On 2/19/19 6:48 AM, Alex Bennée wrote:
>> +++ b/target/arm/translate-a64.c
>> @@ -1841,11 +1841,15 @@ static void handle_sys(DisasContext *s, uint32_t
>> insn, bool isread,
>> /* I/O operations must end the TB here (whether read or write) */
>> gen_io_end();
>> s->base.is_jmp = DISAS_UPDATE;
>> - } else if (!isread && !(ri->type & ARM_CP_SUPPRESS_TB_END)) {
>> + }
>> + if (!isread && !(ri->type & ARM_CP_SUPPRESS_TB_END)) {
> Does this potentially introduce a icount bug (or fix an existing bug)?
>
>> /* We default to ending the TB on a coprocessor register write,
>> * but allow this to be suppressed by the register definition
>> * (usually only necessary to work around guest bugs).
>> */
>> + TCGv_i32 tcg_el = tcg_const_i32(s->current_el);
>> + gen_helper_rebuild_hflags_a64(cpu_env, tcg_el);
>> + tcg_temp_free_i32(tcg_el);
>> s->base.is_jmp = DISAS_UPDATE;
>> }
Neither.
Previously, all that was required was that either icount or !SUPPRESS_TB_END
and a TB, and that was done simply by setting DISAS_UPDATE. Which both if
blocks did.
Now, icount ends a TB (still by setting DISAS_UPDATE). But if !SUPPRESS_TB_END
then we must also rebuild hflags (with a possible harmless double-set of
DISAS_UPDATE).
r~
