runtime: TestFloat64 test fails under arm64 qemu

[wizeman]

What version of Go are you using (go version)?

1.11.4

Does this issue reproduce with the latest release?

Yes.

What operating system and processor architecture are you using (go env)?

I'm running qemu in system emulation mode. Host is (NixOS) Linux x86-64, guest is (NixOS) Linux arm64.

Relevant qemu command line arguments:

# qemu-system-aarch64 -machine virt,highmem=off -cpu cortex-a57 -smp cpus=4 -m 2048 -bios u-boot.bin (...)

What did you do?

Tried to build the go compiler as part of the standard Nix/NixOS build system, which includes running the tests.

What did you expect to see?

I expected all the tests to pass.

What did you see instead?

I ran into this failure:

--- FAIL: TestFloat64 (1.47s)
    softfloat64_test.go:131: 0.0021606686026376387 / -1.136016955471848 = sw -0.001901968621357591, hw -0.0019019686213575912
FAIL
FAIL    runtime 902.126s

Seems to be some kind of rounding issue.

All other tests pass (although admittedly, NixOS disables some tests by default due to them being unreliable or requiring sandboxed features).

[cherrymui]

Can we be sure that the QEMU implementation of the hardware floating point is correct and accurate?

[wizeman]

You may be right that this could be a QEMU problem.
I found this bug report related to QEMU floating point division inaccuracy in IEEE-754 compliant mode: https://bugs.launchpad.net/qemu/+bug/1793119

I'm running QEMU 3.0.0 which doesn't seem to have the following commit, which is mentioned in the bug report and is supposed to fix it: https://git.qemu.org/?p=qemu.git;a=commitdiff;h=5dfbc9e4903c0121140f2

I will try applying it and see if it solves the issue.

[wizeman]

I can confirm that the above QEMU patch seems to have solved this issue. Thanks!