1. Break MUL with specific constant to a SLLI and an ADD/SUB on riscv32
with the M extension.
2. Break MUL with specific constant to two SLLI and an ADD/SUB, if the
constant needs a pair of LUI/ADDI to construct.
Reviewed by: craig.topper
Differential Revision: https://reviews.llvm.org/D93619
Since we're leaving this as ReadWrite with the "write" reflecting
the update to MXCSR, we shouldn't say it only writes arg memory.
Hopefully this fixes the issue reported in post-commit in D93571.
Treat a non-atomic volatile load and store as a relaxed atomic at
system scope for the address spaces accessed. This will ensure all
relevant caches will be bypassed.
A volatile atomic is not changed and still only bypasses caches upto
the level specified by the SyncScope operand.
Differential Revision: https://reviews.llvm.org/D94214
exception thrown during construction in a new-expression.
Instead, when performing deallocation function lookup for a
new-expression, ignore all destroying operator delete candidates, and
fall back to global operator delete if there is no member operator
delete other than a destroying operator delete.
Use of destroying operator delete only makes sense when there is an
object to destroy, which there isn't in this case. The language wording
doesn't cover this case; this oversight has been reported to WG21, with
the approach in this patch as the proposed fix.
this will make sure that all of the functions are using the correct
prototypes. Explained much better in the comments of this diff:
https://reviews.llvm.org/D94195
Summary:
The new macro also inserts the C alias for the C++ implementations
without needing an objcopy based post processing step. The CMake
rules have been updated to reflect this. More CMake cleanup can be
taken up in future rounds and appropriate TODOs have been added for them.
Reviewers: mcgrathr, sivachandra
Subscribers:
We were mishandling the case where both `__tbss` and `__thread_data` sections were
present.
TLVP relocations should be encoded as offsets from the start of `__thread_data`,
even if the symbol is actually located in `__thread_bss`. Previously, we were
writing the offset from the start of the containing section, which doesn't
really make sense since there's no way `tlv_get_addr()` can know which section a
given `tlv$init` symbol is in at runtime.
In addition, this patch ensures that we place `__thread_data` immediately before
`__thread_bss`. This is what ld64 does, likely for performance reasons. Zerofill
sections must also be at the end of their segments; we were already doing this,
but now we ensure that `__thread_bss` occurs before `__bss`, so that it's always
possible to have it contiguous with `__thread_data`.
Fixes llvm.org/PR48657.
Reviewed By: #lld-macho, thakis
Differential Revision: https://reviews.llvm.org/D94329
Match the legacy PM in running various ObjC ARC passes.
This requires making some module passes into function passes. These were
initially ported as module passes since they add function declarations
(e.g. https://reviews.llvm.org/D86178), but that's still up for debate
and other passes do so.
Reviewed By: ahatanak
Differential Revision: https://reviews.llvm.org/D93743
Large integers are generated in Circt commonly which exceed 4kbits. This aligns the maximum bitwidth in MLIR and LLVM.
Reviewed By: rriddle, lattner, mehdi_amini
Differential Revision: https://reviews.llvm.org/D94116
reference binding to an expression.
We need to know the array bound in order to determine whether the
parameter type is reference-compatible with the argument type, so we
need to trigger instantiation in this case.
The pipe signal handler must be installed before any other handlers are
registered. This is because the Unix RegisterHandlers function does not
perform a sigaction() for SIGPIPE unless a one-shot handler is present,
to allow long-lived processes (like lldb) to fully opt-out of llvm's
SIGPIPE handling and ignore the signal safely.
Fixes a bug introduced in D70277.
Tested by running Nick's test case:
% xcrun ./bin/clang -E -fno-integrated-cc1 x.c | tee foo.txt | head
I verified that child cc1 process exits with IO_ERR, and that the parent
recognizes the error code, exiting cleanly.
Differential Revision: https://reviews.llvm.org/D94324
The point of LWG2511 is basically just to make sure that we use
`tuple<Args&&...>` instead of `tuple<Args...>` in a couple of places
inside `scoped_allocator_adaptor` and inside `pair`.
As far as I can tell, this has been true for libc++
since EricWF's D27612 (and maybe even earlier than that).
I accidentally disabled this feature-test macro in my D93830,
due to a rebasing conflict. It had been enabled by my D93815,
and should have remained enabled.
In 7cd67904f7, we removed the unnecessary nullptr checks from the libc++abi
definition of operator delete, but we forgot to update the definition in
libc++ (damn code duplication!). Then, in d4a1e03c5f, I synced the
definitions across libc++ and libc++abi, but I did it the wrong way around.
I re-added the if() checks to libc++abi instead of removing them from libc++.
In ef74f0fdc3, we re-removed the if() check from operator delete, but
only in libc++abi. This patch corrects this mess and removes it
consistently in libc++ and libc++abi.
Differential Revision: https://reviews.llvm.org/D93473
Adding the ability to index the base address brings these operations closer
to the transfer read and write semantics (with lowering advantages), ensures
more consistent use in vector MLIR code (easier to read), and reduces the
amount of code duplication to lower memrefs into base addresses considerably
(making codegen less error-prone).
Reviewed By: ThomasRaoux
Differential Revision: https://reviews.llvm.org/D94278
This revision adds a new `initialize(MLIRContext *)` hook to passes that allows for them to initialize any heavy state before the first execution of the pass. A concrete use case of this is with patterns that rely on PDL, given that PDL is compiled at run time it is imperative that compilation results are cached as much as possible. The first use of this hook is in the Canonicalizer, which has the added benefit of reducing the number of expensive accesses to the context when collecting patterns.
Differential Revision: https://reviews.llvm.org/D93147
Use custom mlir runner init/destroy functions to safely init and destroy shared libraries loaded by the JitRunner.
This mechanism is ignored for Windows builds (for now) because init/destroy functions are not exported, and library unloading relies on static destructors.
Re-submit https://reviews.llvm.org/D94270 with a temporary workaround for windows
Differential Revision: https://reviews.llvm.org/D94312
@ikudrin enabled support for dwarf64 in D87011. Adding a clang flag so it can be used through that compilation pass.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D90507
This annotates the dump methods in APINotes to indicate that they are
unused as they are meant for debugging purposes. This avoids an
unnecessary warning.
This test was failing in our internal CI, since our driver does not default to
C11. Adding this switch fixes the issue.
Differential Revision: https://reviews.llvm.org/D94327
Change the implementation of LinalgOp with TensorReshapeOp by
expansion to be more modular and easier to follow.
Differential Revision: https://reviews.llvm.org/D93748
d9ce31ae7d (D94322) removed the check because I thought it was dead
due to checking the existance of a variable (which always existed).
This causes LLDB tests to fail as they set NO_INSTALL_RPATH because
they're never meant to be installed, but we still would end up using
the install rpath.
Add the check back and make it explicitly check for an empty value
to make the purpose clearer and avoid implicit test for a false/true
value.
Differential Revision: https://reviews.llvm.org/D94326
The existing verification of reshape ops in linalg (linalg.reshape and
linalg.tensor_reshape) allows specification of illegal ops, where
- A dynamic dimension is expanded into multiple dynamic
dimensions. This is ill-specified.
- A static dimension is expanded into dynamic dimension or viceversa,
- The product of extents of the static dimensions in the expanded type
doesnt match the static dimension of the collapsed type.
Making all of these illegal. This also implies that some pessimization
in canonicalization due to incomplete semantics of the operation can
be dropped.
Differential Revision: https://reviews.llvm.org/D93724
Reverted check for empty CMAKE_BUILD_RPATH fixed.
When `BUILD_WITH_INSTALL_RPATH` is enabled it prevents using a custom rpath only
for the build tree as the install rpath will be used. This makes it impossible to run a
runtimes build when compiling with Clang and wanting the installed rpath to be
empty (i.e. `-DCMAKE_BUILD_RPATH="<some path>" -DCMAKE_SKIP_INSTALL_RPATH=ON`).
Disable `BUILD_WITH_INSTALL_RPATH` when `CMAKE_BUILD_RPATH` is non-empty to
allow for such build scenarios.
Differential Revision: https://reviews.llvm.org/D94322
The ISel patterns we have for truncating to i1's under MVE do not seem
to be correct. Instead custom lower to icmp(ne, and(x, 1), 0).
Differential Revision: https://reviews.llvm.org/D94226
Internal subprograms have explicit interfaces. If an internal subprogram has
an alternate return, we check its explicit interface. But we were not
putting the label values of alternate returns into the actual argument.
I fixed this by changing the definition of actual arguments to be able
to contain a common::Label and putting the label for an alternate return
into the actual argument.
I also verified that we were already doing all of the semantic checking
required for alternate returns and removed a "TODO" for this.
I also added the test altreturn06.f90.
Differential Revision: https://reviews.llvm.org/D94017
Remove the stale LLDB-Info.plist which was only used by TestHelp.py. The
latter would try to parse the version number from the plist and use that
to verify the version in the help output. Of course this never matched
so it would fall back to matching any arbitrary version.
This patch does *not* change the real LLDB-Info.plist.in file which is
used for the LLDB Framework.
This patch updates `allocate_shared` to call `allocator_traits::construct`
when creating the object held inside the shared_pointer, and
`allocator_traits::destroy` when destroying it. This resolves
the part of P0674R1 that was originally filed as LWG2070.
This change is landed separately from the rest of P0674R1 because it is
incredibly tricky from an ABI perspective.
This is the reason why this change is so tricky is that we previously
used EBO in a compressed pair to store both the allocator and the object
type stored in the `shared_ptr`. However, starting in C++20, P0674
requires us to use Allocator construction for initializing the object type.
That requirement rules out the use of the EBO for the object type, since
using the EBO implies that the base will be initialized when the control
block is initialized (and hence we can't do it through Allocator construction).
Hence, supporting P0674 requires changing how we store the object type
inside the control block, which we do while being ABI compatible by using
some trickery with a properly aligned char buffer.
Fixes https://llvm.org/PR41900
Supersedes https://llvm.org/D62760
Differential Revision: https://reviews.llvm.org/D91201
This reverts commit 0ebc1fb29f.
The behaviour should have been the same as before unless specifying CMAKE_BUILD_RPATH,
which was previously broken.
However, this seems to have broken builds for some people that don't specify it.
Reverting until I can investigate.
Differential Revision: https://reviews.llvm.org/D94319
This was suggested to prepare for D93975.
By moving the start value creation to widenPHInstruction, we set the
stage to manage the start value directly in VPWidenPHIRecipe, which be
used subsequently to set the 'resume' value for reductions during
epilogue vectorization.
It also moves RdxDesc to the recipe, so we do not have to rely on Legal
to look it up later.
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D94175
Bump the required SWIG version to 3. If my memory serves me well we last
bumped the required SWIG version to 2 for Python 3. At that time SWIG 3
had already been around for a while so everyone I know was already using
that.
It appears that SWIG 3 is the only version that officially supports
C++11 which we're using in the typemap. SWIG 3 was released in 2014 so I
think it's reasonable to make that the minimum required version.
https://bugs.llvm.org/show_bug.cgi?id=48685
Differential revision: https://reviews.llvm.org/D94244
It's still a little confusing because in many cases C++17 and C++20
have different values, and libc++ implements the C++17 behavior but
not the C++20 behavior; 'unimplemented' can't represent that scenario.
Ultimately we probably ought to completely redesign the script to be
in terms of paper numbers, rather than language revisions, and make
it generate the CSV files like "Cxx2aStatusPaperStatus.csv" as well.
Most newly added macros are unimplemented. I've marked a few as implemented,
though, based on my reading of the code; for example I was pretty sure
`__cpp_lib_latch` is implemented since we have `<latch>`.
Differential Revision: https://reviews.llvm.org/D93830
Continue the convergence between LLVM dialect and built-in types by replacing
the bfloat, half, float and double LLVM dialect types with their built-in
counterparts. At the API level, this is a direct replacement. At the syntax
level, we change the keywords to `bf16`, `f16`, `f32` and `f64`, respectively,
to be compatible with the built-in type syntax. The old keywords can still be
parsed but produce a deprecation warning and will be eventually removed.
Depends On D94178
Reviewed By: mehdi_amini, silvas, antiagainst
Differential Revision: https://reviews.llvm.org/D94179
Kazu Hirata