[PATCH] gdb: restrict configure error with all targets and 64 bit bfd to mips

[Maciej W. Rozycki]

On Wed, 5 Feb 2025, Maciej W. Rozycki wrote:

> > I spoke to Maciej, and actually what is going on is that mips really needs 64
> > bit bfd, and requests that in the bfd/configure file.
> > 
> > Somehow that isn't propagated, and both gdb's and opcodes's configure doesn't
> > get that propagated correctly. The proper fix would involve propagating things
> > correctly in several configure files, and I don't even understand why it
> > fails, so it is a bit over my head...
> 
>  There's something fishy going on here that has to be investigated.  It 
> may come from native 32-bit MIPS configurations such as `mips-linux-gnu' 
> requiring 64-bit BFD, uniquely I believe across 32-bit native builds, so 
> if the host and the default targets are `mips-linux-gnu' both at a time, 
> then 64-bit BFD must have somehow been enabled anyway.  Weird that it 
> seems as if not.
> 
>  I'm not sure offhand if my native 32-bit MIPS environment is up to the 
> task toolchain-wise, or whether it would be feasible for me time-wise to 
> set up a crossed build of native 32-bit MIPS GDB, as I reckon there are 
> extra host library requirements for GDB, but I'll see if I can reproduce 
> it sometime.

 So I don't have a C++17 compiler in the first place in my native 32-bit 
MIPS environment, but I was able to trigger this issue with a crossed 
native `mips-linux-gnu' `--enable-targets=all' build.

 And after a little detective work I can see where the configury logic is 
flawed: configuring for a 32-bit host enables 64-bit BFD implicitly via 
this statement in bfd/configure.ac:

case ${host64}-${target64}-${want64} in
  *true*)
    wordsize=64
    [...]
    ;;
  false-false-false)
    wordsize=32
    all_backends='$(BFD32_BACKENDS)'
    ;;
esac

(notice how wordsize=64 can be set regardless of the host pointer size), 
but config/bfd64.m4 only checks for `--enable-64-bit-bfd' and the host 
pointer size, which in preset ${want64} and set ${host64} respectively in 
bfd/configure.ac, ignoring any ${want64} overrides or the ${target64} 
part.

 Following this observation I also triggered this build failure with a 
`--host=i386-linux-gnu --target=riscv64-linux-gnu --enable-targets=all' 
Canadian cross configuration; it has nothing specific to do with the MIPS 
host and said configuration still breaks at link time with our trunk (it's 
libopcodes rather than GDB that fails anyway).

 I think the correct way to get this mess sorted is to move BFD_64_BIT 
code from config/bfd64.m4 over to bfd/configure.ac for its private use 
only and write replacement BFD_64_BIT that just checks for BFD64, having 
included "bfd.h".

 Cc'ing binutils as the problem spans across.  I have posted a proposed 
fix now; cf. 
<https://inbox.sourceware.org/gdb-patches/a7356a9c-ea48-ff18-d99b-4e2840a871fc@redhat.com/>.

  Maciej