On 4 August 2014 05:12, Edgar E. Iglesias <edgar.igles...@gmail.com> wrote:
> On Fri, Aug 01, 2014 at 03:21:08PM +0100, Peter Maydell wrote:
>> On 17 June 2014 09:45, Edgar E. Iglesias <edgar.igles...@gmail.com> wrote:
>> > + case 2:
>> > + if (!arm_dc_feature(s, ARM_FEATURE_EL2) || s->current_pl ==
>> > 0) {
>> > + unallocated_encoding(s);
>> > + break;
>> > + }
>> > + tmp = tcg_const_i32(syn_aa64_hvc(imm16));
>> > + gen_a64_set_pc_im(s->pc);
>>
>> (This set_pc_im is unnecessary.)
>>
>> > + gen_helper_hvc(cpu_env, tmp);
>>
>> This means that exceptions due to HVC are going to be
>> runtime-detected and cause us to go through and retranslate
>> the TB to determine the guest PC. Maybe we should do:
>>
>> /* This helper will raise EXCP_UDEF if HVC is not permitted */
>> gen_helper_hvc_access_check(cpu_env);
>> /* Common case: HVC causes EXCP_HVC */
>> gen_exception_insn(s, 0, EXCP_HVC, syn_aa64_hvc(imm16));
>>
>> Then you only get the overhead of retranslating the TB in the
>> unexpected case where the guest does something dumb and
>> executes an HVC that UNDEFs.
>
> That doesn't match my understanding of what will happen with this kind
> of exception raising. I think the set_pc_im matters and there won't
> be any retranslation of TBs to figure out the guest PC.
Sorry, yes; you're right and I was wrong -- we only retranslate
where we call cpu_restore_state(), which is done only where
tlb_fill() is going to raise an exception.
(I think I need to think a bit about how I'm currently implementing
architectural debug singlestep, since at the moment I assume
that you can tell at translate time whether something is going
to be a valid SMC/HVC/SVC or not, and so whether or not to
advance the singlestep state machine. Maybe I can defer that
to exception entry...)
thanks
-- PMM
