Failures on "Add documentation, tests and better parameter names to Float32x4."

[alexmarkov]

There are new test failures on Add documentation, tests and better parameter names to Float32x4..

The tests

lib/typed_data/float32x4_test/0 RuntimeError (expected Pass)
lib/typed_data/float32x4_test/1 RuntimeError (expected Pass)

are failing on configurations

vm-aot-android-release-arm_x64

Log:

Unhandled exception:
Expect.equals(expected: <0>, actual: <-1>) fails.
#0      Expect._fail (package:expect/expect.dart:887)
#1      Expect._failNotEqual (package:expect/expect.dart:174)
#2      Expect.equals (package:expect/expect.dart:157)
#3      testComparison (file:///b/swarming/w/ir/tests/lib/typed_data/float32x4_test.dart:160)
#4      main (file:///b/swarming/w/ir/tests/lib/typed_data/float32x4_test.dart:651)
#5      _delayEntrypointInvocation.<anonymous closure> (dart:isolate-patch/isolate_patch.dart:315)
#6      _RawReceivePort._handleMessage (dart:isolate-patch/isolate_patch.dart:194)
ExitCode: 255
Time: 0:00:00.132921

--- Re-run this test:
python3 tools/test.py -n vm-aot-android-release-arm_x64 lib/typed_data/float32x4_test/0

https://dart-ci.appspot.com/log/vm-aot-android-release-arm_x64/vm-aot-android-release-arm_x64/4823/lib/typed_data/float32x4_test/0

/cc @lrhn

[alexmarkov]

The following comparison between min and 0.0 fails on 32-bit ARM:

/// Minimal positive 32-bit float value.
const min = 1.401298464324817e-45;

  var p = Float32x4(0.0, nan, min, nan);
  var q = Float32x4(-0.0, nan, 0.0, 3.14);

  cmp = p.lessThanOrEqual(q);

  Expect.equals(0, cmp.z);

The difference in behavior from other platforms is likely caused by flushing denormalized numbers to zero, so min is treated as 0.0 and min <= 0.0 is true/-1. We should probably document this behavior and fix the test to avoid denormalized numbers.

(I assume we don't want to stop using SIMD/NEON instructions on 32-bit ARM due to this non-compliance to IEEE 754.)

Related discussion: #24416.

@lrhn Could you take a look?

[lrhn]

Just like NaN, denormals is a gift that keeps giving.

See fx: WebAssembly/design#1429

I'll check if I can make the tests expect all the possible results, not just the "correct" ones.
(Not sure we've guaranteed any specific semantics for SIMD operations, and this is why we shouldn't.)

[a-siva]

@lrhn I presume you are going to adjust the documentation to reflect this fact.

[lrhn]

That too 😃