Reimplements MisExpect diagnostics from D66324 to reconstruct its
original checking methodology only using MD_prof branch_weights
metadata.
New checks rely on 2 invariants:
1) For frontend instrumentation, MD_prof branch_weights will always be
populated before llvm.expect intrinsics are lowered.
2) for IR and sample profiling, llvm.expect intrinsics will always be
lowered before branch_weights are populated from the IR profiles.
These invariants allow the checking to assume how the existing branch
weights are populated depending on the profiling method used, and emit
the correct diagnostics. If these invariants are ever invalidated, the
MisExpect related checks would need to be updated, potentially by
re-introducing MD_misexpect metadata, and ensuring it always will be
transformed the same way as branch_weights in other optimization passes.
Frontend based profiling is now enabled without using LLVM Args, by
introducing a new CodeGen option, and checking if the -Wmisexpect flag
has been passed on the command line.
Reviewed By: tejohnson
Differential Revision: https://reviews.llvm.org/D115907
Latest OpenMP spec says parallel_data is NULL for initial/implicit-task-end.
We nevertheless need to cleanup the ParallelData here, as there is no other
callback for the end of the implicit parallel region. We can use the reference
stored in the TaskData.
Reviewed By: dreachem
Differential Revision: https://reviews.llvm.org/D114005
Whenever a v_cmp, s_and_saveexec instruction sequence shall be
transformed to an equivalent s_mov, v_cmpx sequence, it needs
to be detected if the v_cmp target register is used between
the two instructions as the v_cmp result gets omitted by
using the v_cmpx instruction, resulting in invalid code.
Reviewed By: foad
Differential Revision: https://reviews.llvm.org/D122797
Bubble up extract_slice above Linalg operation.
A sequence of operations
%0 = linalg.<op> ... arg0, arg1, ...
%1 = tensor.extract_slice %0 ...
can be replaced with
%0 = tensor.extract_slice %arg0
%1 = tensor.extract_slice %arg1
%2 = linalg.<op> ... %0, %1, ...
This results in the reduce computation of the linalg operation.
The implementation uses the tiling utility functions. One difference
from the tiling process is that we don't need to insert the checking
code for the out-of-bound accesses. The use of the slice itself
represents that the code writer is sure about the boundary condition.
To avoid adding the boundary condtion check code, `omitPartialTileCheck`
is introduced for the tiling utility functions.
Differential Revision: https://reviews.llvm.org/D122437
This adds diagnostics for conflicting attributes on the same
declarataion, conflicting attributes on a forward and final
declaration, and defines a more narrowly scoped HLSLEntry attribute
target.
Big shout out to @aaron.ballman for the great feedback and review on
this!
This reverts commit 4cb38bfe76.
Awkwardly enough, this builds Windows buildbots:
http://45.33.8.238/win/55402/step_9.txt
It is yet unclear why this is happening but I will need more time to
diagnose the issue.
Update the hardware CRC32 logic in scudo to support using `-mcrc32`
instead of `-msse4.2`. The CRC32 intrinsics use the former flag
in the newer compiler versions, e.g. in clang since 12fa608af4.
With these compilers, passing `-msse4.2` is insufficient to enable
the instructions and causes build failures when `-march` does not enable
CRC32:
/var/tmp/portage/sys-libs/compiler-rt-sanitizers-14.0.0/work/compiler-rt/lib/scudo/scudo_crc32.cpp:20:10: error: always_inline function '_mm_crc32_u32' requires target feature 'crc32', but would be inlined into function 'computeHardwareCRC32' that is compiled without support for 'crc32'
return CRC32_INTRINSIC(Crc, Data);
^
/var/tmp/portage/sys-libs/compiler-rt-sanitizers-14.0.0/work/compiler-rt/lib/scudo/scudo_crc32.h:27:27: note: expanded from macro 'CRC32_INTRINSIC'
# define CRC32_INTRINSIC FIRST_32_SECOND_64(_mm_crc32_u32, _mm_crc32_u64)
^
/var/tmp/portage/sys-libs/compiler-rt-sanitizers-14.0.0/work/compiler-rt/lib/scudo/../sanitizer_common/sanitizer_platform.h:132:36: note: expanded from macro 'FIRST_32_SECOND_64'
# define FIRST_32_SECOND_64(a, b) (a)
^
1 error generated.
For backwards compatibility, use `-mcrc32` when available and fall back
to `-msse4.2`. The `<smmintrin.h>` header remains in use as it still
works and is compatible with GCC, while clang's `<crc32intrin.h>`
is not.
Originally reported in https://bugs.gentoo.org/835870.
Differential Revision: https://reviews.llvm.org/D122789
The memory_scope enum is not available before OpenCL 2.0, so ensure
the sub_group_barrier overload with a memory_scope argument is
restricted to OpenCL 2.0 and above. This is already the case in
opencl-c.h.
Fixes the issue revealed by https://reviews.llvm.org/D120254
Reported-by: Harald van Dijk (hvdijk)
The alternative outputs of std::put_time and std::strftime are the
easiest to test with the Japanese locale. This is a preparation for the
tests of the chrono formatters.
Note since it takes a while before the Docker file changes propagate to
the build nodes the verification of the locale is done in a separate
patch.
Reviewed By: ldionne, #libc
Differential Revision: https://reviews.llvm.org/D122736
Reduced the details of the non-chrono formatting information. This has
been shipped and these details part of P0645 which is still documented.
Removing this information keeps the information up-to-date.
Adds the formatters required for the types chrono namespace.
Reviewed By: ldionne, #libc
Differential Revision: https://reviews.llvm.org/D122735
This change fixes an assert that occurs in the SMT layer when refuting a
finding that uses pointers of two different sizes. This was found in a
downstream build that supports two different pointer sizes, The CString
Checker was attempting to compute an overlap for the 'to' and 'from'
pointers, where the pointers were of different sizes.
In the downstream case where this was found, a specialized memcpy
routine patterned after memcpy_special is used. The analyzer core hits
on this builtin because it matches the 'memcpy' portion of that builtin.
This cannot be duplicated in the upstream test since there are no
specialized builtins that match that pattern, but the case does
reproduce in the accompanying LIT test case. The amdgcn target was used
for this reproducer. See the documentation for AMDGPU address spaces here
https://llvm.org/docs/AMDGPUUsage.html#address-spaces.
The assert seen is:
`*Solver->getSort(LHS) == *Solver->getSort(RHS) && "AST's must have the same sort!"'
Ack to steakhal for reviewing the fix, and creating the test case.
Reviewed By: steakhal
Differential Revision: https://reviews.llvm.org/D118050
VectorizerStart extension is module callback in old PM, but is function
callback in new PM. We lack a module extension point between end of
buildModuleSimplificationPipeline and the function optimization
(including vectorizer) pipeline. So this patch adds a new module
extension point before the function optimization pipeline.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D122296
This patch fixes a bug in LinearTransform::applyTo where it did not carry the
IdKind information, and instead treated every id as IdKind::Domain.
Reviewed By: arjunp
Differential Revision: https://reviews.llvm.org/D122823
Previously, if a `#pragma clang assume_nonnull begin` was at the
end of a premable with a `#pragma clang assume_nonnull end` at the
end of the main file, clang would diagnose an unterminated begin in
the preamble and an unbalanced end in the main file.
With this change, those errors no longer occur and the case above is
now properly handled. I've added a corresponding test to clangd,
which makes use of preambles, in order to verify this works as
expected.
Differential Revision: https://reviews.llvm.org/D122179
Summary:
To compute the size of a VALU/SALU instruction, we need to check whether an operand
could ever be literal. Previously isLiteralConstant was used, which missed cases
like global variables or external symbols. These misses lead to under-estimation of
the instruction size and branch offset, and thus incorrectly skip the necessary branch
relaxation when the branch offset is actually greater than what the branch bits can hold.
In this work, we use isLiteralConstantLike to check the operands. It maybe conservative,
but it is safe.
Reviewers: arsenm
Differential Revision: https://reviews.llvm.org/D122778
dd9173420f (Add clear_cache implementation for ppc64. Fix buffer to
meet ppc64 alignment., 2017-07-28), adds an implementation for
__builtin___clear_cache on powerpc64, which was promptly ammended to
also be used with big endian mode in f67036b62c (This ppc64 implementation
of clear_cache works for both big and little endian., 2017-08-02)
clang will use this implementation for it's builtin on FreeBSD and result
in an abort() in the cases where 32-bit generation was requested (ex in
macppc or when the big endian powerpc64 build was done with "-m32") and as
reported[1] recently with pcre2, but there is no reason why the same code
couldn't be used in those cases, so use instead the more generic identifier
for the PowerPC architecture.
While at it, update the comment to reflect that POWER8/9 have a 128 byte
wide cache line and so the code could instead use 64 byte windows instead
but that possible optimization has been punted for now.
[1] https://github.com/PhilipHazel/pcre2/issues/92
Reviewed By: jhibbits, #powerpc, MaskRay
Differential Revision: https://reviews.llvm.org/D122640
Instead of first creating a lambda for calculating the range,
then collecting the ranges for the operands, and then calling the
lambda on those ranges, we can first calculate the operand ranges
and then calculate the result directly in the switch.
Previously, when updating the outputs matrix, the local offset was not being considered.
Reviewed By: Groverkss
Differential Revision: https://reviews.llvm.org/D122812
This reverts the revert commit 2760cdc9c6.
This version pulls in the code to create the vector loop object in VPlan
from D121624.
This is needed because otherwise existing LoopInfo verification will
fail, as a loop block doesn't have in-loop successors now that we
do not replace the branch.
Now that we do not add new loops during skeleton construction, there's
also no need to verify LI there.
It seems to have been added back in 761e42fa3d for Clang to use it,
however it seems to have never been used for that purpose, so it is
probably fine to remove it.
Differential Revision: https://reviews.llvm.org/D122330
For some reason, we've been going without a MSAN CI job, even though
even run-buildbot defined a generic-msan job. This must have been an
oversight that went unnoticed. Thanks to @EricWF for the catch.
Differential Revision: https://reviews.llvm.org/D120851
This patch adds limited modeling of the `value_or` method. Specifically, when
used in a particular idiom in a comparison to implicitly check whether the
optional holds a value.
Differential Revision: https://reviews.llvm.org/D122231
This inverts a fold recently added to IR with:
3491f2f4b0
We can put -bidirectional on the Alive2 examples to show that
the reverse transforms work:
https://alive2.llvm.org/ce/z/8iVQwB
The motivation for the IR change was to improve matching to
'fabs' in IR (see https://github.com/llvm/llvm-project/issues/38828 ),
but it regressed x86 codegen for 'not-quite-fabs' patterns like
(X > -X) ? X : -X.
Ie, when there is no fast-math (nsz), the cmp+select is not a proper
fabs operation, but it does map nicely to the unusual NAN semantics
of MINSS/MAXSS.
I drafted this as a target-independent fold, but it doesn't appear to
help any other targets and seems to cause regressions for SystemZ at
least.
Differential Revision: https://reviews.llvm.org/D122726
Unfortunately this just shows that the test case for D47740 never
really tested what it was supposed to test.
Differential Revision: https://reviews.llvm.org/D122664
According to the current design, if a floating point operation is
represented by a constrained intrinsic somewhere in a function, all
floating point operations in the function must be represented by
constrained intrinsics. It imposes additional requirements to inlining
mechanism. If non-strictfp function is inlined into strictfp function,
all ordinary FP operations must be replaced with their constrained
counterparts.
Inlining strictfp function into non-strictfp is not implemented as it
would require replacement of all FP operations in the host function,
which now is undesirable due to expected performance loss.
Differential Revision: https://reviews.llvm.org/D69798
This fixes a TODO in constantArgPropagation() to make it feature complete.
However, I do find myself in agreement with the review comments in
https://reviews.llvm.org/D106426. I don't think we should pursue
specializing such recursive functions as the code size increase becomes
linear to 'max-iters'. Compiling the modified test just with -O3 (no
function specialization) generates the same code.
Differential Revision: https://reviews.llvm.org/D122755
The AMX combiner would store undef or zero to stack and invoke tileload
to load the data to tile register. To avoid the store/load, we can
materialzie undef or zero value to tilezero.
Differential Revision: https://reviews.llvm.org/D122714