Re: [RFC PATCH v3 18/28] target/arm: Move common cpu code into cpu.c

[Fabiano Rosas]

Richard Henderson <richard.henderson@linaro.org> writes:

> On 1/13/23 06:04, Fabiano Rosas wrote:
>> The cpu_tcg.c file about to be moved into the tcg directory. Move the
>> code that is needed for cpus that also work with KVM into cpu.c.
>> 
>> Signed-off-by: Fabiano Rosas <farosas@suse.de>
>> ---
>>   target/arm/cpu.c     | 76 +++++++++++++++++++++++++++++++++++++++++++
>>   target/arm/cpu_tcg.c | 77 --------------------------------------------
>>   2 files changed, 76 insertions(+), 77 deletions(-)
>
> Actually, not true.  This is tcg-only.  As is the bulk of aarch64_max_initfn 
> from which 
> this is called -- note the first 4 lines of that function:
>
>      if (kvm_enabled() || hvf_enabled()) {
>          /* With KVM or HVF, '-cpu max' is identical to '-cpu host' */
>          aarch64_host_initfn(obj);
>          return;
>      }
>
> Thus the rest of the function is only reachable for tcg.

Sigh... It seems it's not that simple:

We can currently have a QEMU invocation with "-accel qtest -cpu max" and
no other accelerator. Currently, it falls into this implicit else
branch. So this is actually "else if (tcg_enabled() || qtest_enabled())".

If I move the "TCG-only" code under CONFIG_TCG, the qtests that use -cpu
max will break.

So I have chosen to move the code which depends on aa32_max_features as
you suggest into tcg/ but kept the cortex-a57 as a baseline for
qtest. This has the effect of causing "-cpu max" for the tests to be a
slightly different CPU depending on whether TCG is built in (which
perhaps is ok because if the tests depended on cpu features they should
specify an accel?).