Re: [PATCH v12 02/15] accel: collecting TB execution count

[Wu, Fei]

On 5/23/2023 8:45 AM, Richard Henderson wrote:
> On 5/18/23 06:57, Fei Wu wrote:
>> +void HELPER(inc_exec_freq)(void *ptr)
>> +{
>> +    TBStatistics *stats = (TBStatistics *) ptr;
>> +    tcg_debug_assert(stats);
>> +    ++stats->executions.normal;
>> +}
> ...
>> +static inline void gen_tb_exec_count(TranslationBlock *tb)
>> +{
>> +    if (tb_stats_enabled(tb, TB_EXEC_STATS)) {
>> +        TCGv_ptr ptr = tcg_temp_new_ptr();
>> +        tcg_gen_movi_ptr(ptr, (intptr_t)tb->tb_stats);
>> +        gen_helper_inc_exec_freq(ptr);
>> +    }
>> +}
> 
> This is 3 host instructions, easily expanded inline:
> 
> --- a/accel/tcg/translator.c
> +++ b/accel/tcg/translator.c
> @@ -11,6 +11,7 @@
>  #include "qemu/error-report.h"
>  #include "tcg/tcg.h"
>  #include "tcg/tcg-op.h"
> +#include "tcg/tcg-temp-internal.h"
>  #include "exec/exec-all.h"
>  #include "exec/gen-icount.h"
>  #include "exec/log.h"
> @@ -18,6 +19,30 @@
>  #include "exec/plugin-gen.h"
>  #include "exec/replay-core.h"
> 
> +
> +static void gen_tb_exec_count(TranslationBlock *tb)
> +{
> +    if (tb_stats_enabled(tb, TB_EXEC_STATS)) {
> +        TCGv_ptr ptr = tcg_temp_ebb_new_ptr();
> +
> +        tcg_gen_movi_ptr(ptr, (intptr_t)&tb->tb_stats->executions.normal);
> +        if (sizeof(tb->tb_stats->executions.normal) == 4) {
> +            TCGv_i32 t = tcg_temp_ebb_new_i32();
> +            tcg_gen_ld_i32(t, ptr, 0);
> +            tcg_gen_addi_i32(t, t, 1);
> +            tcg_gen_st_i32(t, ptr, 0);
> +            tcg_temp_free_i32(t);
> +        } else {
> +            TCGv_i64 t = tcg_temp_ebb_new_i64();
> +            tcg_gen_ld_i64(t, ptr, 0);
> +            tcg_gen_addi_i64(t, t, 1);
> +            tcg_gen_st_i64(t, ptr, 0);
> +            tcg_temp_free_i64(t);
> +        }
> +        tcg_temp_free_ptr(ptr);
> +    }
> +}
> +
>  bool translator_use_goto_tb(DisasContextBase *db, target_ulong dest)
>  {
>      /* Suppress goto_tb if requested. */
> 
> 
> I'm not expecially keen on embedding the TBStatistics pointer directly
> like this; for most hosts we will have to put this constant into the
> constant pool.  Whereas the pointer already exists at tb->tb_stats, and
> tb is at a constant displacement prior to the code, so we already have
> mechanisms for generating pc-relative addresses.
> 
> However, that's premature optimization.  Let's get it working first.
> 
Here is the coremark results on a 4c qemu-system-riscv64, coremark takes
1cpu in this test.

(tb_stats stop)                         Iterations/Sec   : 5358.012664
(tb_stats start all)
    helper - qatomic_inc(normal)        Iterations/Sec   : 2416.626390
    helper - ++normal                   Iterations/Sec   : 4307.559767
    no helper - inline add              Iterations/Sec   : 5168.930031

Also if coremark runs on 4cpu, tb_stats will cost much more even in the
inline case, it's an extreme case though.

Thanks,
Fei.

> 
> r~
>