Re: [PATCH, v2] Fix 32-bit build for --enable-targets=all

[Luis Machado <luis.machado@arm.com>]

On 4/26/22 03:52, Alan Modra wrote:
> On Mon, Apr 25, 2022 at 08:40:54AM +0100, Luis Machado wrote:
>> On 4/24/22 15:58, Joel Brobecker wrote:
>>> Looking at this patch, I think you you may not have seen Alan's
>>> comment, which je sent on Apr 18, saying:
>>>
>>>       https://sourceware.org/pipermail/gdb-patches/2022-April/187960.html
>>>       | Anything that requires 64-bit BFD support does not belong in
>>>       | TARGET32_LIBOPCODES_CFILES.  In fact, the whole point of
>>>       | TARGET32_LIBOPCODES_CFILES was to fix --enable-targets=all breakage on
>>>       | 32-bit hosts without --enable-64-bit-bfd.  Why would you want to put
>>>       | bpf here?  It's a 64-bit target!
>>>
>>> (I see that you weren't in the list of direct recipients for that email)
>>>
>>
>> Yes, it looks that way.
>>
>> Unfortunately --enable-targets=all never really worked OK for 32-bit builds
>> after splitting 64/32 targets. It is not clear to me if there are bugs
>> elsewhere that are preventing a clean build, but right now it doesn't look
>> buildable at all.
>>
>> Alan?
> 
> The major problem I have with your patch is that all it does is sweep
> a problem under the rug.  While it may fix a build breakage I doubt
> that it actually improves anything for users.  For example, if I apply
> your patch for a 32-bit --enable-targets=all binutils build, then
> attempt to disassemble one of the bpf gas testsuite objects:
> 
> $ ~/build/gas/all32/binutils/objdump -dr tmpdir/lddw.o
> /home/alan/build/gas/all32/binutils/objdump: tmpdir/lddw.o: file format not recognized
>

Yeah, I agree. Doing some more research on the code, it is the wrong 
approach. Though I'm wondering what an appropriate fix would look like, 
one that wouldn't require a lot of work and would still make things 
build cleanly. A compromise?

> That's due to lack of the required support from bfd/elf64-bpf.c to
> load bpf object files into BFD.  I think you'll find a similar result
> for the other targets your patch touches, and not just with objdump
> but with everything else that uses libbfd.
> 
> There is also a minor problem with the patch in that it adds entries
> to TARGET32_LIBOPCODES_CFILES without removing the corresponding
> entries from TARGET64_LIBOPCODES_CFILES.  Similarly for the defines in
> opcodes/disassemble.c.
>

Maybe some targets shouldn't be accepted as part of a 32-bit 
--enable-targets=all for GDB/sim. Removing those "problematic" targets 
from the list might be another possible approach I suppose.