Instructions between two adjacent loops will be hoisted above the first
loop, or sunk below the second to facilitate loop fusion. Hoisting will
be attempted for an instruction that dominates the first loop.
Otherwise, sinking this instructions will be attempted.
Instructions with side effects will not be considered for sinking or
hoisting. Hoisting/sinking of any instructions between loops will only
be performed if all the instructions can be moved. As well,
sinking/hoisting is considered for each instruction in isolation,
without taking into account sinking/hoisting decisions for other
instructions in the preheader.
Differential Revision: https://reviews.llvm.org/D118076
I wondered why I hadn't seen the link failures in D130571
<https://reviews.llvm.org/D130571> on Solaris/sparcv9 and found that
Solaris, unlike Linux, doesn't use `-Wl,-z,defs` when linking shared
objects. The three affected shared objects indeed have undefined
references to several 64-bit atomics.
This patch will expose such issues in the future by defaulting to
`-Wl,-z,defs` on Solaris, as recommended in `ld(1)`.
Tested on `sparcv9-sun-solaris2.11` and `amd64-pc-solaris2.11`.
Differential Revision: https://reviews.llvm.org/D130572
Otherwise we get invalid results for ODR checks. See changed test for an
example: despite the fact that we merge the first concept, its **uses**
were considered different by `Profile`, leading to redefinition errors.
After this change, canonical decl for a concept can come from a
different module and may not be visible. This behavior looks suspicious,
but does not break any tests. We might want to add a mechanism to make
the canonical concept declaration visible if we find code that relies on
this invariant.
Additionally make sure we always merge with the canonical declaration to
avoid chains of merged concepts being reported as redefinitions. An
example was added to the test.
Also change the order of includes in the test. Importing a moduralized
header before its textual part causes the include guard macro to be
exported and the corresponding `#include` becomes a no-op.
Reviewed By: ChuanqiXu
Differential Revision: https://reviews.llvm.org/D130585
We have seen the llvm test fail every so often on our bots:
https://lab.llvm.org/buildbot/#/builders/173/builds/6711
This happens because `flang-new -fc1 -S` does not wait for the
pipe to finish being written to/does not read out all the content.
You can see this in the output, cat comes after flang:
+ /home/tcwg-buildbot/worker/flang-aarch64-release/build/bin/not /home/tcwg-buildbot/worker/flang-aarch64-release/build/bin/flang-new -fc1 -S - -o -
+ cat /home/tcwg-buildbot/worker/flang-aarch64-release/llvm-project/flang/test/Driver/input-from-stdin-llvm.ll
error: Invalid input type - expecting a Fortran file
This means that cat gets SIGPIPE which causes it to exit with
code 141 and that's the final result due to pipefail.
flang isn't wrong to exit early (I think some modes of grep also do this)
and we only care about flang's exit code. So disable pipefail for
the stdin testing.
Reviewed By: awarzynski
Differential Revision: https://reviews.llvm.org/D130619
When building on Linux/sparc64, the 32-bit `libclang_rt.asan.so`,
`libclang_rt.ubsan_minimal.so`, and `libclang_rt.ubsan_standalone.so`
failed to link with undefined references to 64-bit atomics, which `clang`
cannot inline. Even D130569 <https://reviews.llvm.org/D130569> didn't help
because those libraries are linked with `-nodefaultlibs`, so dependent
libraries need to be added explicitly.
That's what this patch does.
Tested on `sparc64-unknown-linux-gnu` and `sparcv9-sun-solaris2.11`.
Differential Revision: https://reviews.llvm.org/D130571
This patch refactors the runtime support for GET_COMMAND_ARGUMENT to
have a single entry point instead of 2. It also updates lowering
accordingly.
This makes it easier to handle dynamically optional arguments. See also
https://reviews.llvm.org/D118777
Differential Revision: https://reviews.llvm.org/D130475
While working on D118450 <https://reviews.llvm.org/D118450>, I noticed that
`sys::getHostCPUName` lacks SPARC support.
This patch implements it. The code is taken from/inspired by GCC's
`gcc/config/sparc/driver-sparc.cc`. There's one caveat: since LLVM, unlike
GCC, doesn't support the SPARC-M7, -S7, and -M8 CPUs, I map all those to
the latest supported one (UltraSparc T4/`niagara4`).
Tested on `sparcv9-sun-solaris2.11` and `sparc64-unknown-linux-gnu` by
running `savcov --version` on
- Netra SPARC S7-2 (SPARC-S7, Solaris 11.4)
- SPARC T5-2 (SPARC T5, Solaris 11.4)
- SPARC Enterprise T5220 (UltraSPARC T2, Solaris 11.3)
- SPARC T5 (UltraSPARC T5, Debian sid)
- SPARC T3 (UltraSPARC T3, Debian sid)
- SPARC Enterprise T5220 (Debian sid)
Differential Revision: https://reviews.llvm.org/D130272
SimplifyDemandedBits currently early-outs for multi-use values beyond the root node (just returning the knownbits), which is missing a number of optimizations as there are plenty of cases where we can still simplify when initially demanding all elements/bits.
@lenary has confirmed that the test cases in aea-erratum-fix.ll need refactoring and the current increase codegen is not a major concern.
Differential Revision: https://reviews.llvm.org/D129765
Normally, generic processor does not have any SubtargetFeature. And it
can just generate most basic instructions which have no Predicates to
guard.
But it needs to enbale predicate for the btsti16 instruction as one of the most basic instructions.
Or the generic processor can't finish codegen process. So Add FeatureBTST16 SubtargetFeature to generic ProcessorModel.
Check that 8-bit and 16-bit atomics also work as expected. Also
fix the alignment on the 64-bit tests -- testing unaligned atomics
wasn't intended here.
The SGI page doesn't exist anymore and isn't really relevant at this day
and age.
While at it, added the "other" main C++ website and moved all URLs to
HTTPS.
Without this patch, clang will not wrap in an ElaboratedType node types written
without a keyword and nested name qualifier, which goes against the intent that
we should produce an AST which retains enough details to recover how things are
written.
The lack of this sugar is incompatible with the intent of the type printer
default policy, which is to print types as written, but to fall back and print
them fully qualified when they are desugared.
An ElaboratedTypeLoc without keyword / NNS uses no storage by itself, but still
requires pointer alignment due to pre-existing bug in the TypeLoc buffer
handling.
---
Troubleshooting list to deal with any breakage seen with this patch:
1) The most likely effect one would see by this patch is a change in how
a type is printed. The type printer will, by design and default,
print types as written. There are customization options there, but
not that many, and they mainly apply to how to print a type that we
somehow failed to track how it was written. This patch fixes a
problem where we failed to distinguish between a type
that was written without any elaborated-type qualifiers,
such as a 'struct'/'class' tags and name spacifiers such as 'std::',
and one that has been stripped of any 'metadata' that identifies such,
the so called canonical types.
Example:
```
namespace foo {
struct A {};
A a;
};
```
If one were to print the type of `foo::a`, prior to this patch, this
would result in `foo::A`. This is how the type printer would have,
by default, printed the canonical type of A as well.
As soon as you add any name qualifiers to A, the type printer would
suddenly start accurately printing the type as written. This patch
will make it print it accurately even when written without
qualifiers, so we will just print `A` for the initial example, as
the user did not really write that `foo::` namespace qualifier.
2) This patch could expose a bug in some AST matcher. Matching types
is harder to get right when there is sugar involved. For example,
if you want to match a type against being a pointer to some type A,
then you have to account for getting a type that is sugar for a
pointer to A, or being a pointer to sugar to A, or both! Usually
you would get the second part wrong, and this would work for a
very simple test where you don't use any name qualifiers, but
you would discover is broken when you do. The usual fix is to
either use the matcher which strips sugar, which is annoying
to use as for example if you match an N level pointer, you have
to put N+1 such matchers in there, beginning to end and between
all those levels. But in a lot of cases, if the property you want
to match is present in the canonical type, it's easier and faster
to just match on that... This goes with what is said in 1), if
you want to match against the name of a type, and you want
the name string to be something stable, perhaps matching on
the name of the canonical type is the better choice.
3) This patch could expose a bug in how you get the source range of some
TypeLoc. For some reason, a lot of code is using getLocalSourceRange(),
which only looks at the given TypeLoc node. This patch introduces a new,
and more common TypeLoc node which contains no source locations on itself.
This is not an inovation here, and some other, more rare TypeLoc nodes could
also have this property, but if you use getLocalSourceRange on them, it's not
going to return any valid locations, because it doesn't have any. The right fix
here is to always use getSourceRange() or getBeginLoc/getEndLoc which will dive
into the inner TypeLoc to get the source range if it doesn't find it on the
top level one. You can use getLocalSourceRange if you are really into
micro-optimizations and you have some outside knowledge that the TypeLocs you are
dealing with will always include some source location.
4) Exposed a bug somewhere in the use of the normal clang type class API, where you
have some type, you want to see if that type is some particular kind, you try a
`dyn_cast` such as `dyn_cast<TypedefType>` and that fails because now you have an
ElaboratedType which has a TypeDefType inside of it, which is what you wanted to match.
Again, like 2), this would usually have been tested poorly with some simple tests with
no qualifications, and would have been broken had there been any other kind of type sugar,
be it an ElaboratedType or a TemplateSpecializationType or a SubstTemplateParmType.
The usual fix here is to use `getAs` instead of `dyn_cast`, which will look deeper
into the type. Or use `getAsAdjusted` when dealing with TypeLocs.
For some reason the API is inconsistent there and on TypeLocs getAs behaves like a dyn_cast.
5) It could be a bug in this patch perhaps.
Let me know if you need any help!
Signed-off-by: Matheus Izvekov <mizvekov@gmail.com>
Differential Revision: https://reviews.llvm.org/D112374
Now the prefixes have the same length
and they'll be easier to distinguish when
`GlobalOp` tests will be added
Depends on D129778
Reviewed By: clementval
Differential Revision: https://reviews.llvm.org/D130087
This change decouples common functionality for convering abstract
results, so it can be reused later.
Depends on D129485
Reviewed By: clementval, jeanPerier
Differential Revision: https://reviews.llvm.org/D129778
Combine the recently added utilities for folded-by-construction affine
operations with the attribute-based Range to enable more folding. This
decreases the amount of emitted code but has little effect on test
precisely because the tests are not checking for the spurious constants.
The difference in the shape of affine maps comes from the internals of
affine folding.
Depends on D129633
Reviewed By: nicolasvasilache
Differential Revision: https://reviews.llvm.org/D130167
While most of methods in ViewLikeInterface accept an `OpFoldResult` for
the offset/size/stride that may be static, represented as `Attribute`,
or dynamic, represented as `Value`, the `Range` abstraction only
accepted `Values`. This can often lead to known-constant
offset/size/strides being materialized into constant operations and
hinder further constant propagation without explicitly running the
constant folding pass. This often leads to a more complicated than
necessary addressing code being emitted. Switch `Range` to use
`OpFoldResult`. Code that uses `Range` currently keeps materializing the
constants to minimize the effect of this change on the IR. Further
commits will make use of this.
Reviewed By: nicolasvasilache, mravishankar
Differential Revision: https://reviews.llvm.org/D129633
The structured op splitting transformation is conceptually similar to
tiling in the sense that it decomposes the iteration space of the
original op into several parts. Therefore, it is possible to implement
it using the TilingInterface to operate on iteration spaces and their
parts. However, the implementation also requires to pass updated input
operands, which is not supported by the interface, so the implementation
currently remains Linalg-specific.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D129564
Given a patch like D129506, using instructions not valid for the current
feature set becomes an error. This updates the Arm hint-space
instructions for pac/bti to require thumbv7m as opposed to 8.1-m.main, to
make them valid when compiling for thumbv7m with -mbranch-protection.
Differential Revision: https://reviews.llvm.org/D129692
This adds a +atomic-32 target feature, which instructs LLVM to assume
that lock-free 32-bit atomics are available for this target, even
if they usually wouldn't be.
If only atomic loads/stores are used, then this won't emit libcalls.
If atomic CAS is used, then the user is responsible for providing
any necessary __sync implementations (e.g. by masking interrupts
for single-core privileged use cases).
See https://reviews.llvm.org/D120026#3674333 for context on this
change. The tl;dr is that the thumbv6m target in Rust has
historically made atomic load/store only available, which is
incompatible with the change from D120026, which switched these to
use libatomic.
Differential Revision: https://reviews.llvm.org/D130480
- Place rules under rule::lhs::rhs__rhs__rhs
- Change mangling of keywords to ALL_CAPS (needed to turn keywords that appear
alone on RHS into valid identifiers)
- Make enums implicitly convertible to underlying type (though still scoped,
using alias tricks)
In principle this lets us exhaustively write a switch over all rules of a NT:
switch ((rule::declarator)N->rule()) {
case rule::declarator::noptr_declarator:
...
}
In practice we don't do this anywhere yet as we're often switching over multiple
nonterminal kinds at once.
Differential Revision: https://reviews.llvm.org/D130414
This is still undocumented and unsupported, but if someone passed it
before you would end up with a missing file error since this takes an
argument that wouldn't be handled.
Differential Revision: https://reviews.llvm.org/D130606
We are supporting quadword lock free atomics on AIX. For the situation that users on AIX are using a libatomic that is lock-based for quadword types, we can't enable quadword lock free atomics by default on AIX in case user's new code and existing code accessing the same shared atomic quadword variable, we can't guarentee atomicity. So we need an option to enable quadword lock free atomics on AIX, thus we can build a quadword lock-free libatomic(also for advanced users considering atomic performance critical) for users to make the transition smooth.
Reviewed By: shchenz
Differential Revision: https://reviews.llvm.org/D127189
Added alloca optimization which was missed during the implemenation of D112098.
Reviewed By: vitalybuka
Differential Revision: https://reviews.llvm.org/D130503
The IR Verifier requires that every call instruction to an inlineable
function (among other things, its implementation must be visible in the
translation unit) must also have !dbg metadata attached to it. When
parallelizing, Polly emits calls to OpenMP runtime function out of thin
air, or at least not directly derived from a bounded list of previous
instruction. While we could search for instructions in the SCoP that has
some debug info attached to it, there is no guarantee that we find any.
Our solution is to generate a new DILocation that points to line 0 to
represent optimized code.
The OpenMP function implementation is usually not available in the
user's translation unit, but can become visible in an LTO build. For
the bug to appear, libomp must also be built with debug symbols.
IMHO, the IR verifier rule is too strict. Runtime functions can
also be inserted by other optimization passes, such as
LoopIdiomRecognize. When inserting a call to e.g. memset, it uses the
DebugLoc from a StoreInst from the unoptimized code. It is not
required to have !dbg metadata attached either.
Fixes#56692
Jakob Johnson