Re: [Qemu-devel] Re: [PATCH] target-ppc: ext32u instead of andi with constant

[Alexander Graf]

Am 22.03.2011 um 10:09 schrieb Aurelien Jarno <address@hidden>:

> On Tue, Mar 22, 2011 at 08:36:26AM +0100, Alexander Graf wrote:
>> 
>> On 22.03.2011, at 07:41, Aurelien Jarno wrote:
>> 
>>> Cc: Alexander Graf <address@hidden>
>>> Signed-off-by: Aurelien Jarno <address@hidden>
>>> ---
>>> target-ppc/translate.c |    8 ++++----
>>> 1 files changed, 4 insertions(+), 4 deletions(-)
>>> 
>>> diff --git a/target-ppc/translate.c b/target-ppc/translate.c
>>> index 3d265e3..49eab28 100644
>>> --- a/target-ppc/translate.c
>>> +++ b/target-ppc/translate.c
>>> @@ -6975,7 +6975,7 @@ static inline void gen_evmergelo(DisasContext *ctx)
>>> #if defined(TARGET_PPC64)
>>>    TCGv t0 = tcg_temp_new();
>>>    TCGv t1 = tcg_temp_new();
>>> -    tcg_gen_andi_tl(t0, cpu_gpr[rB(ctx->opcode)], 0x00000000FFFFFFFFLL);
>>> +    tcg_gen_ext32u_tl(t0, cpu_gpr[rB(ctx->opcode)]);
>>>    tcg_gen_shli_tl(t1, cpu_gpr[rA(ctx->opcode)], 32);
>>>    tcg_gen_or_tl(cpu_gpr[rD(ctx->opcode)], t0, t1);
>>>    tcg_temp_free(t0);
>>> @@ -6994,7 +6994,7 @@ static inline void gen_evmergehilo(DisasContext *ctx)
>>> #if defined(TARGET_PPC64)
>>>    TCGv t0 = tcg_temp_new();
>>>    TCGv t1 = tcg_temp_new();
>>> -    tcg_gen_andi_tl(t0, cpu_gpr[rB(ctx->opcode)], 0x00000000FFFFFFFFLL);
>>> +    tcg_gen_ext32u_tl(t0, cpu_gpr[rB(ctx->opcode)]);
>>>    tcg_gen_andi_tl(t1, cpu_gpr[rA(ctx->opcode)], 0xFFFFFFFF0000000ULL);
>> 
>> Wouldn't deposit make sense here? But that can be a later optimization.
> 
> Indeed it makes sense here, the thing is that I don't really know how 
> deposit is going. We have merged it, but we don't have the expected 
> performance (ie no improvement).

Yup, as on x86 we have to special-case a lot. The current x86 deposit 
implementation doesn't implement too many of those :).

On ppc for example, the world looks different. And even without speedup, I like 
the cleanup it provides.

Btw, I do use deposit on the s390 target implementation that is finally able to 
run Debian guests as well. All I need to do to finish it up is to properly 
split it up into smaller readable patches. So please decide on deposit soon, 
otherwise we'll have a user merged :).


Alex

>