Arthur Eubanks [Tue, 4 Oct 2022 22:13:16 +0000 (15:13 -0700)]
[PassTimingInfo] Stop double (or worse) counting passes/analyses
If we nest timers, we end up double counting anything nested.
The most egregious is ModuleInlinerWrapperPass/DevirtSCCRepeatedPass showing up as >20% of the total time when they're just wrappers.
Analyses also end up getting counted multiple times because they're nested inside wrappers and passes.
Ignore ModuleInlinerWrapperPass/DevirtSCCRepeatedPass and put analyses into their own TimerGroup.
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D135219
Craig Topper [Thu, 13 Oct 2022 15:52:58 +0000 (08:52 -0700)]
[RISCV] Match (select C, -1, X)->(or -C, X) during lowerSelect
Same with (select C, X, -1), (select C, 0, X), and (select C, X, 0).
There's a DAGCombine after we turn the select into select_cc, but
that may introduce a setcc that didn't previously exist. We could
add more DAGCombines to remove the extra setcc, but this seemed lower
effort.
Reviewed By: reames
Differential Revision: https://reviews.llvm.org/D135833
Nico Weber [Thu, 13 Oct 2022 15:18:59 +0000 (11:18 -0400)]
[lld/mac] Convert undef-spell-corrector.s test to split-file
Differential Revision: https://reviews.llvm.org/D135888
Nico Weber [Thu, 13 Oct 2022 14:36:08 +0000 (10:36 -0400)]
[lld/mac] Make two local variables const
While reading this code, I was wondering if we change these variables in the
loop. We don't, so make them const to make this easier to see next time.
No behavior change.
Differential Revision: https://reviews.llvm.org/D135877
Ramiro Leal-Cavazos [Thu, 13 Oct 2022 16:02:01 +0000 (09:02 -0700)]
Fix handling of rank-1 tensors in tosa.reduce_sum
The conversion of `tosa.reduce_sum` to linalg creates a
`linalg.generic` op that produces a tensor of rank `input_rank -
1`. This tensor is then expanded back into a tensor of rank
`input_rank`. In the case where the tensor being expanded is rank-0,
the reassociation map used must be empty. However, the current
implementation indexes and modifies the reassociation map independent
of the rank of the tensor being expanded, resulting in out-of-bounds
indexing when the tensor being expanded is rank-0. This commit adds a
guard to the reassociation map indexing.
Reviewed By: jpienaar
Differential Revision: https://reviews.llvm.org/D135828
Mehdi Amini [Thu, 6 Oct 2022 19:31:50 +0000 (19:31 +0000)]
Apply clang-tidy fixes for performance-unnecessary-value-param in TosaCanonicalizations.cpp (NFC)
Mehdi Amini [Thu, 6 Oct 2022 19:31:17 +0000 (19:31 +0000)]
Apply clang-tidy fixes for modernize-use-equals-default in TosaCanonicalizations.cpp (NFC)
David Green [Thu, 13 Oct 2022 15:41:44 +0000 (16:41 +0100)]
[CodeGenPrep] Handle constants in ConvertPhiType
This is a simple addition to the convertPhiTypes in CodeGenPrepare to
consider and convert constants as it converts the phi type. Someone
fixed the bug in the motivating example, so the undef is now a constant
0. This does mean converting between integer and floating point
constants, which may have different materialization.
Differential Revision: https://reviews.llvm.org/D135561
Nicolas Vasilache [Thu, 13 Oct 2022 15:13:25 +0000 (08:13 -0700)]
[mlir][Transform] Fix dropReverseMapping early exit condition
Previously, the erasure would not trigger and result in surprising behavior.
Differential Revision: https://reviews.llvm.org/D135881
Nikita Popov [Thu, 13 Oct 2022 15:29:37 +0000 (17:29 +0200)]
[MemCpyOpt] Add additional tests with lifetime intrinsics (NFC)
David Green [Thu, 13 Oct 2022 15:23:34 +0000 (16:23 +0100)]
[AArch64] Add ConvertPhiType constant tests. NFC
Nikita Popov [Thu, 13 Oct 2022 15:02:16 +0000 (17:02 +0200)]
[MemCpyOpt] Add additional debug output (NFC)
Nikita Popov [Thu, 13 Oct 2022 14:47:04 +0000 (16:47 +0200)]
[MemCpyOpt] Don't run full pipeline in test (NFC)
Just memcpyopt is enough for this test.
Aaron Ballman [Thu, 13 Oct 2022 14:58:52 +0000 (10:58 -0400)]
Update developer policy on potentially breaking changes
We've recently had issues appropriately notifying users and
stakeholders of changes to projects that may be potentially disruptive
when upgrading. This led to discussion on Discourse about how to
improve the situation, which can be found at:
https://discourse.llvm.org/t/rfc-add-new-discourse-channel-for-potentially-breaking-disruptive-changes-for-clang/65251
Ultimately, it sounds like we want to encourage three things:
* Alert vendors during the code review so they can provide early
feedback on potentially breaking changes that would be unacceptable
for them.
* Alert vendors and users after committing the changes so they can
perform pre-release testing on a completed change to determine if it
causes unreasonable problems for them.
* Alert users more clearly through the release notes so that it's
easier to determine how disruptive an upgrade might be.
This updates the developer policy accordingly.
Differential Revision: https://reviews.llvm.org/D134878
Florian Hahn [Thu, 13 Oct 2022 14:56:26 +0000 (15:56 +0100)]
[LV] Add epilogue test with variable induction start value.
Add additional test mentioned by @venkataramanan.kumar.llvm in
D92132.
Guillaume Chatelet [Thu, 13 Oct 2022 14:49:58 +0000 (14:49 +0000)]
Revert "[libc] New version of the mem* framework"
This reverts commit
d55f2d8ab076298cfd745c05c1b4dfd5583f8b9e.
Guillaume Chatelet [Thu, 13 Oct 2022 14:43:44 +0000 (14:43 +0000)]
[libc] New version of the mem* framework
This version is more composable and also simpler at the expense of being more explicit and more verbose. It also provides minimal implementations for ARM platforms.
Codegen can be checked here https://godbolt.org/z/x19zvE59v
Differential Revision: https://reviews.llvm.org/D135134
Nico Weber [Wed, 12 Oct 2022 00:40:26 +0000 (20:40 -0400)]
Reland "[lld/mac] Port typo correction for undefined symbols from ELF port"
The only difference in the reland is that the loop at the top of
getAlternativeSpelling() now calls dyn_cast_or_null() instead
of dyn_cast() -- a file's symbols list can contain null entries.
The test for this might be slightly involved, so I'll land it in
a follow-up, to make the reland similar to the original commit.
Originally reviewed at:
Differential Revision: https://reviews.llvm.org/D135038
This reverts commit
317b5582b813c51d1fb6723fd44b227b7f274bc7.
Alexey Bataev [Tue, 4 Oct 2022 18:14:38 +0000 (11:14 -0700)]
[SLP]Improve costs of vectorized loads/stores by analyzing GEPs.
When generating masked gathers nodes, SLP vectorizer accounts the cost
of the GEPs for loads as part of the scalar-vector transformation cost
estimation. But it does not do it for vectorized loads/stores, while it
may completely remove some of the GEPs completely. Because of this in
some cases masked gather operation can be much more profitable rather
than regular vectorization (masked-gather cost + vector GEP - scalar
loads + GEPs comparing to vectorized loads - scalar loads).
Added the analysis of the removed scalarGEPs for vectorized load/store nodes for better cost estimation.
Differential Revision: https://reviews.llvm.org/D135282
Philip Reames [Thu, 13 Oct 2022 14:08:35 +0000 (07:08 -0700)]
Revert "Remove PlaceSafepoints pass"
This reverts commit
cb66e123c6bc82a793300b6fb3ecbed79c58f557. It was reported via https://reviews.llvm.org/rGcb66e123c6bc82a793300b6fb3ecbed79c58f557#1132969 that the Microsoft.NET compiler is still using this pass.
Matt Devereau [Thu, 13 Oct 2022 14:06:24 +0000 (14:06 +0000)]
[VectorCombine] Add insertelement-shufflevector VectorCombine tests
This is a precommit which adds some tests to show the functionality of an
upcoming VectorCombine optimization
Nemanja Ivanovic [Thu, 13 Oct 2022 14:06:26 +0000 (09:06 -0500)]
[PowerPC] Stash GPR to VSR if emergency spill slot is not reachable
When removing frame indices on PowerPC, we need to scavenge
a GPR to materialize a large constant if the stack offset
for the spill/reload cannot be reached by a D-Form
instruction. However, in a perfect storm of conditions,
we may not have GPR's available to scavenge, thereby
requiring an emergency spill. If such an emergency
spill also needs to be spilled to a location with a
large offset, it would itself require register scavenging
thereby creating an infinite loop.
This patch detects when the scavenger cannot scavenge
a register and the spill/reload is to a location with
a large offset. It then stashes a GPR into a VSR so
that it can use the GPR to materialize the constant
(rather than scavenging a GPR).
Fixes: https://github.com/llvm/llvm-project/issues/52894
Differential revision: https://reviews.llvm.org/D124841
Anton Sidorenko [Thu, 13 Oct 2022 13:56:14 +0000 (14:56 +0100)]
[NFC] Use forward decl of MachineCombinerPattern enum to reduce dependencies
Differential Revision: https://reviews.llvm.org/D135776
Joseph Huber [Thu, 13 Oct 2022 13:44:07 +0000 (08:44 -0500)]
[Libomp] Do not error on undefined version script symbols
Summary:
A recent patch made all of the places we used version scripts in OpenMP
start throwing errors when linking with LLD. This is hopefully the last
one required to get the build bots to pass. This patch simply adds the
`--undefined-version` flag to the places where the version scripts are
used.
Simon Pilgrim [Thu, 13 Oct 2022 13:34:01 +0000 (14:34 +0100)]
[X86] Attempt to combine binary shuffles where both operands come from the same larger vector
Allows us to use combineX86ShuffleChainWithExtract to combine targetshuffle(low_subvector(x),high_subvector(x)) -> low_subvector(targetshuffle(x)) style patterns
This is currently very limited (it must have a v2i64/v2f64 result), but while triaging I noticed we might be able to extend this to allow more types for targets with suitable variable cross lane shuffle support.
Fixes #58339
Pavel Labath [Thu, 13 Oct 2022 13:22:22 +0000 (15:22 +0200)]
[lldb] Place PlatformQemu Properties into anonymous namespace
It's fine right now, but will break as soon as someone else declares a
PluginProperties class in the same way.
Also tighten up the scope of the anonymous namespaces surrounding the
other PluginProperties classes.
Joseph Huber [Thu, 13 Oct 2022 13:08:46 +0000 (08:08 -0500)]
[Libomptarget] Make the plugins ingore undefined exported symbols
Summary:
Recent changes made the default behaviour to error when given an
undefined symbol in a version script. A previous patch fixed this for
`libomptarget` by removing the single undefined symbol. However, the
plguins are expected to only define a subset of the availible functions
so we shouldn't treat it as an error. This patch updates the build flags
to work appropriately.
Florian Hahn [Thu, 13 Oct 2022 12:57:43 +0000 (13:57 +0100)]
[ConstraintElim] Use MulOverflow to avoid UB on signed overflow.
This fixes an UBSan failure after
359bc5c541ae. For inbounds GEP with
index sizes <= 64, having the coefficients overflowing is fine.
Valentin Clement [Thu, 13 Oct 2022 12:48:20 +0000 (14:48 +0200)]
[flang] Lower tbp dispatch calls for polymorphic array element
When calling a type-bound procedure from a polymoprhic array element,
the dynamic type needs to be extracted from the array descriptor
and passed to the embox operation for the pass-object.
Depends on D135809
Reviewed By: jeanPerier
Differential Revision: https://reviews.llvm.org/D135856
Joseph Huber [Thu, 13 Oct 2022 12:39:04 +0000 (07:39 -0500)]
[Libomptarget] Remove undefined 'omp_get_interop_rc_desc' symbol from exports list
Summary:
A recent patch made undefined symbols in version scripts cause errors by
default. The `omp_get_interop_rc_desc` function is declared but not
defined, so it is undefined in the final link unit. This patch removes
it from the exports list, it should be added back in when actually
defined and used.
Tobias Gysi [Thu, 13 Oct 2022 11:45:13 +0000 (14:45 +0300)]
[mlir][llvm] Use TableGen to import compare ops from LLVM IR.
The revision imports compare operations using TableGen generated
builders, instead of using the special handlers defined by the Importer.
It therefore adds a new llvmArgIndexes field that allows to specify
a mapping between MLIR argument and LLVM IR operand indexes if they do
not match. Additionally, the FCmp op is extended with an additional
builder and all compare operations are extended with verification
traits to ensure the operands types match. These extensions simplify
the logic of the newly introduced builders and are in line with the
compare operations define by the arithmetic dialect.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D135855
Jean Perier [Thu, 13 Oct 2022 12:24:56 +0000 (14:24 +0200)]
[flang][RFC] Adding higher level FIR ops to ease expression lowering
This document describes a new HLFIR dialect with a new value type
(hlfir.expr) and some new higher level FIR operations to make lowering
to FIR more straightforward and make pattern matching of high level
Fortran concepts (array and character assignments, character operations,
and transformational intrinsics) easier in FIR.
This should allow implementing the remaining gaps in Fortran 95 features
without increasing the lowering code complexity too much, and get a
clean start for the remaining F2003 and F2018 features.
Differential Revision: https://reviews.llvm.org/D134285
Weining Lu [Thu, 13 Oct 2022 06:37:53 +0000 (14:37 +0800)]
[Clang][LoongArch] Pass "f" and "d" features to cc1 to enable hard float
On LoongArch, currently neither of "f" and "d" feature is passed from
clang driver to cc1 by default. This means the backend generates code
for soft float.
In order to run programs in current LoongArch machines (hard float
environment) this patch temporarily enables "f" and "d" features.
In future, we should conditionally turn on these features depend on
various clang options, e.g. -mdouble-float, -msingle-float,
-msoft-float and -mfpu.
WANG Xuerui [Thu, 13 Oct 2022 11:29:17 +0000 (19:29 +0800)]
[test][LoongArch] Mark old JIT ExecutionEngine tests as unsupported
The old MCJIT won't be accepting new architectures so it can only be
appropriate to add ourselves to the exclude list.
Differential Revision: https://reviews.llvm.org/D134367
WANG Xuerui [Thu, 13 Oct 2022 11:19:49 +0000 (19:19 +0800)]
[LoongArch] Add support for llvm.trap and llvm.debugtrap
Similar to D69390 for RISCV, use a guaranteed non-existing insn for
llvm.trap and the break insn for llvm.debugtrap.
Differential Revision: https://reviews.llvm.org/D134365
Rupert Nash [Thu, 13 Oct 2022 11:22:50 +0000 (07:22 -0400)]
[libc++] <sstream>: update references to standard paragraphs
The paragraph numbers used previously did not match the C++20 nor
C++17 standards. I have updated them to the textual dotted names to
avoid this problem in future.
Reviewed By: #libc, ldionne
Differential Revision: https://reviews.llvm.org/D135861
WANG Xuerui [Thu, 13 Oct 2022 06:56:52 +0000 (14:56 +0800)]
[LoongArch] Updates for the LoongArch ELF psABI v2.01 revision
The e_flags of existing object files are all 0x3 which happens to be
compatible. From this commit on, all LoongArch objects produced with
upstream LLVM will be of object file ABI v1, which is already supported
by binutils' master branch (to be released as 2.40), and is allowed by
the same binutils version to interlink with v0 objects so the existing
distributions have time to migrate.
Differential Revision: https://reviews.llvm.org/D134601
Sheng [Thu, 13 Oct 2022 10:57:00 +0000 (18:57 +0800)]
[AArch64] Improve codegen for "trunc <4 x i64> to <4 x i8>" for all cases
To achieve this, we need this observation:
`uzp1` is just a `xtn` that operates on two registers
For example, given the following register with type v2i64:
LSB_______MSB
x0 x1 x2 x3
Applying xtn on it we get:
x0 x2
This is equivalent to bitcast it to v4i32, and then applying uzp1 on it:
x0 x1 x2 x3
|
uzp1
v
x0 x2 <value from other register>
We can transform xtn to uzp1 by this observation, and vice versa.
This observation only works on little endian target. Big endian target has
a problem: the uzp1 cannot be replaced by xtn since there is a discrepancy
in the behavior of uzp1 between the little endian and big endian.
To illustrate, take the following for example:
LSB____________________MSB
x0 x1 x2 x3
On little endian, uzp1 grabs x0 and x2, which is right; on big endian, it
grabs x3 and x1, which doesn't match what I saw on the document. But, since
I'm new to AArch64, take my word with a pinch of salt. This bevavior is
observed on gdb, maybe there's issue in the order of the value printed by it ?
Whatever the reason is, the execution result given by qemu just doesn't match.
So I disable this on big endian target temporarily until we find the crux.
Fixes #57502
Reviewed By: dmgreen, mingmingl
Co-authored-by: Mingming Liu <mingmingl@google.com>
Differential Revision: https://reviews.llvm.org/D133850
Sam McCall [Tue, 11 Oct 2022 17:10:53 +0000 (19:10 +0200)]
[pseudo] Document disambiguation design progress
Need to take a break from this, so write down where we got to.
Differential Revision: https://reviews.llvm.org/D135696
Daniel Grumberg [Wed, 12 Oct 2022 17:18:15 +0000 (18:18 +0100)]
[clang][ExtractAPI] Ignore fully anonymous RecordDecls
ExtractAPI was emitting a separate symbol for anonymous record declaration
that define the type of a member of another record declaration. Now
ExtractAPI ignores these declarations and just records the existence of
the actual member.
Differential Revision: https://reviews.llvm.org/D135804
Simon Pilgrim [Thu, 13 Oct 2022 10:36:14 +0000 (11:36 +0100)]
[X86][AVX2] Add shuffle test case where we fail to merge vpunpcklqdq(vextracti128(x,0),vextracti128(x,1)) -> vpermq
These are likely to appear during truncation
Caroline Concatto [Thu, 6 Oct 2022 11:38:42 +0000 (12:38 +0100)]
[AArch64] Add SME 2 target feature for Armv8-A and Armv9-A 2022 Architecture Extension
First patch in a series adding MC layer support for Scalable Matrix
Extension 2 (SME2).
This patch adds the following feature:
sme2
The 2022 Architecture Extension release adds other feature flags(eg.:sme2.1),
that will be in follow-up patches.
The reference can be found here:
https://developer.arm.com/documentation/ddi0602/2022-09
Differential Revision: https://reviews.llvm.org/D135448
Nikita Popov [Thu, 13 Oct 2022 10:08:08 +0000 (12:08 +0200)]
Revert "[FunctionAttrs] Make location classification more precise"
This reverts commit
b05f5b90a12098660a4fc16da0b4d421ddfe14e2.
There are thread sanitizer buildbot failures in simple_stack.c.
I think that's because this ended up affecting the handling of
volatile accesses to allocas. Reverting for now.
Nikita Popov [Mon, 10 Oct 2022 12:12:18 +0000 (14:12 +0200)]
[ModRef] Move ModRefInfo and FunctionModRefBehavior into IR (NFC)
This is in preparation for
https://discourse.llvm.org/t/rfc-unify-memory-effect-attributes/65579.
FunctionModRefBehavior will be used by attributes, and as such has
to be available in IR, not just Analysis.
Differential Revision: https://reviews.llvm.org/D135589
Sam McCall [Wed, 12 Oct 2022 23:43:49 +0000 (01:43 +0200)]
[clangd] Block clang-tidy misc-const-correctness check
This check performs an extremely large amount of work (for each variable, it
runs very many full matcher-driven traversals of the whole scope the variable
is defined in).
When (inadvertently) enabled for Fuchsia, it regressed BuildAST times by >10x
(400ms -> 7s on my machine).
Differential Revision: https://reviews.llvm.org/D135829
Valentin Clement [Thu, 13 Oct 2022 09:40:40 +0000 (11:40 +0200)]
[flang] Add optional tdesc operand to fir.embox for polymorphic entities
In some cases, it is useful to be able to embox a polymorphic entity
together with its dynamic type. This patch adds an optional tdesc operand
to fir.embox so the dynamic type can be provided while emboxing the entity.
This will be used in follow-up patch that lowers tbp calls for element of
polymorphic entities array.
Reviewed By: jeanPerier
Differential Revision: https://reviews.llvm.org/D135809
Archibald Elliott [Tue, 11 Oct 2022 13:02:33 +0000 (14:02 +0100)]
[ARM] Support fp16/bf16 using w constraint
fp16 and bf16 values can be used in GCC's inline assembly using the "w"
constraint, which means "VFP floating-point registers d0-d31" - fp16 and
bf16 values are stored in S registers (which alias the D registers).
This change ensures that LLVM is compatible with GCC for programs that
use fp16 and the 'w' constraint.
Differential Revision: https://reviews.llvm.org/D135662
Nikita Popov [Thu, 13 Oct 2022 08:48:18 +0000 (10:48 +0200)]
[FunctionAttrs] Make location classification more precise
Don't add argmem if the pointer is clearly not an argument (e.g.
a global). I don't think this makes a difference right now, but
gives more obvious results with D135780.
Nikita Popov [Thu, 13 Oct 2022 09:16:24 +0000 (11:16 +0200)]
[FunctionAttrs] Regenerate test checks (NFC)
Florian Hahn [Thu, 13 Oct 2022 09:19:29 +0000 (10:19 +0100)]
[ConstraintElim] Bail out for GEPs when index size > 64 bits.
Limit pointer decomposition to pointers with index sizes of at most 64
bits. int64_t is used for coefficients, so as long as the index size <=
64 bits we should be able to represent all pointer offsets.
Pointer decomposition is limited to inbounds GEPs, so if a index
computation would overflow the result is poison, so it doesn't matter
that the coefficient overflows.
This allows replacing MulOverflow with regular multiplications.
Nikita Popov [Thu, 13 Oct 2022 09:13:30 +0000 (11:13 +0200)]
[FunctionAttrs] Handle potential access of captured argument
We have to account for accesses to argument memory via captures.
I don't think there's any way to make this produce incorrect
results right now (because as soon as "other" is set, we lose the
ability to infer argmemonly), but this avoids incorrect results
once we have more precise representation.
Haojian Wu [Thu, 13 Oct 2022 08:01:47 +0000 (10:01 +0200)]
[clangd] Fix a crash in ExtractFunction tweak.
Differential Revision: https://reviews.llvm.org/D135857
Bjorn Pettersson [Wed, 12 Oct 2022 17:59:50 +0000 (19:59 +0200)]
[test][SLPVectorizer] Use -passes syntax in RUN lines. NFC
Bjorn Pettersson [Wed, 12 Oct 2022 17:17:02 +0000 (19:17 +0200)]
[test][IndVarSimplify] Use -passes syntax in RUN lines. NFC
Bjorn Pettersson [Wed, 12 Oct 2022 17:10:09 +0000 (19:10 +0200)]
[test][AggressiveInstCombine] Use -passes syntax in RUN lines. NFC
Bjorn Pettersson [Wed, 12 Oct 2022 14:26:27 +0000 (16:26 +0200)]
[test][DSE] Use -passes=dse instead of -dse in lit tests. NFC
Simon Tatham [Thu, 13 Oct 2022 08:18:31 +0000 (09:18 +0100)]
Propagate tied operands when copying a MachineInstr.
MachineInstr's copy constructor works by calling the addOperand method
to add each operand of the old MachineInstr to the new one, one by
one. But addOperand deliberately avoids trying to replicate ties
between operands, on the grounds that the tie refers to operands by
index, and the indices aren't necessarily finalized yet.
This led to a code generation fault when the machine pipeliner cloned
an Arm conditional instruction, and lost the tie between the output
register and the input value to be used when the condition failed to
execute.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D135434
Valentin Clement [Thu, 13 Oct 2022 08:39:29 +0000 (10:39 +0200)]
[flang] Embox derived-type directly to fir.class without fir.convert
non-polymorphic derived-type can call type-bound procedure with passed-object.
In that case, the derived-type is emboxed in order to be passed to the call.
Until now the emboxing was done to a fir.box followed by a fir.convert.
This patch update the createBox function so that we can directly embox to
a fir.class and avoid the extra fir.convert.
Reviewed By: PeteSteinfeld
Differential Revision: https://reviews.llvm.org/D135790
Nikita Popov [Thu, 13 Oct 2022 08:36:09 +0000 (10:36 +0200)]
[FunctionAttrs] Convert tests to use opaque pointers (NFC)
Conversion performed using the script at:
https://gist.github.com/nikic/
98357b71fd67756b0f064c9517b62a34
Nikita Popov [Thu, 13 Oct 2022 08:31:07 +0000 (10:31 +0200)]
[FunctionAttrs] Regenerate test checks (NFC)
Nikita Popov [Wed, 12 Oct 2022 15:22:11 +0000 (17:22 +0200)]
[FunctionAttrs] Account for memory effects of inalloca/preallocated
The code for inferring memory attributes on arguments claims that
inalloca/preallocated arguments are always clobbered:
https://github.com/llvm/llvm-project/blob/
d71ad4108056d685f48407447095d8d92fd7685d/llvm/lib/Transforms/IPO/FunctionAttrs.cpp#L640-L642
However, we would still infer memory attributes for the whole
function without taking this into account, so we could still end
up inferring readnone for the function. This adds an argument
clobber if there are any inalloca/preallocated arguments.
Differential Revision: https://reviews.llvm.org/D135783
Andi Drebes [Thu, 13 Oct 2022 08:11:43 +0000 (10:11 +0200)]
[mlir] MemRefToLLVM: Save / restore stack when lowering memref.copy
The MemRef to LLVM conversion pass emits `llvm.alloca` operations to promote MemRef descriptors to the stack when lowering `memref.copy` operations for operands which do not have a contiguous layout in memory. The original stack position is never restored after the allocations, which creates an issue when the copy operation is embedded into a loop with a high trip count, ultimately resulting in a segmentation fault due to the stack growing too large.
Below is as a minimal example illustrating the issue:
```
module {
func.func @main() {
%arg0 = memref.alloc() : memref<32x64xi64>
%arg1 = memref.alloc() : memref<16x32xi64>
%lb = arith.constant 0 : index
%ub = arith.constant 100000 : index
%step = arith.constant 1 : index
%slice = memref.subview %arg0[16,32][16,32][1,1] :
memref<32x64xi64> to memref<16x32xi64, #map>
scf.for %i = %lb to %ub step %step {
memref.copy %slice, %arg1 :
memref<16x32xi64, #map> to memref<16x32xi64>
}
return
}
}
```
When running the code above, e.g., with mlir-cpu-runner, the execution crashes with a segmentation fault:
```
$ mlir-opt \
--convert-scf-to-cf \
--convert-memref-to-llvm \
--convert-func-to-llvm
--convert-cf-to-llvm \
--reconcile-unrealized-casts <file> | \
mlir-cpu-runner \
-e main -entry-point-result=void \
--shared-libs=$PWD/build/lib/libmlir_c_runner_utils.so
[...]
Segmentation fault
```
This patch causes the code lowering a `memref.copy` operation in the MemRefToLLVM pass to emit a pair of matching `llvm.intr.stacksave` and `llvm.intr.stackrestore` operations around the promotion of memory descriptors and the subsequent call to `memrefCopy` in order to restore the stack to its original position after the call.
Reviewed By: ftynse
Differential Revision: https://reviews.llvm.org/D135756
Oleg Shyshkov [Thu, 13 Oct 2022 07:28:46 +0000 (07:28 +0000)]
[mlir] Remove iterator_types() method from LinalgStructuredInterface.
`getIteratorTypesArray` should be used instead. It's a better substitute for all the current usages of the interface.
The current `ArrayAttr iterator_types()` has a few problems:
* It creates an assumption operation has iterators types as an attribute, but it's not always the case. Sometime iterator types can be inferred from other attribute, or they're just static.
* ArrayAttr is an obscure contained and required extracting values in the client code.
* Makes it hard to migrate iterator types from strings to enums ([RFC](https://discourse.llvm.org/t/rfc-enumattr-for-iterator-types-in-linalg/64535/9)).
Concrete ops, like `linalg.generic` will still have iterator types as an attribute if needed.
As a side effect, this change helps a bit with migration to prefixed accessors.
Differential Revision: https://reviews.llvm.org/D135765
Benjamin Kramer [Thu, 13 Oct 2022 07:46:32 +0000 (09:46 +0200)]
[HLSL] Simplify code and fix unused variable warnings. NFC.
Leon Clark [Thu, 13 Oct 2022 07:02:24 +0000 (08:02 +0100)]
Precommit for SWDEV-353076: Add check directives to existing tests.
Add FileCheck directives to existing tests in preparation for new tests.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D135788
Martin Storsjö [Mon, 8 Aug 2022 09:55:57 +0000 (12:55 +0300)]
[MC] [Win64EH] Check that ARM64 prologs and epilogs have the right matching number of instructions
This matches what was done for the ARM implementation (where getting
the instruction sizes right is even more tricky, and hence needed
tighter testing).
This will allow catching any future cases where prologs and epilogs
don't match the instructions within them.
Differential Revision: https://reviews.llvm.org/D131394
Martin Storsjö [Wed, 12 Oct 2022 21:30:33 +0000 (00:30 +0300)]
[AArch64] Use encodeLogicalImmediate for forming the immediate to an AND. NFC.
Differential Revision: https://reviews.llvm.org/D135817
Martin Storsjö [Wed, 12 Oct 2022 12:03:16 +0000 (15:03 +0300)]
[AArch64] Remove dead code for inserting SEH opcodes for realignment. NFC.
If the stack is realigned, we've emitted a frame pointer and
already terminated the SEH prologue, making this dead code since
a07787c9a50c046e45921dd665f5a53a752bbc31.
The immediate to this SEH opcode was entirely bogus - we don't
know how many bytes the AND operation adjusts the SP, and by
doing "NumBytes & andMaskEncoded" (where andMaskEncoded was the
immediate bitpattern for the AND instruction), the immediate to the
opcode was total gibberish.
This hasn't had any practical effect, since the original stack
pointer always was restored from the frame pointer afterwards anyway.
Differential Revision: https://reviews.llvm.org/D135815
Martin Storsjö [Tue, 11 Oct 2022 12:20:59 +0000 (15:20 +0300)]
[AArch64] Fix aligning the stack after calling __chkstk
Whenever a call to __chkstk was made, the frame lowering previously
omitted the aligning (as NumBytes was reset to zero before doing
alignment).
This fixes https://github.com/llvm/llvm-project/issues/56182.
Differential Revision: https://reviews.llvm.org/D135687
Balazs Benics [Thu, 13 Oct 2022 06:41:31 +0000 (08:41 +0200)]
[analyzer] Workaround crash on encountering Class non-type template parameters
The Clang Static Analyzer will crash on this code:
```lang=C++
struct Box {
int value;
};
template <Box V> int get() {
return V.value;
}
template int get<Box{-1}>();
```
https://godbolt.org/z/5Yb1sMMMb
The problem is that we don't account for encountering `TemplateParamObjectDecl`s
within the `DeclRefExpr` handler in the `ExprEngine`.
IMO we should create a new memregion for representing such template
param objects, to model their language semantics.
Such as:
- it should have global static storage
- for two identical values, their addresses should be identical as well
http://eel.is/c%2B%2Bdraft/temp.param#8
I was thinking of introducing a `TemplateParamObjectRegion` under `DeclRegion`
for this purpose. It could have `TemplateParamObjectDecl` as a field.
The `TemplateParamObjectDecl::getValue()` returns `APValue`, which might
represent multiple levels of structures, unions and other goodies -
making the transformation from `APValue` to `SVal` a bit complicated.
That being said, for now, I think having `Unknowns` for such cases is
definitely an improvement to crashing, hence I'm proposing this patch.
Reviewed By: xazax.hun
Differential Revision: https://reviews.llvm.org/D135763
Xiang Li [Thu, 13 Oct 2022 00:02:56 +0000 (17:02 -0700)]
[HLSL] CodeGen hlsl cbuffer/tbuffer.
cbuffer A {
float a;
float b;
}
will be translated to a global variable.
Something like
struct CB_Ty {
float a;
float b;
};
CB_Ty A;
And all use of a and b will be replaced with A.a and A.b.
Only support none-legacy cbuffer layout now.
CodeGen for Resource binding will be in separate patch.
In the separate patch, resource binding will map the resource information to the global variable.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D130131
Kai Luo [Thu, 13 Oct 2022 03:51:50 +0000 (11:51 +0800)]
[clang][Module][AIX] Mark test unsupported since objc doesn't have xcoff support
Fixed error
```
Command Output (stderr):
--
fatal error: error in backend: Objective-C support is unimplemented for object file format
```
Source code in `clang/lib/CodeGen/CGObjCMac.cpp:5080`
```
case llvm::Triple::Wasm:
case llvm::Triple::GOFF:
case llvm::Triple::SPIRV:
case llvm::Triple::XCOFF:
case llvm::Triple::DXContainer:
llvm::report_fatal_error(
"Objective-C support is unimplemented for object file format");
}
```
Reviewed By: hubert.reinterpretcast
Differential Revision: https://reviews.llvm.org/D135848
Jordan Rupprecht [Thu, 13 Oct 2022 03:43:45 +0000 (20:43 -0700)]
[test] Fix a test that wasn't running
The functionality is fine (we don't run the breakpoint command twice), but this test forgot to call `FileCheck` and isn't checking the same value isn't there twice..
Yaxun (Sam) Liu [Tue, 11 Oct 2022 22:53:59 +0000 (18:53 -0400)]
[HIP] Detect HIP for Debian/Fedora
HIP is installed at /usr or /usr/local on Debin/Fedora,
and the version file is at {root}/share/hip/version.
Reviewed by: Artem Belevich
Differential Revision: https://reviews.llvm.org/D135796
Mehdi Amini [Thu, 6 Oct 2022 19:24:31 +0000 (19:24 +0000)]
Apply clang-tidy fixes for llvm-qualified-auto in SparseTensorCodegen.cpp (NFC)
Mehdi Amini [Thu, 6 Oct 2022 19:23:43 +0000 (19:23 +0000)]
Apply clang-tidy fixes for readability-identifier-naming in SparseBufferRewriting.cpp (NFC)
Kazu Hirata [Thu, 13 Oct 2022 02:20:38 +0000 (19:20 -0700)]
[clang] Use std::underlying_type_t (NFC)
Jan Svoboda [Tue, 11 Oct 2022 23:06:08 +0000 (16:06 -0700)]
[clang][deps] Respect VFS overlays in canonical preprocessing mode
The `-ivfsoverlay` flag was only being respected when the scanner was instructed to minimize the inputs. This patch respects that flag even in canonical preprocessing mode.
Depends on D135414.
Reviewed By: Bigcheese
Differential Revision: https://reviews.llvm.org/D135416
Jan Svoboda [Tue, 11 Oct 2022 23:02:10 +0000 (16:02 -0700)]
[clang][deps] Don't share in-memory VFS across scans
The dependency scanning worker may create an in-memory VFS and inject that into its `FileManager`. (Implemented in D109485.) This VFS currently persists between scans, which is not correct: in-memory files created in one scan could affect subsequent scans. This patch changes things so that the in-memory VFS is local to `DependencyScanningWorker::computeDependencies()`.
To test this, one would first scan for dependencies of a named module and then run second scan (on the same worker) for something unrelated, which would pick up the in-memory file created in the first scan. This set up is impossible to achieve with `clang-scan-deps`. We could set this up in `ToolingTests`, but the scanner needs to access the physical filesystem to store `.pcm` files. AFAIK we don't have a good way to propagate something like `%t` into unit tests to make that feasible. (This could be achieved with something like the virtualized `OutputManager`.)
Reviewed By: Bigcheese
Differential Revision: https://reviews.llvm.org/D135414
Rafael Auler [Sat, 27 Aug 2022 01:39:36 +0000 (18:39 -0700)]
[BOLT] Add pass to fix ambiguous memory references
This adds a round of checks to memory references, looking for
incorrect references to jump table objects. Fix them by replacing the
jump table reference with another object reference + offset.
This solves bugs related to regular data references in code
accidentally being bound to a jump table, and this reference being
updated to a new (incorrect) location because we moved this jump
table.
Fixes #55004
Reviewed By: #bolt, maksfb
Differential Revision: https://reviews.llvm.org/D134098
Rafael Auler [Fri, 16 Sep 2022 23:20:00 +0000 (16:20 -0700)]
[BOLT][NFC] Refactor creation of symbol+addend references
Put code that creates references to symbol+addend behind MCPlusBuilder.
Will use this later in validate memory references pass.
Reviewed By: #bolt, maksfb, yota9
Differential Revision: https://reviews.llvm.org/D134097
Matthias Springer [Thu, 13 Oct 2022 01:26:25 +0000 (10:26 +0900)]
[mlir][transforms] TopologicalSort: Support ops from different blocks
This change allows analyzing ops from different block, in particular when used in programs that have `cf` branches.
Differential Revision: https://reviews.llvm.org/D135644
Jeff Niu [Sat, 8 Oct 2022 02:28:41 +0000 (19:28 -0700)]
[mlir][scf] Match any constants instead of arith.constant
By matching `arith.constant` specifically, SCF canonicalizers/folders
are incompatible with other kinds of constants. Use the generic
matchers instead.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D135517
Matt Arsenault [Tue, 11 Oct 2022 01:03:46 +0000 (18:03 -0700)]
llvm-reduce: Color output of child processes
When reducing llvm-reduce with llvm-reduce, it can be confusing
to figure out what lines are printed by the child or parent. Not
sure this is the most reliable way to set and restore this.
Jeff Niu [Tue, 11 Oct 2022 17:40:23 +0000 (10:40 -0700)]
[mlir][ods] Assert on static getAttrName methods
This patch makes ODS insert an assert when calling static `get*AttrName`
methods with a `OperationName` that the name is the same as the op's.
This prevents accidentally passing the wrong kind of name and getting an
erroneous attribute name.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D135698
Jeff Niu [Sat, 1 Oct 2022 00:02:05 +0000 (17:02 -0700)]
[mlir][ods] Allow custom directives in optional groups
Attributes and types only (so far). Since `struct` and `params` are
allowed, it makes sense to allow custom directives as long as their
arguments contain at least one bound argument.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D135001
LLVM GN Syncbot [Thu, 13 Oct 2022 00:34:23 +0000 (00:34 +0000)]
[gn build] Port
573a5de7551c
Matt Arsenault [Tue, 11 Oct 2022 02:42:17 +0000 (19:42 -0700)]
llvm-reduce: Add opcode reduction pass
Try some dumb strength reductions to "simpler" opcodes.
Make some opcode substitutions I typically try to get smaller
MIR out of codegen. This is a bit target specific and I have a
lot of increasingly target specific modifications I try
during manual reduction.
Sanjoy Das [Wed, 12 Oct 2022 05:33:45 +0000 (22:33 -0700)]
Correctly model undefined behavior in {tensor|memref}.dim
These operations have undefined behavior if the index is not less than the rank of the source tensor / memref, so they cannot be freely speculated like they were before this patch. After this patch we speculate them only if we can prove that they don't have UB.
Depends on D135505.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D135748
Henrique Bucher [Mon, 10 Oct 2022 16:31:06 +0000 (11:31 -0500)]
Summary: This documentation patch adds information to allow remote users to also use the plugin as it will be invisible to them using the current instructions. It solves issue #58252.
This documentation patch adds information to allow remote users to also use the plugin as it will be invisible to them using the current instructions. It solves issue #58252.
Reviewed By: JDevlieghere
Differential Revision: https://reviews.llvm.org/D135577
Matt Arsenault [Wed, 12 Oct 2022 22:27:28 +0000 (15:27 -0700)]
AMDGPU: Fix assertion on <1 x i16> vectors
Fixes issue 58331.
Matt Arsenault [Tue, 11 Oct 2022 17:41:09 +0000 (10:41 -0700)]
AMDGPU: Fix hazard with v_accvgpr_write_b32 and inline asm VGPR defs
If inline asm has a VGPR def, it must have come from a VGPR write
somewhere inside the asm. This should be further extended to all
read after write hazards.
Matt Arsenault [Sat, 8 Oct 2022 17:21:33 +0000 (10:21 -0700)]
llvm-reduce: Don't write out IR to score IR complexity
In a testcase I'm working on, the old write out and count IR lines
was taking about 200-300ms per iteration. This drops it out of the
profile.
This doesn't account for everything, but it doesn't seem to matter.
We should probably try to account for metadata and constantexpr tree
depths.
Matt Arsenault [Mon, 10 Oct 2022 18:02:47 +0000 (11:02 -0700)]
llvm-reduce: Improve delta pass flag handling
Verify all the requested passes exist before trying to run any.
For long reductions, it was really annoying for it to get halfway through
and then I come back later to an incomplete reduction.
Also add a new skip-delta-passes flag. Most of the time I want to opt out
of specific reductions, rather than run a select few.
Jonathon Penix [Tue, 19 Jul 2022 18:47:25 +0000 (11:47 -0700)]
[Flang] Add -fconvert option to swap endianness for unformatted files.
To accomplish this, this patch creates an optional list of environment
variable default values to be set by the runtime to allow directly using
the existing runtime implementation of FORT_CONVERT for I/O conversions.
Vitaly Buka [Wed, 12 Oct 2022 23:45:07 +0000 (16:45 -0700)]
[test][openmp] Relax check for tsan reports count
I see 2 reports time to time.
Wiktor Garbacz [Wed, 12 Oct 2022 23:36:37 +0000 (16:36 -0700)]
[LeakSanitizer] Capture calling thread SP early to avoid false negatives.
As shown in https://github.com/llvm/llvm-project/issues/42932 dead
pointers might be overlapped by a new stack frame inside CheckForLeaks,
which does not use bytes with pointers. This leads to false negatives.
It's not a full solution for the problem as it does not solve
"overlapping" new/old frames for frames below the CheckForLeaks and in
other threads. It should improve leaks found in direct callers of
__lsan_do_leak_check.
Differential Revision: https://reviews.llvm.org/D130237
Alex Lorenz [Wed, 12 Oct 2022 22:56:44 +0000 (15:56 -0700)]
[clang][test] NFC, check in darwin-ld-platform-version-macos-requires-darwin.c should be more permissive
some Darwin CI uses older SDK that reports different SDK version here
https://green.lab.llvm.org/green/job/clang-stage1-RA/
Florian Mayer [Wed, 12 Oct 2022 22:52:27 +0000 (15:52 -0700)]
[MC] only run ELF/AArch64 if supported
Steven Wu [Wed, 12 Oct 2022 22:25:43 +0000 (15:25 -0700)]
[CMake] Fix FindGRPC cmake module to allow different layering
Take the library target out of `generate_protos` function so the caller
can decide where to layer the library using the source generated from
the function.
Fixes: https://github.com/llvm/llvm-project/issues/58075
Reviewed By: sammccall
Differential Revision: https://reviews.llvm.org/D135712