Re: [Qemu-devel] [PATCH for 1.4 0/4] Simplify Makefile.objs some more

[Paolo Bonzini]

Il 20/01/2013 00:22, Andreas Färber ha scritto:
> Am 19.01.2013 23:06, schrieb Paolo Bonzini:
>> Il 19/01/2013 19:11, Andreas Färber ha scritto:
>>>>> The patches are mostly mechanical substitutions, and there is no
>>>>> user-visible change---neither in total build time, nor in the files that
>>>>> are linked into the executables.
>>> Without having tested this yet I want to remind that it is necessary for
>>> qom/cpu.c to be built twice
>>
>> Hmm, it's not anymore actually (since libuser was removed).  It hasn't
>> been built twice for a month and apparently nothing broke.
> 
> I surely didn't ack that.

At the time I did the move, there was no conditional CPUState field
(commit 8e98e2e80b92e08e79e27a0c20a172906cfa12d2).  The only difference
between the user and softmmu qom/cpu.c was that the user version ended
up in libuser.  ISTR I asked around about libuser, and the conclusion
was that it was only needed for --enable-user-pie.

> Have you actually tested linux-user to verify
> it works? It might lead to unexpected CPUState field accesses.

As I said, at the time I did the move there was no difference between
the two.  Now, I am indeed quite surprised that it works, but yes---it
does, at least trivial /bin/ls.

> You are judging based on master. I have some more code movements queued
> (qom-cpu-8) and I believe it was Anthony who insisted on suppressing
> those unneeded user-only fields even if they were unconditional in
> CPU_COMMON before.
> 
> qom/cpu.c is not intended to remain so small forever - any cpu_* code
> that does not depend on CPUArchState can find a new home there.
> cpu_interrupt() is being moved to qom/cpu.h and cpu_reset_interrupt() to
> qom/cpu.c for instance. And I'm working on refactoring CPU VMState, that
> either requires #ifdef'ery or lots of new stubs beyond what Eduardo added.

You can place those in a qom/cpu-softmmu.c.

What I care about is having an understandable build system.  At some
point we had 12 recursive -obj-y variables and a few other non-recursive
ones.  The interactions and inclusions were completely impossible to
understand.

Paolo