Re: git: 4c0dfd5959bd - main - arm: fail early on gcc builds

> On 17 May 2024, at 17:27, Brooks Davis <brooks@freebsd.org> wrote:
> 
> On Fri, May 17, 2024 at 09:14:46AM -0700, John Baldwin wrote:
>> On 5/17/24 9:01 AM, Brooks Davis wrote:
>>> The branch main has been updated by brooks:
>>> 
>>> URL: https://cgit.FreeBSD.org/src/commit/?id=4c0dfd5959bd499754e3fc9a7ce37c1d5ece31f7
>>> 
>>> commit 4c0dfd5959bd499754e3fc9a7ce37c1d5ece31f7
>>> Author:     Brooks Davis <brooks@FreeBSD.org>
>>> AuthorDate: 2024-05-17 15:47:22 +0000
>>> Commit:     Brooks Davis <brooks@FreeBSD.org>
>>> CommitDate: 2024-05-17 16:01:19 +0000
>>> 
>>>     arm: fail early on gcc builds
>>>     Since at least 2022 (see https://reviews.freebsd.org/D36754), it has
>>>     not been possible to build armv6/armv7 with gcc due to atomics macros
>>>     gcc doesn't like.  Prevent developers doing due diligance from wasting
>>>     time and CPU cycles on this combination as it just fails to build in
>>>     libc.
>>>     Reviewed by:    imp, andrew
>>>     Differential Revision:  https://reviews.freebsd.org/D45193
>> 
>> As I said in the review, this is a weird change.  Only amd64 and i386 build
>> with GCC.  You should blacklist all the others if you are going to blacklist
>> any, not just 32-bit arm.  I hadn't blacklisted any to make it possible for
>> other developers to work on it.  Also, the review in question is probably not
>> the right fix, it's probably that we need to pass different CFLAGS to GCC for
>> 32-bit arm.

I found too many issues with gcc 13 to get a world or kernel built. I have fixes for some of them, but it looked like the effort to fix the remaining issues wasn’t worth it.

> 
> I think I misunderstood your point in the review.  I've wasted days
> trying to do the right thing to let me reapply 4510f2ca9170 (Make
> WITHOUT_UNDEFINED_VERSION the default) only to find that it's secret
> knowledge that gcc builds are broken except on x86.
> 
> IMO there's less than zero value in universe targets that don't work
> unless they are clearly fenced.
> 
> I propose disabling all the others and adding a flag to ignore this
> guard.
> 
> -- Brooks
> 
> P.S. I've got aarch64 building except for GOOGLETEST and BOOT which both
> fail to link due to relocation complaints.

I have fixes in [1] to build world and kernel for arm64 with GCC. I’ve pushed some of the simple fixes & can put the remaining in review next week.

Andrew

[1] https://github.com/zxombie/freebsd/tree/gcc-arm64