Re: [Qemu-devel] [PATCH 03/14] target-ppc: use separate indices for various translation modes

[Richard Henderson <rth@twiddle.net>]

On 09/16/2014 10:20 AM, Tom Musta wrote:
> On 9/15/2014 10:03 AM, Paolo Bonzini wrote:
>> PowerPC TCG flushes the TLB on every IR/DR change, which basically
>> means on every user<->kernel context switch.  Encode IR/DR in the
>> MMU index.
>>
>> This brings the number of TLB flushes down from ~900000 to ~50000
>> for starting up the Debian installer, which is in line with x86
>> and gives a ~10% performance improvement.
>>
>> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
>> ---
>>  target-ppc/cpu.h         |  7 ++-----
>>  target-ppc/excp_helper.c |  3 ---
>>  target-ppc/helper_regs.h | 11 ++++++-----
>>  3 files changed, 8 insertions(+), 13 deletions(-)
>>
>> diff --git a/target-ppc/cpu.h b/target-ppc/cpu.h
>> index b64c652..c29ce3b 100644
>> --- a/target-ppc/cpu.h
>> +++ b/target-ppc/cpu.h
>> @@ -922,7 +922,8 @@ struct ppc_segment_page_sizes {
>>  
>>  /*****************************************************************************/
>>  /* The whole PowerPC CPU context */
>> -#define NB_MMU_MODES 3
>> +#define NB_MMU_MODES 12
>> +#define MMU_USER_IDX 3  /* PR=IR=DR=1 */
> 
> This doesn't build for me:
> 
>   CC    ppc64-softmmu/tcg/tcg.o
> In file included from /bghome/tmusta/powerisa/qemu/qemu/tcg/tcg.c:264:
> /bghome/tmusta/powerisa/qemu/qemu/tcg/ppc/tcg-target.c: In function ?tcg_out_tlb_read?:
> /bghome/tmusta/powerisa/qemu/qemu/tcg/ppc/tcg-target.c:1394: error: size of array ?qemu_build_bug_on__1396? is negative
> make[1]: *** [tcg/tcg.o] Error 1
> make: *** [subdir-ppc64-softmmu] Error 2
> 
> which correlates with this:
> 
>   1389      /* Compensate for very large offsets.  */
>   1390      if (add_off >= 0x8000) {
>   1391          /* Most target env are smaller than 32k; none are larger than 64k.
>   1392             Simplify the logic here merely to offset by 0x7ff0, giving us a
>   1393             range just shy of 64k.  Check this assumption.  */
>   1394          QEMU_BUILD_BUG_ON(offsetof(CPUArchState,
>   1395                                     tlb_table[NB_MMU_MODES - 1][1])
>   1396                            > 0x7ff0 + 0x7fff);
>   1397          tcg_out32(s, ADDI | TAI(TCG_REG_TMP1, base, 0x7ff0));
>   1398          base = TCG_REG_TMP1;
>   1399          cmp_off -= 0x7ff0;
>   1400          add_off -= 0x7ff0;
>   1401      }

Ouch, yes indeed.

While we could probably fix this for ppc (using addis), it's not nearly so
easily fixable for arm -- without impacting performance anyway.

Does 96k worth of TLBs really help that much?  Are all 12 of them actually
used?  Can we use a more complex encoding scheme for the mmu_idx and use less?


r~