Update Ubuntu (and therefore Qemu) Version

[Alexhuszagh]

Currently, on images provided by cross v0.2.1, we get the following after loading into the image:

Cross Information:

$ cross --version
cross 0.2.1
cargo 1.51.0 (43b129a20 2021-03-16)

Image Information

$ $ docker images
REPOSITORY           TAG                                     IMAGE ID       CREATED         SIZE
rustembedded/cross   powerpc-unknown-linux-gnu-0.2.1         e18a73ede16f   10 months ago   752MB
$ docker run -it e18a73ede16f /bin/bash
# lsb_release -a
No LSB modules are available.
Distributor ID: Ubuntu
Description:    Ubuntu 16.04.6 LTS
Release:        16.04
Codename:       xenial
# qemu-img --version
qemu-img version 3.0.1
Copyright (c) 2003-2017 Fabrice Bellard and the QEMU Project developers

The latest LTS version for Ubuntu is 20.04, which has newer versions of important dependencies, such as Qemu.

$ docker run -it ubuntu:20.04 /bin/bash
# lsb_release -a
No LSB modules are available.
Distributor ID: Ubuntu
Description:    Ubuntu 20.04.2 LTS
Release:        20.04
Codename:       focal
# qemu-img --version
qemu-img version 4.2.1 (Debian 1:4.2-3ubuntu6.16)
Copyright (c) 2003-2019 Fabrice Bellard and the QEMU Project developers

Issue

This actually matters, as numerous bugs, such as for float-point emulation, have been fixed in newer Qemu versions:

• Inaccurate Fmod on i386

In addition, here's a sample code snippet that fails on Qemu 3.0.1 but works on Qemu 4.2.1:

#include <inttypes.h>
#include <stdio.h>
#include <stdint.h>
#include <string.h>

double dn1() {
    double num = 1303.0;
    double den = 1e-20;
    return num * den;
}

double dn2() {
    double num = 1303.0;
    double den = 1e20;
    return num / den;
}


float fn1() {
    float num = 1303.0;
    float den = 1e-20;
    return num * den;
}

float fn2() {
    float num = 1303.0;
    float den = 1e20;
    return num / den;
}

int main() {
  float f = fn1();
  uint32_t bf;
  memcpy(&bf, &f, 8);
  printf("%u\n", bf);

  f = fn2();
  memcpy(&bf, &f, 8);
  printf("%u\n", bf);

  double d = dn1();
  uint64_t bd;
  memcpy(&bd, &d, 8);
  printf("%" PRIu64 "\n", bd);

  d = dn2();
  memcpy(&bd, &d, 8);
  printf("%" PRIu64 "\n", bd);

  return 0;
}

If we run the above using Qemu 3.0.1, we get:

# qemu-ppc --version
qemu-ppc version 3.0.1
Copyright (c) 2003-2017 Fabrice Bellard and the QEMU Project developers
# qemu-ppc ./main-ppc
594566255
594566255
4354430593920867240
4354430593920867241

If we run the above using Qemu 4.2.1, we get:

# qemu-ppc --version
qemu-ppc version 4.2.1 (qemu-4.2.1-1.fc32)
Copyright (c) 2003-2019 Fabrice Bellard and the QEMU Project developers
# qemu-ppc ./main-ppc
594566255
594566255
4354430593920867240
4354430593920867240

In short, we have a rounding error of 1 ULP due to cross using outdated Ubuntu (and therefore Qemu) versions. Ideally, we'd use version 6.0.1 or v5.0.0 (the latter requiring the following PPA, the former requiring compiling from source).

[senden9]

Relates to #517

[Alexhuszagh]

The system packages for Focal (20.04) are still at 4.2.X, and v6.1.0 of Qemu has recently been released, and there's been numerous bug fixes in 5.X and 6.X. Since we already build from source, it should be trivial to update the local versions and enure all targets build.

[VictorKoenders]

I'm having an issue while running powerpc64le-unknown-linux-gnu where all floats are 0.0 and are not equal to each other

Is this related or should I open a separate issue?

[Alexhuszagh]

@VictorKoenders How recent is the version you are running? It should be fixed in Qemu versions, although I'm not sure a cross-rs release has been published since then, as shown in #586.

[VictorKoenders]

I think my issue is closer related to #644

[Emilgardis]

we're using 5.1.0 on upstream cross right now, try using that instead while we get ready for a release.

cargo install cross --git https://github.com/cross-rs/cross

or point the image you're using to the newest (unreleased) image

[target.{{target}}]
image = "ghcr.io/cross-rs/{{target}}:edge"

[Alexhuszagh]

@VictorKoenders They're related (I know from the Qemu bug tracker): the floating point implementation was somewhat bugged in older versions of Qemu, and pinning it to v3.0.1 (which I did) fixed some issues but causes other ones that existed in 4.x. In Qemu versions 5.x+, (I believe) all of these are fixed. So unpinning the older version of Qemu and using a suitable, modern version should fix these issues (which is done in upstream).

You can test upstream or compile and use a system installation of a modern Qemu version to test. (Right now I'm getting relocation errors on my home PC cross-compiling and don't have time to test).

[Alexhuszagh]

@Emilgardis Also this PR is ready but is waiting on the merge of #593, since 2 of the images fail with interactive installed due to tzdata. Any review of that and this PR should be fit to be merged.