Re: [PATCH v2] aarch64: Add fallback if ARM_CC_FOR_TARGET not set (was: Re: [PATCH 5/5] gdb: native target invalid architecture detection)

[John Baldwin <jhb@FreeBSD.org>]

On 6/8/22 2:48 PM, Pedro Alves wrote:
> On 2022-06-08 20:01, John Baldwin wrote:
> 
>>> This commit adds a fallback.  If ARM_CC_FOR_TARGET is not set, and
>>> testing for Linux, try arm-linux-gnueabi-gcc,
>>> arm-none-linux-gnueabi-gcc, arm-linux-gnueabihf-gcc as 32-bit
>>> compilers, making sure that the produced executable runs on the target
>>> machine before claiming that the compiler produces useful executables.
>>
>> Note that if the target compiler is clang, it will probably already
>> support 32-bit ARM via a -target option.  Not sure if you want to try
>> to handle that case here as well?
> 
> Sure, why not.
> 
> Do these multi-arch tests work on FreeBSD?  At least gdb.multi/multi-arch.exp should,
> since FreeBSD supports multi-process nowadays, right?

Yes-ish.  There are some failures that are probably bugs in the FreeBSD
native target (or ptrace behavior with now signal catching works on
FreeBSD that breaks some test assumptions).
  
>> The other wrinkle perhaps is that just because you have a capable
>> compiler around may not mean that you have 32-bit libraries available
>> to link against?  I'm not sure if the compilers you list in your patch
>> ensure that you have an ARM libc around as well via dependencies.
>>
> 
> That is exactly the same as on x86-64.  Just because you have gcc installed,
> which can compile 32-bit x86 with "gcc -m32", it doesn't mean you have the 32-bit
> runtime libs installed.  So any test using -m32 today may run into that issue already.
> 
> However, note the new arm_cc_for_target proc compiles/links AND tries to run the resulting
> program.  If there's some basic 32-bit runtime dependency missing, then the probing
> program won't run successfully, and we'll try the next compiler in the list of potential
> compilers.  If we run out of compilers, then the proc returns "" and the testcase that calls
> arm_cc_for_target ends up bailing out with UNSUPPORTED, like so:
> 
> .    if {[arm_cc_for_target] != ""} {
>          lappend options "compiler=[arm_cc_for_target]"
>      } else {
>          unsupported "ARM compiler is not known"
>          return -1
>      }
> 
> On the aarch64 ubuntu 18.04 machine on the compile farm, the first probed compiler
> (arm-linux-gnueabi-gcc) compiles&links successfully, but then running the built probe
> program fails.  The second alternative doesn't exist on the system (arm-none-linux-gnueabi-gcc),
> so it's skipped, and then the last alternative (arm-linux-gnueabihf-gcc) compiles&links&runs
> the probe program successfully, so that's then the compiler that is used by the actual testcases.

Ok, that works great then.  Right now FreeBSD doesn't build and ship 32-bit
libraries for aarch64 by default.  If that changes in the future I will take
a stab at updating this function to support clang by trying to use -target
if the current CC is already clang.

-- 
John Baldwin