Re: [PATCH-for-5.0] roms/edk2-funcs.sh: Use available GCC for ARM/Aarch64 targets

[Laszlo Ersek]

On 12/05/19 17:50, Ard Biesheuvel wrote:
> On Thu, 5 Dec 2019 at 16:27, Philippe Mathieu-Daudé <address@hidden> wrote:
>>
>> On 12/5/19 5:13 PM, Laszlo Ersek wrote:
>>> Hi Phil,
>>>
>>> (+Ard)
>>>
>>> On 12/04/19 23:12, Philippe Mathieu-Daudé wrote:
>>>> Centos 7.7 only provides cross GCC 4.8.5, but the script forces
>>>> us to use GCC5. Since the same machinery is valid to check the
>>>> GCC version, remove the $emulation_target check.
>>>>
>>>>    $ cat /etc/redhat-release
>>>>    CentOS Linux release 7.7.1908 (Core)
>>>>
>>>>    $ aarch64-linux-gnu-gcc -v 2>&1 | tail -1
>>>>    gcc version 4.8.5 20150623 (Red Hat 4.8.5-16) (GCC)
>>>
>>> this patch is not correct, in my opinion. ARM / AARCH64 support in edk2
>>> requires GCC5 as a minimum. It was never tested with an earlier
>>> toolchain, to my understanding. Not on my part, anyway.
>>>
>>> To be more precise: when I tested cross-gcc toolchains earlier than
>>> that, the ArmVirtQemu builds always failed. Minimally, those toolchains
>>> didn't recognize some of the AARCH64 system registers.
>>>
>>> If CentOS 7.7 does not provide a suitable (>=GCC5) toolchain, then we
>>> can't build ArmVirtQemu binaries on CentOS 7.7, in my opinion.
>>>
>>> Personally, on my RHEL7 laptop, over time I've used the following
>>> toolchains, to satisfy the GCC5 requirement of ArmVirtQemu (which
>>> requirement I took as experimental evidence):
>>>
>>> - Initially (last quarter of 2014), I used binary distributions --
>>>    tarballs -- of cross-binutils and cross-gcc, from Linaro.
>>>
>>> - Later (last quarter of 2016), I rebuilt some SRPMs that were at the
>>>    time Fedora-only for RHEL7. Namely:
>>>
>>>    - cross-binutils-2.27-3.fc24
>>>      https://koji.fedoraproject.org/koji/buildinfo?buildID=801348
>>>
>>>    - gcc-6.1.1-2.fc24
>>>      https://koji.fedoraproject.org/koji/buildinfo?buildID=761767
>>>
>>> - Most recently, I've been using cross-binutils updated from EPEL7:
>>>
>>>    - cross-binutils-2.27-9.el7.1
>>>      https://koji.fedoraproject.org/koji/buildinfo?buildID=918474
>>>
>>> To my knowledge, there is still no suitable cross-compiler available on
>>> RHEL7, from any trustworthy RPM repository. So, to this day, I use
>>> gcc-6.1.1-2 for cross-building ArmVirtQemu, on my RHEL7 laptop.
>>>
>>> Again: I believe it does not matter if the gcc-4.8.5-based
>>> cross-compiler in CentOS 7 "happens" to work. That's a compiler that I
>>> have never tested with, or vetted for, upstream ArmVirtQemu.
>>>
>>> Now, I realize that in edk2, we have stuff like
>>>
>>>    GCC48_AARCH64_CC_FLAGS
>>>
>>> in "BaseTools/Conf/tools_def.template" -- coming from commit
>>> 7a9dbf2c94d1 ("BaseTools/Conf/tools_def.template: drop ARM/AARCH support
>>> from GCC46/GCC47", 2019-01-08). That doesn't change the fact that I've
>>> never built or tested ArmVirtQemu with such a compiler. And so this
>>> patch makes me quite uncomfortable.
>>>
>>> If that rules out CentOS 7 as a QEMU project build / CI platform for the
>>> bundled ArmVirtQemu binaries, then we need a more recent platform
>>> (perhaps CentOS 8, not sure).
>>
>> Unfortunately CentOS 8 is not available as a Docker image, which is a
>> convenient way to build EDK2 in a CI.
>>
>>> I think it's also educational to check the origin of the code that your
>>> patch proposes to remove. Most recently it was moved around from a
>>> different place, in QEMU commit 65a109ab4b1a ('roms: lift
>>> "edk2-funcs.sh" from "tests/uefi-test-tools/build.sh"', 2019-04-17).
>>>
>>> In that commit, for some reason I didn't keep the original code comments
>>> (perhaps it would have been too difficult or messy to preserve the
>>> comments sanely with the restructured / factored-out code). But, they
>>> went like this (originally from commit 77db55fc8155,
>>> "tests/uefi-test-tools: add build scripts", 2019-02-21):
>>>
>>> # Expose cross_prefix (which is possibly empty) to the edk2 tools. While at 
>>> it,
>>> # determine the suitable edk2 toolchain as well.
>>> # - For ARM and AARCH64, edk2 only offers the GCC5 toolchain tag, which 
>>> covers
>>> #   the gcc-5+ releases.
>>> # - For IA32 and X64, edk2 offers the GCC44 through GCC49 toolchain tags, in
>>> #   addition to GCC5. Unfortunately, the mapping between the toolchain tags 
>>> and
>>> #   the actual gcc releases isn't entirely trivial. Run "git-blame" on
>>> #   "OvmfPkg/build.sh" in edk2 for more information.
>>> # And, because the above is too simple, we have to assign cross_prefix to an
>>> # edk2 build variable that is specific to both the toolchain tag and the 
>>> target
>>> # architecture.
>>>
>>> So... unless Ard feels it is really totally safe to retro-actively rely
>>> on the gcc-4.8.5-based compiler in CentOS 7, I'd rather we picked a more
>>> recent build platform (OS) instead. For example, we build ArmVirtQemu on
>>> RHEL8 regularly, so that's a reality-based "plus" for CentOS 8.
>>>
>>>
>>> Independently of all of the above, the OVMF toolchain selection logic
>>> that this patch proposes to reuse with ArmVirtQemu, is *really*
>>> x86-specific. Please run "git blame" on "OvmfPkg/build.sh" in upstream
>>> edk2, to see where the various branches come from (as the comments in
>>> this shell script suggest as well). There had been mess like commit
>>> 656ac0c7d8ea ('Revert "OvmfPkg/build.sh: select the GCC49 toolchain
>>> settings for gcc-7.*"', 2017-08-25).
>>
>> Thanks for all the pointers, very educative indeed :)
>>
>> I'll see other setups I can use with GCC5+ available.
>>
>> I still have to figure if there are free tier CI with less limitations
>> than Travis/Shippable/GitLab, so we can keep the full EDK2 build output log.
>>
> 
> My CI job for ArmVirtQemu/EDK2 build tested GCC48 and GCC49 until very
> recently, and I never experienced any issues when running those
> images, although it's been much longer that I actually tried that. So
> I wouldn't recommend against it, and if we do identify any issues, we
> should either deprecate GCC48 (for ArmVirtQemu or for AArch64
> altogether) or fix them.
> 

OK, thank you, I'm fully satisfied with this addition. :)

Phil, in this case I think we can indeed replace the hard-coded "GCC5"
with a bit of dynamic detection. Two remarks:

- Please CC Ard on v2, so he can ACK. I'd like that. :)

- Again, we shouldn't blindly reapply the x86 (OVMF) quirk(s). I mean
mainly the special casing of "6.[0-2].*" to GCC49, which comes from
upstream edk2 commit 432f1d83f77a ("OvmfPkg/build.sh: Use GCC49
toolchains with GCC 6.[0-2]", 2016-12-06).

... or is that GCC bug target-independent in fact? I can't really tell;
the upstream GCC bug
<https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70955> is ISA-specific
(x86-64), and so are function calling conventions.

I'd suggest *not* applying the quirk for ArmVirtQemu, initially.

Thanks
Laszlo