Instead of looking up the vector loop using the header, keep track of
the current vector loop in VPTransformState. This removes the
requirement for the vector header block being part of the loop up front.
A follow-up patch will move the code to generate the Loop object for the
vector loop to VPRegionBlock.
Depends on D121619.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D121621
Sometimes we would like to run post-processing repeatedly on the original sample profile for tuning. In order to avoid regenerating the original profile from scratch every time, this change adds the support of reading in the original profile (called symbolized profile) and running the post-processor on it.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D121655
When two local symbols (think: file-scope static functions, or functions in
unnamed namespaces) with the same name in two different translation units
both needed thunks, ld64.lld previously created external thunks for both
of them. These thunks ended up with the same name, leading to a duplicate
symbol error for the thunk symbols.
Instead, give thunks for local symbols local visibility.
(Hitting this requires a jump to a local symbol from over 128 MiB away.
It's unlikely that a single .o file is 128 MiB large, but with ICF
you can end up with a situation where the local symbol is ICF'd with
a symbol in a separate translation unit. And that can introduce a
large enough jump to require a thunk.)
Fixes PR54599.
Differential Revision: https://reviews.llvm.org/D122624
Add documentation describing how to
- Use `llvm-remark-size-diff`
- Interpret the output from the tool
Differential Revision: https://reviews.llvm.org/D122744
Currently, we have two different lists of features each CPU supports...
and those lists aren't consistent. This patch assumes AArch64.td is
right, and tries to fix AArch64TargetParser to match.
It's hard to find documentation for the right features, but reviewers
have confirmed these changes.
Probably we should try to unify the two lists at some point, but
synchronizing them seems like a prerequisite to that anyway.
Differential Revision: https://reviews.llvm.org/D122274
Updating the diagnostics as per the feedback on
https://reviews.llvm.org/D122627.
This change correctly handles missing argument lists, and changes the
subject for the `numthreads` attribute to be global functions.
I did not handle applying the attribute to template functions because
that currently fails parsing in a way that is consisetent with the
current DXC codebase (Microsoft attributes are not supported on
templates).
A future improvement to the diagnostic maybe warranted.
This patch changes the requirement for getting the declaration of the
assertion handler from including <__assert> to including any public
C++ header of the library. Note that C compatibility headers are
excluded because we don't implement all the C headers ourselves --
some of them are taken straight from the C library, like assert.h.
It also adds a generated test to check it. Furthermore, this new
generated test is designed in a way that will make it possible to
replace almost all the existing test-generation scripts with this
system in upcoming patches.
Differential Revision: https://reviews.llvm.org/D122506
If the `ExternalFS` has already remapped a path then the
`RedirectingFileSystem` should not change it to the originally provided
path. This fixes the original path always being used if multiple VFS
overlays were provided and the path wasn't found in the highest (ie.
first in the chain).
This also renames `IsVFSMapped` to `ExposesExternalVFSPath` and only
sets it if `UseExternalName` is true. This flag then represents that the
`Status` has an external path that's different from its virtual path.
Right now the contained path is still the external path, but further PRs
will change this to *always* be the virtual path. Clients that need the
external can then request it specifically.
Note that even though `ExposesExternalVFSPath` isn't set for all
VFS-mapped paths, `IsVFSMapped` was only being used by a hack in
`FileManager` that was specific to module searching. In that case
`UseExternalNames` is always `true` and so that hack still applies.
Resolves rdar://90578880 and llvm-project#53306.
Differential Revision: https://reviews.llvm.org/D122549
Previously, these isel optimizations were disabled if the AND could
be selected as a ANDI instruction. This patch disables the optimizations
only if the immediate is valid for C.ANDI. If we can't use C.ANDI,
we might be able to compress the shift instructions instead.
I'm not checking the C extension since we have relatively poor test
coverage of the C extension. Without C extension the code size
should be equal. My only concern would be if the shift+andi had
better latency/throughput on a particular CPU.
I did have to add a peephole to match SRLIW if the input is zexti32
to prevent a regression in rv64zbp.ll.
Reviewed By: luismarques
Differential Revision: https://reviews.llvm.org/D122701
The splat_vector will be legalized to build_vector eventually
anyway. This patch makes it take fewer steps.
Unfortunately, this results in some codegen changes. It looks
like it comes down to how the nodes were ordered in the topological
sort for isel. Because the build_vector is created earlier we end up
with a different ordering of nodes.
Reviewed By: frasercrmck
Differential Revision: https://reviews.llvm.org/D122185
Avoids merge errors when opaque pointers are loaded into different types.
Reviewed by: jcranmer-intel, hiraditya
Differential Revision: https://reviews.llvm.org/D122521
D122053 set the ExtendType for ConstantSDNodes in getCopyToRegs to
ZERO_EXTEND to match assumptions in ComputePHILiveOutRegInfo. PHIs
are probably not the only way ConstantSDNodeNodes can get to
getCopyToRegs.
This patch adds an ExtendType parameter to CopyValueToVirtualRegister and
has HandlePHINodesInSuccessorBlocks pass ISD::ZERO_EXTEND for ConstantInts.
This way we only affect ConstantSDNodes for PHIs.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D122171
To achieve this we hook into the preprocessor during the
ExtractAPIAction and record definitions for macros that don't get
undefined during preprocessing.
Make the API records a property of the action instead of the ASTVisitor
so that it can be accessed outside the AST visitation and push back
serialization to the end of the frontend action.
This will allow accessing and modifying the API records outside of the
ASTVisitor, which is a prerequisite for supporting macros.
Now that all dependencies on creating the latch block up-front have been
removed, there is no need to create it early.
Depends on D121618.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D121619
This patch mostly follows up on D121292 which introduced the vp.fcmp
intrinsic.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D122729
Extend D120185 to also log the node being matched on in case of a crash.
This can help if a matcher is causing a crash or there are not enough interesting nodes bound.
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D122529
D86142 introduced --fortran-common and defaulted it to true (matching GNU ld
but deviates from gold/macOS ld64). The default state was motivated by transparently
supporting some FORTRAN 77 programs (Fortran 90 deprecated common blocks).
Now I think it again. I believe we made a mistake to change the default:
* this is a weird and legacy rule, though the breakage is very small
* --fortran-common introduced complexity to parallel symbol resolution and will slow down it
* --fortran-common more likely causes issues when users mix COMMON and
STB_GLOBAL definitions (see https://github.com/llvm/llvm-project/issues/48570 and
https://maskray.me/blog/2022-02-06-all-about-common-symbols).
I have seen several issues in our internal projects and Android.
On the other hand, --no-fortran-common is safer since
COMMON/STB_GLOBAL have the same semantics related to archive member extraction.
Therefore I think we should switch back, not punishing the common uage.
A platform wanting --fortran-common can implement ld.lld as a shell script
wrapper around `lld -flavor gnu --fortran-common "$@"`.
Reviewed By: ikudrin, sfertile
Differential Revision: https://reviews.llvm.org/D122450
Implement a demangleable strong ownership symbol mangling.
* The original module symbol mangling scheme turned out to be
undemangleable.
* The hoped-for C++17 compatibility of weak ownership turns out to be
fragile
* C++20 now has better ways of controlling C++17 compatibility
The issue is captured on the ABI list at:
https://github.com/itanium-cxx-abi/cxx-abi/issues/134
GCC implements this new mangling.
The old mangling is unceremoniously dropped. No backwards
compatibility, no deprectated old-mangling flag. It was always
labelled experimental. (Old and new manglings cannot be confused.)
Reviewed By: dblaikie
Differential Revision: https://reviews.llvm.org/D122256
We only want to do the upgrade from named to anonymous struct
return if the intrinsic is declared to return a struct, but not
if it has an overloaded return type that just happens to be a
struct. In that case the struct type will be mangled into the
intrinsic name and there is no problem.
This should address the problem reported in
https://reviews.llvm.org/D122471#3416598.
This allows us to detect whether we're being compiled with LLVM's libunwind
more easily, without CMake having to set explicit variables.
As discussed in https://llvm.org/D119538.
Differential Revision: https://reviews.llvm.org/D121015
This is an extension of D70965 to avoid creating a mathlib
call where it did not exist in the original source. Also see
D70852 for discussion about an alternative proposal that was
abandoned.
In the motivating bug report:
https://github.com/llvm/llvm-project/issues/54554
...we also have a more general issue about handling "no-builtin" options.
Differential Revision: https://reviews.llvm.org/D122610
This recommits dddf4ce03, which was reverted because of a couple of test
failures on macos. The reason behind the failures was that the patch
inadvertenly changed the value returned by the host platform from
"macosx" to "darwin". The new version fixes that.
Original commit message was:
The decision which categories are relevant for a particular test run
happen very early in the test setup process. They use the SBPlatform
object to determine which categories should be skipped. The platform
object created for this purpose transcends individual test runs.
This setup is not compatible with the direction discussed in
<https://discourse.llvm.org/t/multiple-platforms-with-the-same-name/59594>
-- when platform objects are tied to a specific (SB)Debugger, they need
to be created alongside it, which currently happens in the test setUp
method.
This patch is the first step in that direction -- it rewrites the
category skipping logic to avoid depending on a global SBPlatform
object. Fortunately, the skipping logic is fairly simple (and I believe
it outght to stay that way) and mainly consists of comparing the
platform name against some hardcoded lists. This patch bases this
comparison on the platform name instead of the os part of the triple (as
reported by the platform).
Differential Revision: https://reviews.llvm.org/D121605
Statically checking for overflow with
if constexpr (sizeof(std::size_t) <= sizeof(std::int64_t)) {
return static_cast<std::int64_t>(length);
}
Doesn't work if `sizeof(std::size_t) == sizeof(std::int64_t)` because std::size_t
is unsigned.
if `length == std::numeric_limits<size_t>` casting it to `int64_t` is going to overflow.
This code would be much simpler if returning a `uint64_t` instead of a signed
value...
Differential Revision: https://reviews.llvm.org/D122705
This patch adds the first support for vector-predicated comparison
intrinsics, starting with vp.fcmp. It uses metadata to encode its
condition code, like the llvm.experimental.constrained.fcmp intrinsic.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D121292
When shifting by a byte-multiple:
bswap (shl X, C) --> lshr (bswap X), C
bswap (lshr X, C) --> shl (bswap X), C
This is the backend version of D122010 and an alternative
suggested in D120648.
There's an extra check to make sure the shift amount is
valid that was not in the rough draft.
I'm not sure if there is a larger motivating case for RISCV (bug report?),
but the ARM diffs show a benefit from having a late version of the
transform (because we do not combine the loads in IR).
Differential Revision: https://reviews.llvm.org/D122655
This patch fixes a (seemingly very rare) crash during vector constant
folding introduced in D113300.
Normally, during legalization, if we create an illegally-typed node during
a failed attempt at constant folding it's cleaned up before being
visited, due to it having no uses.
If, however, an illegally-typed node is created during one round of
legalization and isn't cleaned up, it's possible for a second round of
legalization to create new illegally-typed nodes which add extra uses to
the old illegal nodes. This means that we can end up visiting the old
nodes before they're known to be dead, at which point we crash.
I'm not happy about this fix. Creating illegal types at all seems like a
bad idea, but we all-too-often rely on illegal constants being
successfully folded and being fixed up afterwards. However, we can't
rely on constant folding actually happening, and we don't have a
foolproof way of peering into the future.
Perhaps the correct fix is to revisit the node-iteration order during
legalization, ensuring we visit all uses of nodes before the nodes
themselves. Or alternatively we could try and clean up dead nodes
immediately after failing constant folding.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D122382
Florian Hahn