing IR mutations.
APIs are systematized around RewriterBase and relevant debug messages are added.
Deliberate use of OpBuilder::InsertionGuard is added where needed.

Differential Revision: https://reviews.llvm.org/D143738

20 months ago[Support/BLAKE3] Rename blake3_* -> llvm_blake3_* to avoid symbol collisions
Jeremy Maitin-Shepard [Wed, 15 Feb 2023 00:21:42 +0000 (16:21 -0800)]
[Support/BLAKE3] Rename blake3_* -> llvm_blake3_* to avoid symbol collisions

While some blake3 symbols are already prefixed, a number of symbols with hidden visibility have been left without an `llvm_` prefix.  This results in symbol collisions when statically linking llvm into a binary that also uses the external blake3 library.

Reviewed By: akyrtzi, MaskRay

Differential Revision: https://reviews.llvm.org/D143981