[gardening][vm][fuzz] divergence clamp in SIMARM and SIMARM64

[dcharkes]

solate (/b/s/w/itxb1yll64/dart_fuzzUXAVAV) FP NO-FFI FLAT : JIT-SLOWPATH-DET-ReleaseSIMARM64 - JIT-OPTCOUNTER-DET-DebugSIMARM64C: !DIVERGENCE! 1.101:1203086137 (output)

out1 and out2 are different:

is different.
Expected: ... 0.565459, 0.000000,  ...
  Actual: ... 0.565459, -0.000000, ...
                        ^
 Differ at offset 6137

-- BEGIN REPRODUCE  --

DART SDK REVISION: 

dart runtime/tools/dartfuzz/dartfuzz.dart --fp --no-ffi --flat --seed 1203086137 fuzz.dart

-- RUN 1 --

out/ReleaseSIMARM64/dart --use-slow-path --deterministic --old_gen_heap_size=128 /b/s/w/itxb1yll64/dart_fuzzUXAVAV/fuzz.dart

-- RUN 2 --

out/DebugSIMARM64C/dart --optimization_counter_threshold=1 --deterministic --old_gen_heap_size=128 /b/s/w/itxb1yll64/dart_fuzzUXAVAV/fuzz.dart

-- END REPRODUCE  --

https://logs.chromium.org/logs/dart/buildbucket/cr-buildbucket/8738390725973939873/+/u/collect_shards/dartfuzz_-_generated_programs_shard_42/task_stdout_stderr:_dartfuzz_-_generated_programs_shard_42

Repros locally on 0329ebf.

Smaller repro: https://dart-review.googlesource.com/c/sdk/+/382600

 main() {
  final float32x4 = Float32x4(-0.0, 0.0, -0.0, 0.0);
  final lowerLimit = Float32x4(-1.0, -1.0, -0.0, -0.0);
  final result = float32x4.clamp(
    lowerLimit,
    Float32x4.zero(),
  );
  print(result);
  final resultAsInt = Int32x4.fromFloat32x4Bits(result);
  print(resultAsInt);
  Expect.equals(0x0, resultAsInt.x);
  Expect.equals(0x0, resultAsInt.y);
  Expect.equals(-0x80000000, resultAsInt.z);
  Expect.equals(-0x80000000, resultAsInt.w);
}

[dcharkes]

This fails on all arm64: https://dart-ci.firebaseapp.com/cl/382600/1

vm-aot-android-release-arm64c
vm-aot-linux-debug-simarm_x64
vm-aot-mac-release-arm64
vm-aot-win-debug-arm64
vm-checked-mac-release-arm64
vm-linux-release-simarm
vm-mac-debug-arm64

On Mac also reproduces without VM flags.

The implementation:

https://github.com/dart-lang/sdk/blob/main/runtime/lib/simd128.cc#L186-L219

  // ARM semantics are different from X86/X64 at an instruction level. Ensure
  // that we match the semantics of the architecture in the C version.

[dcharkes]

According to

• Inconsistent results for -Float32x4.zero().clamp  #40426

inconsistent results are expected on different architectures.

https://dart-review.googlesource.com/c/sdk/+/134573 prevents the fuzzer to compare float results with each other for various architectures. However, in this repro the two simulators for the same architecture produce different results.

I can also make a single simulator produce different results by adding prints before the code in question: https://dart-review.googlesource.com/c/sdk/+/382600/5.

Inconsistencies between the intrinsics and the simulator were already addressed: https://dart-review.googlesource.com/c/sdk/+/244320, and the IL in the repro has both as intrinsics, so something else is causing the divergence.

The runtime entry doesn't implement the arm64 (and arm) semantics faithfully.