https://lab.llvm.org/buildbot/#/builders/174/builds/13432

XCOFFObjectFile.cpp:805:12: error: reinterpret_cast from 'unsigned long' to 'uintptr_t' (aka 'unsigned int') is not allowed
    return reinterpret_cast<uintptr_t>(0ul);

The test requires the AArch64 backend, so move it to the right subdir.

Includes a test for the miscompile in #57712.

Use opaqueptr for test case
llvm/test/Transforms/SimplifyCFG/preserve-llvm-loop-metadata.ll.

- Adjust variable number accordingly since bitcast between different pointer
  types are not necessary.

Differential Revision: https://reviews.llvm.org/D134159

https://lab.llvm.org/buildbot/#/builders/216/builds/9977

XCOFFOtFile.cpp: error C3487: 'unsigned long': all return expressions must deduce to the same type: previously it was 'uintptr_t'

Write a semantics test for the atomic intrinsic subroutine,
atomic_and.

Reviewed By: rouson

Differential Revision: https://reviews.llvm.org/D133727

This was achieved with the 'cost-tables vs llvm-mca' script D103695

After BOLT's merge to LLVM, there are two (almost identical) versions of the
code layout algorithm. The diff unifies the implementations by keeping the one
in LLVM.

There are mild changes in the resulting block orders. I tested the changes
extensively both on the clang binary and on prod services. Didn't see stat sig
differences on average.

Reviewed By: Amir

Differential Revision: https://reviews.llvm.org/D129895

Summary:

according nm in AIX OS , https://www.ibm.com/docs/en/aix/7.2?topic=n-nm-command

In AIX OS, The default is to process 32-bit object files (ignore 64-bit objects). The mode can also be set with the OBJECT_MODE environment variable. For example, OBJECT_MODE=64 causes nm to process any 64-bit objects and ignore 32-bit objects. The -X flag overrides the OBJECT_MODE variable.

In non AIX OS. The default is to process all support object files. and not support the OBJECT_MODE environment variable.

Reviewers:  James Henderson

Differential Revision: https://reviews.llvm.org/D132494

In https://llvm.org/D56913, we added an emulation for the __atomic_always_lock_free
compiler builtin when compiling in Freestanding mode. However, the emulation
did (and could not) give exactly the same answer as the compiler builtin,
which led to a potential ABI break for e.g. enum classes.

After speaking to the original author of D56913, we agree that the correct
behavior is to instead always use the compiler builtin, since that provides
a more accurate answer, and __atomic_always_lock_free is a purely front-end
builtin which doesn't require any runtime support. Furthermore, it is
available regardless of the Standard mode (see https://godbolt.org/z/cazf3ssYY).

However, this patch does constitute an ABI break. As shown by https://godbolt.org/z/1eoex6zdK:
- In LLVM <= 11.0.1, an atomic<enum class with 1 byte> would not contain a lock byte.
- In LLVM >= 12.0.0, an atomic<enum class with 1 byte> would contain a lock byte.

This patch breaks the ABI again to bring it back to 1 byte, which seems
like the correct thing to do.

Fixes #57440

Differential Revision: https://reviews.llvm.org/D133377

This feature relies on Relations in the index being complete.
An out-of-tree index implementation is missing some override relations, so
such renames end up breaking the code.
We plan to fix it, but this flag is a cheap band-aid for now.

Differential Revision: https://reviews.llvm.org/D133440

This was achieved with the 'cost-tables vs llvm-mca' script D103695

Summary:

llvm-readobj support a new option --exception-section for xcoff object file.

https://www.ibm.com/docs/en/aix/7.2?topic=formats-xcoff-object-file-format#XCOFF__iua3i23ajbau

Reviewers:  James Henderson,Paul Scoropan

Differential Revision: https://reviews.llvm.org/D133030

When we aim a hint at some expanded tokens, we're only willing to attach it
to spelled tokens that exactly corresponde.

e.g.
int zoom(int x, int y, int z);
int dummy = zoom(NUMBERS);

Here we want to place a hint "x:" on the expanded "1", but we shouldn't
be willing to place it on NUMBERS, because it doesn't *exactly*
correspond (it has more tokens).

Fortunately we don't even have to implement this algorithm from scratch,
TokenBuffer has it.

Fixes https://github.com/clangd/clangd/issues/1289
Fixes https://github.com/clangd/clangd/issues/1118
Fixes https://github.com/clangd/clangd/issues/1018

Differential Revision: https://reviews.llvm.org/D133982

This revision adds a new op `map_nested_foreach_thread_to_gpu_threads` to transform dialect. The op searches `scf.foreach_threads` inside the `gpu_launch` and distributes them with `gpu.thread_id` attribute.

Loop mapping is explicit and given by the `map_nested_foreach_thread_to_gpu_threads` op. Mapping is done one-to-one, therefore the loops dissappear.

The dynamic trip count or trip count that are larger than thread size are not supported for the time being. However, we can indeed support them by generating a loop inside with cyclic scheduling.

For the time being, trip counts that are dynamic or bigger than thread sizes are not supported. However, in the future the compiler can indeed generate a loop with static cyclic scheduling to support these cases.

Current mechanism allows `scf.foreach_threads` to be siblings or nested. There cannot be interleaving code between the loops when they are nested.

Reviewed By: nicolasvasilache

Differential Revision: https://reviews.llvm.org/D133950

Implement exp10f function correctly rounded to all rounding modes.

Algorithm: perform range reduction to reduce
```
  10^x = 2^(hi + mid) * 10^lo
```
where:
```
  hi is an integer,
  0 <= mid * 2^5 < 2^5
  -log10(2) / 2^6 <= lo <= log10(2) / 2^6
```
Then `2^mid` is stored in a table of 32 entries and the product `2^hi * 2^mid` is
performed by adding `hi` into the exponent field of `2^mid`.
`10^lo` is then approximated by a degree-5 minimax polynomials generated by Sollya with:
```
  > P = fpminimax((10^x - 1)/x, 4, [|D...|], [-log10(2)/64. log10(2)/64]);
```
Performance benchmark using perf tool from the CORE-MATH project on Ryzen 1700:
```
$ CORE_MATH_PERF_MODE="rdtsc" ./perf.sh exp10f
GNU libc version: 2.35
GNU libc release: stable
CORE-MATH reciprocal throughput   : 10.215
System LIBC reciprocal throughput : 7.944

LIBC reciprocal throughput        : 38.538
LIBC reciprocal throughput        : 12.175   (with `-msse4.2` flag)
LIBC reciprocal throughput        : 9.862    (with `-mfma` flag)

$ CORE_MATH_PERF_MODE="rdtsc" ./perf.sh exp10f --latency
GNU libc version: 2.35
GNU libc release: stable
CORE-MATH latency   : 40.744
System LIBC latency : 37.546

BEFORE
LIBC latency        : 48.989
LIBC latency        : 44.486   (with `-msse4.2` flag)
LIBC latency        : 40.221   (with `-mfma` flag)
```
This patch relies on https://reviews.llvm.org/D134002

Reviewed By: orex, zimmermann6

Differential Revision: https://reviews.llvm.org/D134104

This should fix the expensive checks build. Ideally we would not
have invalid loops in LoopDispositions.

Without LZCNT/BMI, the *_ZERO_UNDEF costs are cheaper as they can avoid the zero handling.

This reverts commit e5581df6.

This causes major compile-time regressions, about 2-3% end-to-end
on CTMark.

Optimize the core part of `tanhf` implementation that is to compute `e^x`
similar to https://reviews.llvm.org/D133870.  Factor the constants and
polynomial approximation out so that it can be used for `exp10f`

Performance benchmark using perf tool from the CORE-MATH project on Ryzen 1700:
```
$ CORE_MATH_PERF_MODE="rdtsc" ./perf.sh tanhf
GNU libc version: 2.35
GNU libc release: stable
CORE-MATH reciprocal throughput   : 13.377
System LIBC reciprocal throughput : 55.046

BEFORE:
LIBC reciprocal throughput        : 75.674
LIBC reciprocal throughput        : 33.242    (with `-msse4.2` flag)
LIBC reciprocal throughput        : 25.927    (with `-mfma` flag)

AFTER:
LIBC reciprocal throughput        : 26.359
LIBC reciprocal throughput        : 18.888    (with `-msse4.2` flag)
LIBC reciprocal throughput        : 14.243    (with `-mfma` flag)

$ CORE_MATH_PERF_MODE="rdtsc" ./perf.sh tanhf --latency
GNU libc version: 2.35
GNU libc release: stable
CORE-MATH latency   : 43.365
System LIBC latency : 123.499

BEFORE
LIBC latency        : 112.968
LIBC latency        : 104.908   (with `-msse4.2` flag)
LIBC latency        : 92.310    (with `-mfma` flag)

AFTER
LIBC latency        : 69.828
LIBC latency        : 63.874    (with `-msse4.2` flag)
LIBC latency        : 57.427    (with `-mfma` flag)
```

Reviewed By: orex, zimmermann6

Differential Revision: https://reviews.llvm.org/D134002

Only AVX512 has decent CTTZ/CTLZ vector ops, add tests to ensure we definitely vectorize these

This spotted a mistake with the original patch, so it puts the status
back to "partial" in the C status tracking page.

This amends 51038362.

Allows to determine known zero elements, which particularly helps simplification of DIV/REM by constant patterns

Given an opOperand uniquely determined by the operation `%op` and the operand number `num`,
the `transform.get_producer_of_operand %op[num]` returns the handle to the unique operation
that produced the SSA value used as opOperand.

The transform fails if the operand is a block argument.

Differential Revision: https://reviews.llvm.org/D134171

[mlir][Linalg] NFC - Cleanup internal transform APIs and produce better messages on failure to apply.

This is required because if there is a pure loop-invariant instruction, Loop Rotation
may decide to not clone it and just hoist it instead. If SCEV has previously cached
that it was loop-variant (not being smart enough to prove invariance), we may end
up with inconsistent cache state (which may later trigger false-negative assertion
failures checking that something was invariant).

This is a conservative fix that unconditionally drops the dispositions. We could
only drop it if the hoisting has actually happened, but it should take some time
understanding whether it's safe with all other things this function does.

Differential Revision: https://reviews.llvm.org/D134167
Reviewed By: fhahn

Alive2 doesn't support verification of optimizations that use inter-procedural analyses.
Right now, clang uses GlobalsAA by default and there's no way to disable it.
This leads to Alive2 producing false positives.
The added flag allows us to skip global analyses altogether.

Differential Revision: https://reviews.llvm.org/D134139

Add tests for the core static shuffle pattern match helpers

The batch-reduce GEMM kernel essentially multiplies a sequence of input tensor
blocks (which form a batch) and the partial multiplication results are reduced
into a single output tensor block.

See: https://ieeexplore.ieee.org/document/9139809 for more details.

Reviewed By: nicolasvasilache

Differential Revision: https://reviews.llvm.org/D134163

This bug was found by recent improvement in SCEV verifier. The code in LoopFuse
directly reassigns blocks to be a part of a different loop, which should automatically
invalidate all related cached loop dispositions.

Differential Revision: https://reviews.llvm.org/D134173
Reviewed By: nikic

It seems that it is sometimes broken. Initial motivation for this was
investigation of https://github.com/llvm/llvm-project/issues/56260, but
it also seems that we have found an unrelated bug in LoopFusion that leaves
broken caches.

Differential Revision: https://reviews.llvm.org/D134158
Reviewed By: nikic

Previously only using the UnsafeFPMath option, this now looks for the
fast moth flags on the instructions, using the same flag flags as other
backends.