Re: [Linaro-TCWG-CI] llvmorg-20-init-4622-g16900d3b98e6: FAIL: 1 regressions on aarch64

Takahiro Kawashima (Fujitsu) t-kawashima@fujitsu.com wrote:

I took a look. This is probably caused by a floating-point arithmetic error.

The Fortran source program is https://github.com/fujitsu/compiler-test-suite/blob/main/Fortran/0354/0354_0... . The compilation flags are `-O3 -ffast-math`.

Changing `1.0E-6` to `1.0E-3` in the first line resolves this failure. Adding `-fno-associative-math` also resolves this failure.

If this floating-point arithmetic error is intended to occur, you don't need to investigate it.

Thanks for investigating. Under `-ffast-math`, we're allowed to reassociate floating-point ops, and indeed we do, so I think the source code change is the right fix.

Warm regards, Ram