Re: [Qemu-devel] [PATCH 09/10] ppc: Move exception generation code out of line

["Cédric Le Goater" <clg@kaod.org>]

On 06/13/2016 09:44 AM, Thomas Huth wrote:
> On 13.06.2016 07:24, Cédric Le Goater wrote:
>> From: Benjamin Herrenschmidt <benh@kernel.crashing.org>
>>
>> There's no point inlining this, if you hit the exception case you exit
>> anyway, and not inlining saves about 100K of code size (and cache
>> footprint).
>>
>> Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
>> ---
>>  target-ppc/translate.c | 9 ++++++---
>>  1 file changed, 6 insertions(+), 3 deletions(-)
>>
>> diff --git a/target-ppc/translate.c b/target-ppc/translate.c
>> index f211d175c09c..600d5db2bb9a 100644
>> --- a/target-ppc/translate.c
>> +++ b/target-ppc/translate.c
>> @@ -283,7 +283,8 @@ void gen_update_current_nip(void *opaque)
>>      tcg_gen_movi_tl(cpu_nip, ctx->nip);
>>  }
>>  
>> -static inline void gen_exception_err(DisasContext *ctx, uint32_t excp, uint32_t error)
>> +static void __attribute__((noinline))
>> +gen_exception_err(DisasContext *ctx, uint32_t excp, uint32_t error)
>>  {
>>      TCGv_i32 t0, t1;
>>      if (ctx->exception == POWERPC_EXCP_NONE) {
>> @@ -297,7 +298,8 @@ static inline void gen_exception_err(DisasContext *ctx, uint32_t excp, uint32_t
>>      ctx->exception = (excp);
>>  }
>>  
>> -static inline void gen_exception(DisasContext *ctx, uint32_t excp)
>> +static void __attribute__((noinline))
>> +gen_exception(DisasContext *ctx, uint32_t excp)
>>  {
>>      TCGv_i32 t0;
>>      if (ctx->exception == POWERPC_EXCP_NONE) {
>> @@ -309,7 +311,8 @@ static inline void gen_exception(DisasContext *ctx, uint32_t excp)
>>      ctx->exception = (excp);
>>  }
>>  
>> -static inline void gen_debug_exception(DisasContext *ctx)
>> +static void __attribute__((noinline))
>> +gen_debug_exception(DisasContext *ctx)
>>  {
>>      TCGv_i32 t0;
> 
> Do you get the same results if you just remove the "inline" keyword,
> without adding the "__attribute__((noinline))" ? If yes, I'd suggest to
> do this patch without the "__attribute__((noinline))" - that's easier to
> read, and the compiler can still decide to inline something in case it's
> better one a certain architecture.

Yes. They are no differences. 

The interesting part though is that the .text is about the same size. 
There is even a slight increase of ~2K with gcc 4.9.2 (intel host) and 
a slight decrease of ~1K with gcc 5.3.1 (ppc64le host).

I guess we can just drop that patch. It does not seem to bring much.

Thanks,

C.