Re: [Qemu-arm] [PATCH 66/67] target/arm: Move singlestep check from gen_jmp to gen_goto_tb

[Richard Henderson]

On 7/26/19 11:13 AM, Peter Maydell wrote:
> On Fri, 26 Jul 2019 at 18:51, Richard Henderson
> <address@hidden> wrote:
>>
>> We miss quite a number of single-step events by having
>> the check in the wrong place.
>>
>> Signed-off-by: Richard Henderson <address@hidden>
>> ---
>>  target/arm/translate.c | 16 ++++++----------
>>  1 file changed, 6 insertions(+), 10 deletions(-)
>>
>> diff --git a/target/arm/translate.c b/target/arm/translate.c
>> index c2b8b86fd2..9ae9b23823 100644
>> --- a/target/arm/translate.c
>> +++ b/target/arm/translate.c
>> @@ -2740,7 +2740,10 @@ static void gen_goto_ptr(void)
>>   */
>>  static void gen_goto_tb(DisasContext *s, int n, target_ulong dest)
>>  {
>> -    if (use_goto_tb(s, dest)) {
>> +    if (unlikely(is_singlestepping(s))) {
>> +        gen_set_pc_im(s, dest);
>> +        gen_singlestep_exception(s);
>> +    } else if (use_goto_tb(s, dest)) {
>>          tcg_gen_goto_tb(n);
>>          gen_set_pc_im(s, dest);
>>          tcg_gen_exit_tb(s->base.tb, n);
>> @@ -2751,16 +2754,9 @@ static void gen_goto_tb(DisasContext *s, int n, 
>> target_ulong dest)
>>      s->base.is_jmp = DISAS_NORETURN;
>>  }
>>
>> -static inline void gen_jmp (DisasContext *s, uint32_t dest)
>> +static inline void gen_jmp(DisasContext *s, uint32_t dest)
>>  {
>> -    if (unlikely(is_singlestepping(s))) {
>> -        /* An indirect jump so that we still trigger the debug exception.  
>> */
>> -        if (s->thumb)
>> -            dest |= 1;
>> -        gen_bx_im(s, dest);
>> -    } else {
>> -        gen_goto_tb(s, 0, dest);
>> -    }
>> +    gen_goto_tb(s, 0, dest);
>>  }
>>
>>  static inline void gen_mulxy(TCGv_i32 t0, TCGv_i32 t1, int x, int y)
> 
> I haven't tested but I'm not sure this change is right.
> The idea of the current code is that we handle generating
> the actual singlestep exception in arm_tr_tb_stop() at the
> end, rather than in the process of generating code for the
> insn. This change means we now do a gen_singlestep_exception()
> in the gen_jmp()/gen_goto_tb() part of the code, but we haven't
> removed the handling of it in arm_tr_tb_stop(), so we're now
> doing this in two places. Why doesn't the current design work?

Note that the existing gen_goto_tb does not test for single-step, and always
emits NORETURN code, either via tcg_gen_goto_tb() or
tcg_gen_lookup_and_goto_ptr().  The singlestep exception is neither generated
nor would it be reachable.

Another way to do handle this would be to test single-step but set DISAS_JUMP
instead of actually generate the singlestep exception in gen_goto_tb.  Do you
prefer that method?

> And if we need to do something different for the route to
> "change the PC via gen_jmp()" why don't we need to do it
> also when we change the PC via eg gen_bx_im() ?

See patch 67.

>  * four places for barrier insns where we use a
>    gen_goto_tb to end the TB -- this isn't consistent
>    with how we end the TB for other situations...)

Fixing this sounds like a good cleanup.


r~