Re: [Qemu-devel] [RFC v6 10/14] softmmu: Simplify helper_*_st_name, wrap unaligned code

[Alex Bennée]

alvise rigo <address@hidden> writes:

> On Thu, Jan 7, 2016 at 5:35 PM, Alex Bennée <address@hidden> wrote:
>>
>> alvise rigo <address@hidden> writes:
>>
>>> On Thu, Jan 7, 2016 at 3:46 PM, Alex Bennée <address@hidden> wrote:
>>>>
>>>> Alvise Rigo <address@hidden> writes:
>>>>
>>>>> Attempting to simplify the helper_*_st_name, wrap the
>>>>> do_unaligned_access code into an inline function.
>>>>> Remove also the goto statement.
>>>>
>>>> As I said in the other thread I think these sort of clean-ups can come
>>>> before the ll/sc implementations and potentially get merged ahead of the
>>>> rest of it.
>>>>
>>>>>
>>>>> Suggested-by: Jani Kokkonen <address@hidden>
>>>>> Suggested-by: Claudio Fontana <address@hidden>
>>>>> Signed-off-by: Alvise Rigo <address@hidden>
>>>>> ---
>>>>>  softmmu_template.h | 96 
>>>>> ++++++++++++++++++++++++++++++++++--------------------
>>>>>  1 file changed, 60 insertions(+), 36 deletions(-)
>>>>>
>>>>> diff --git a/softmmu_template.h b/softmmu_template.h
>>>>> index d3d5902..92f92b1 100644
>>>>> --- a/softmmu_template.h
>>>>> +++ b/softmmu_template.h
>>>>> @@ -370,6 +370,32 @@ static inline void glue(io_write, 
>>>>> SUFFIX)(CPUArchState *env,
>>>>>                                   iotlbentry->attrs);
>>>>>  }
>>>>>
>>>>> +static inline void glue(helper_le_st_name, _do_unl_access)(CPUArchState 
>>>>> *env,
>>>>> +                                                           DATA_TYPE val,
>>>>> +                                                           target_ulong 
>>>>> addr,
>>>>> +                                                           TCGMemOpIdx 
>>>>> oi,
>>>>> +                                                           unsigned 
>>>>> mmu_idx,
>>>>> +                                                           uintptr_t 
>>>>> retaddr)
>>>>> +{
>>>>> +    int i;
>>>>> +
>>>>> +    if ((get_memop(oi) & MO_AMASK) == MO_ALIGN) {
>>>>> +        cpu_unaligned_access(ENV_GET_CPU(env), addr, MMU_DATA_STORE,
>>>>> +                             mmu_idx, retaddr);
>>>>> +    }
>>>>> +    /* XXX: not efficient, but simple */
>>>>> +    /* Note: relies on the fact that tlb_fill() does not remove the
>>>>> +     * previous page from the TLB cache.  */
>>>>> +    for (i = DATA_SIZE - 1; i >= 0; i--) {
>>>>> +        /* Little-endian extract.  */
>>>>> +        uint8_t val8 = val >> (i * 8);
>>>>> +        /* Note the adjustment at the beginning of the function.
>>>>> +           Undo that for the recursion.  */
>>>>> +        glue(helper_ret_stb, MMUSUFFIX)(env, addr + i, val8,
>>>>> +                                        oi, retaddr + GETPC_ADJ);
>>>>> +    }
>>>>> +}
>>>>
>>>> There is still duplication of 99% of the code here which is silly given
>>>
>>> Then why should we keep this template-like design in the first place?
>>> I tried to keep the code duplication for performance reasons
>>> (otherwise how can we justify the two almost identical versions of the
>>> helper?), while making the code more compact and readable.
>>
>> We shouldn't really - code duplication is bad for all the well known
>> reasons. The main reason we need explicit helpers for the be/le case are
>> because they are called directly from the TCG code which encodes the
>> endianess decision in the call it makes. However that doesn't stop us
>> making generic inline helpers (helpers for the helpers ;-) which the
>> compiler can sort out.
>
> I thought you wanted to make conditional all the le/be differences not
> just those helpers for the helpers...

That would be nice for it all but that involves tweaking the TCG->helper
calls themselves. However if we are re-factoring common stuff from those
helpers into inlines then we can at least reduce the duplication there.

> So, if we are allowed to introduce this small overhead, all the
> helper_{le,be}_st_name_do_{unl,mmio,ram}_access can be squashed to
> helper_generic_st_do_{unl,mmio,ram}_access. I think this is want you
> proposed in the POC, right?

Well in theory it shouldn't introduce any overhead. However my proof is
currently waiting on a bug fix to GDB's dissas command so I can show you
the side by side assembly dump ;-)

>>>> the compiler inlines the code anyway. If we gave the helper a more
>>>> generic name and passed the endianess in via args I would hope the
>>>> compiler did the sensible thing and constant fold the code. Something
>>>> like:
>>>>
>>>> static inline void glue(helper_generic_st_name, _do_unl_access)
>>>>                         (CPUArchState *env,
>>>>                         bool little_endian,
>>>>                         DATA_TYPE val,
>>>>                         target_ulong addr,
>>>>                         TCGMemOpIdx oi,
>>>>                         unsigned mmu_idx,
>>>>                         uintptr_t retaddr)
>>>> {
>>>>     int i;
>>>>
>>>>     if ((get_memop(oi) & MO_AMASK) == MO_ALIGN) {
>>>>         cpu_unaligned_access(ENV_GET_CPU(env), addr, MMU_DATA_STORE,
>>>>                              mmu_idx, retaddr);
>>>>     }
>>>>     /* Note: relies on the fact that tlb_fill() does not remove the
>>>>      * previous page from the TLB cache.  */
>>>>     for (i = DATA_SIZE - 1; i >= 0; i--) {
>>>>         if (little_endian) {
>>>
>>> little_endian will have >99% of the time the same value, does it make
>>> sense to have a branch here?
>>
>> The compiler should detect that little_endian is constant when it
>> inlines the code and not bother generating a test/branch case for
>> something that will never happen.
>>
>> Even if it did though I doubt a local branch would stall the processor
>> that much, have you counted how many instructions we execute once we are
>> on the slow path?
>
> Too many :)

Indeed, that is why its SLOOOOW ;-)
>
> Regards,
> alvise
>
>>
>>>
>>> Thank you,
>>> alvise
>>>
>>>>                 /* Little-endian extract.  */
>>>>                 uint8_t val8 = val >> (i * 8);
>>>>         } else {
>>>>                 /* Big-endian extract.  */
>>>>                 uint8_t val8 = val >> (((DATA_SIZE - 1) * 8) - (i * 8));
>>>>         }
>>>>         /* Note the adjustment at the beginning of the function.
>>>>            Undo that for the recursion.  */
>>>>         glue(helper_ret_stb, MMUSUFFIX)(env, addr + i, val8,
>>>>                                         oi, retaddr + GETPC_ADJ);
>>>>     }
>>>> }
>>>>
>>>>
>>>>> +
>>>>>  void helper_le_st_name(CPUArchState *env, target_ulong addr, DATA_TYPE 
>>>>> val,
>>>>>                         TCGMemOpIdx oi, uintptr_t retaddr)
>>>>>  {
>>>>> @@ -433,7 +459,8 @@ void helper_le_st_name(CPUArchState *env, 
>>>>> target_ulong addr, DATA_TYPE val,
>>>>>              return;
>>>>>          } else {
>>>>>              if ((addr & (DATA_SIZE - 1)) != 0) {
>>>>> -                goto do_unaligned_access;
>>>>> +                glue(helper_le_st_name, _do_unl_access)(env, val, addr, 
>>>>> mmu_idx,
>>>>> +                                                        oi, retaddr);
>>>>>              }
>>>>>              iotlbentry = &env->iotlb[mmu_idx][index];
>>>>>
>>>>> @@ -449,23 +476,8 @@ void helper_le_st_name(CPUArchState *env, 
>>>>> target_ulong addr, DATA_TYPE val,
>>>>>      if (DATA_SIZE > 1
>>>>>          && unlikely((addr & ~TARGET_PAGE_MASK) + DATA_SIZE - 1
>>>>>                       >= TARGET_PAGE_SIZE)) {
>>>>> -        int i;
>>>>> -    do_unaligned_access:
>>>>> -        if ((get_memop(oi) & MO_AMASK) == MO_ALIGN) {
>>>>> -            cpu_unaligned_access(ENV_GET_CPU(env), addr, MMU_DATA_STORE,
>>>>> -                                 mmu_idx, retaddr);
>>>>> -        }
>>>>> -        /* XXX: not efficient, but simple */
>>>>> -        /* Note: relies on the fact that tlb_fill() does not remove the
>>>>> -         * previous page from the TLB cache.  */
>>>>> -        for (i = DATA_SIZE - 1; i >= 0; i--) {
>>>>> -            /* Little-endian extract.  */
>>>>> -            uint8_t val8 = val >> (i * 8);
>>>>> -            /* Note the adjustment at the beginning of the function.
>>>>> -               Undo that for the recursion.  */
>>>>> -            glue(helper_ret_stb, MMUSUFFIX)(env, addr + i, val8,
>>>>> -                                            oi, retaddr + GETPC_ADJ);
>>>>> -        }
>>>>> +        glue(helper_le_st_name, _do_unl_access)(env, val, addr, oi, 
>>>>> mmu_idx,
>>>>> +                                                retaddr);
>>>>>          return;
>>>>>      }
>>>>>
>>>>> @@ -485,6 +497,32 @@ void helper_le_st_name(CPUArchState *env, 
>>>>> target_ulong addr, DATA_TYPE val,
>>>>>  }
>>>>>
>>>>>  #if DATA_SIZE > 1
>>>>> +static inline void glue(helper_be_st_name, _do_unl_access)(CPUArchState 
>>>>> *env,
>>>>> +                                                           DATA_TYPE val,
>>>>> +                                                           target_ulong 
>>>>> addr,
>>>>> +                                                           TCGMemOpIdx 
>>>>> oi,
>>>>> +                                                           unsigned 
>>>>> mmu_idx,
>>>>> +                                                           uintptr_t 
>>>>> retaddr)
>>>>> +{
>>>>> +    int i;
>>>>> +
>>>>> +    if ((get_memop(oi) & MO_AMASK) == MO_ALIGN) {
>>>>> +        cpu_unaligned_access(ENV_GET_CPU(env), addr, MMU_DATA_STORE,
>>>>> +                             mmu_idx, retaddr);
>>>>> +    }
>>>>> +    /* XXX: not efficient, but simple */
>>>>> +    /* Note: relies on the fact that tlb_fill() does not remove the
>>>>> +     * previous page from the TLB cache.  */
>>>>> +    for (i = DATA_SIZE - 1; i >= 0; i--) {
>>>>> +        /* Big-endian extract.  */
>>>>> +        uint8_t val8 = val >> (((DATA_SIZE - 1) * 8) - (i * 8));
>>>>> +        /* Note the adjustment at the beginning of the function.
>>>>> +           Undo that for the recursion.  */
>>>>> +        glue(helper_ret_stb, MMUSUFFIX)(env, addr + i, val8,
>>>>> +                                        oi, retaddr + GETPC_ADJ);
>>>>> +    }
>>>>> +}
>>>>
>>>> Not that it matters if you combine to two as suggested because anything
>>>> not called shouldn't generate the code.
>>>>
>>>>>  void helper_be_st_name(CPUArchState *env, target_ulong addr, DATA_TYPE 
>>>>> val,
>>>>>                         TCGMemOpIdx oi, uintptr_t retaddr)
>>>>>  {
>>>>> @@ -548,7 +586,8 @@ void helper_be_st_name(CPUArchState *env, 
>>>>> target_ulong addr, DATA_TYPE val,
>>>>>              return;
>>>>>          } else {
>>>>>              if ((addr & (DATA_SIZE - 1)) != 0) {
>>>>> -                goto do_unaligned_access;
>>>>> +                glue(helper_be_st_name, _do_unl_access)(env, val, addr, 
>>>>> mmu_idx,
>>>>> +                                                        oi, retaddr);
>>>>>              }
>>>>>              iotlbentry = &env->iotlb[mmu_idx][index];
>>>>>
>>>>> @@ -564,23 +603,8 @@ void helper_be_st_name(CPUArchState *env, 
>>>>> target_ulong addr, DATA_TYPE val,
>>>>>      if (DATA_SIZE > 1
>>>>>          && unlikely((addr & ~TARGET_PAGE_MASK) + DATA_SIZE - 1
>>>>>                       >= TARGET_PAGE_SIZE)) {
>>>>> -        int i;
>>>>> -    do_unaligned_access:
>>>>> -        if ((get_memop(oi) & MO_AMASK) == MO_ALIGN) {
>>>>> -            cpu_unaligned_access(ENV_GET_CPU(env), addr, MMU_DATA_STORE,
>>>>> -                                 mmu_idx, retaddr);
>>>>> -        }
>>>>> -        /* XXX: not efficient, but simple */
>>>>> -        /* Note: relies on the fact that tlb_fill() does not remove the
>>>>> -         * previous page from the TLB cache.  */
>>>>> -        for (i = DATA_SIZE - 1; i >= 0; i--) {
>>>>> -            /* Big-endian extract.  */
>>>>> -            uint8_t val8 = val >> (((DATA_SIZE - 1) * 8) - (i * 8));
>>>>> -            /* Note the adjustment at the beginning of the function.
>>>>> -               Undo that for the recursion.  */
>>>>> -            glue(helper_ret_stb, MMUSUFFIX)(env, addr + i, val8,
>>>>> -                                            oi, retaddr + GETPC_ADJ);
>>>>> -        }
>>>>> +        glue(helper_be_st_name, _do_unl_access)(env, val, addr, oi, 
>>>>> mmu_idx,
>>>>> +                                                retaddr);
>>>>>          return;
>>>>>      }
>>>>
>>>>
>>>> --
>>>> Alex Bennée
>>
>>
>> --
>> Alex Bennée


--
Alex Bennée