Enable more CLIF tests on AArch64

The tests for the SIMD floating-point maximum and minimum operations
require particular care because the handling of the NaN values is
non-deterministic and may vary between platforms. There is no way to
match several NaN values in a test, so the solution is to extract the
non-deterministic test cases into a separate file that is subsequently
replicated for every backend under test, with adjustments made to the
expected results.

Copyright (c) 2021, Arm Limited.

cranelift/filetests/filetests/runtests/simd-arithmetic-nondeterministic-x86_64.clif

@@ -0,0 +1,28 @@
+; Test the non-deterministic aspects of the SIMD arithmetic operations.
+; If you change this file, you should most likely update
+; simd-arithmetic-nondeterministic*.clif as well.
+test run
+set enable_simd
+target x86_64 machinst skylake
+
+function %fmax_f64x2(f64x2, f64x2) -> f64x2 {
+block0(v0: f64x2, v1: f64x2):
+    v2 = fmax v0, v1
+    return v2
+}
+
+; note below how NaNs are quieted but (unlike fmin), retain their sign: this discrepancy is allowed by non-determinism
+; in the spec, see https://webassembly.github.io/spec/core/bikeshed/index.html#nan-propagation%E2%91%A0.
+; run: %fmax_f64x2([NaN:0x42 0.0], [0x1.0 0.0]) == [NaN 0.0]
+
+function %fmin_f64x2(f64x2, f64x2) -> f64x2 {
+block0(v0: f64x2, v1: f64x2):
+    v2 = fmin v0, v1
+    return v2
+}
+
+; note below how NaNs are quieted and negative: this is due to non-determinism in the spec for NaNs, see
+; https://webassembly.github.io/spec/core/bikeshed/index.html#nan-propagation%E2%91%A0.
+; run: %fmin_f64x2([-NaN 0x100.0], [0.0 NaN]) == [-NaN -NaN]
+; run: %fmin_f64x2([NaN 0.0], [0.0 0.0]) == [-NaN 0.0]
+; run: %fmin_f64x2([NaN:0x42 0.0], [0x1.0 0.0]) == [-NaN 0.0]