Wrong floating-point emulation on AArch64 with FPCR set to zero

[Koutheir Attouchi]

On AArch64, with FPCR set to Zero (i.e., FPU set to IEEE-754 compliant mode), floating-point emulation does not produce the same results as real hardware (e.g., Raspberry Pi 3 with AArch64 Linux).

I attached a sample that reproduces the issue. It divides `x` by `y` and puts the result in `r`. The expected result of the operation is `q`.

Output on real hardware:
=========================================================
fpcr = 0x07000000.
x = 0x03250f416dcdc6d0. y = 0x00029f4e5837c977. r = 0x7ff0000000000000. q = 0x43300fde9cbcf023.
fpcr = 0x00000000.
x = 0x03250f416dcdc6d0. y = 0x00029f4e5837c977. r = 0x43300fde9cbcf023. q = 0x43300fde9cbcf023.
=========================================================

Notice that after setting FPCR to zero, `r` equals `q`.

Output on qemu 3.0.0 (Linux user-mode emulation):
=========================================================
fpcr = 0x07000000.
x = 0x03250f416dcdc6d0. y = 0x00029f4e5837c977. r = 0x7ff0000000000000. q = 0x43300fde9cbcf023.
fpcr = 0x00000000.
x = 0x03250f416dcdc6d0. y = 0x00029f4e5837c977. r = 0x43300fde9cbcf024. q = 0x43300fde9cbcf023.
=========================================================

Notice that after setting FPCR to zero, `r` is not equal to `q`.

Also notice that, using another proprietary operating system, the same issue arises between a real board and QEMU. This might be an issue in emulation of the AArch64 instruction "fdiv".

Build command line: aarch64-linux-gnu-gcc -static -O0 -o sample1 sample1.c

[Emilio G. Cota (cota)]

Thanks for your report. This is a known regression on our implementation of f64_div, introduced by cf07323d49 ("fpu/softfloat: re-factor div", 2018-02-21).

We are working on improving FP tests to limit regressions, e.g. see this thread, where the bug you report is first mentioned: https://lists.gnu.org/archive/html/qemu-devel/2018-09/msg01703.html

[Koutheir Attouchi (koutheir)]

Thanks for the update. Is there a fix/patch for the issue?

[Emilio G. Cota (cota)]

Not yet. There should be a fix before 3.1 is released.

Both 2.12 and 3.0 have this bug, so you might want to consider using 2.11 until the bug gets fixed.

[Peter Maydell (pmaydell)]

On 18 September 2018 at 19:18, Emilio G. Cota
<email address hidden> wrote:
> Not yet. There should be a fix before 3.1 is released.
>
> Both 2.12 and 3.0 have this bug, so you might want to consider using
> 2.11 until the bug gets fixed.

On the other hand 2.11 has a different set of slightly-inaccurate-fp-result
bugs that are fixed in 3.0, so you get to pick your choice of bug there.

thanks
-- PMM

[Koutheir Attouchi (koutheir)]

Neither will be sufficient in my use case. IEEE-754 conformance is essential. Thank you for the hints.

[Koutheir Attouchi (koutheir)]

Thanks a lot :-)

[Thomas Huth (th-huth)]

https://git.qemu.org/?p=qemu.git;a=commitdiff;h=5dfbc9e4903c0121140f2