Re: [PATCH v3 09/20] cputlb: Replace switches in load/store_helper with callback

[David Hildenbrand]

On 23.09.19 11:51, Paolo Bonzini wrote:
> On 22/09/19 05:54, Richard Henderson wrote:
>> +/* Wrap the unaligned load helpers to that they have a common signature.  */
>> +static inline uint64_t wrap_ldub(const void *haddr)
>> +{
>> +    return ldub_p(haddr);
>> +}
>> +
>> +static inline uint64_t wrap_lduw_be(const void *haddr)
>> +{
>> +    return lduw_be_p(haddr);
>> +}
>> +
>> +static inline uint64_t wrap_lduw_le(const void *haddr)
>> +{
>> +    return lduw_le_p(haddr);
>> +}
>> +
>> +static inline uint64_t wrap_ldul_be(const void *haddr)
>> +{
>> +    return (uint32_t)ldl_be_p(haddr);
>> +}
>> +
>> +static inline uint64_t wrap_ldul_le(const void *haddr)
>> +{
>> +    return (uint32_t)ldl_le_p(haddr);
>> +}
> 
> Any reason to have these five functions (plus five for stores) instead
> of a pair
> 
> static uint64_t ld_memop(const void *haddr, MemOp op)
> {
> }
> 
> static uint64_t st_memop(void *haddr, MemOp op, uint64_t val)
> {
> }
> 
> that includes the switches?  Everything should be inlined just the same
> if you do
> 
>         if (unlikely(tlb_addr & TLB_BSWAP)) {
>             st_memop(haddr, op ^ MO_BSWAP, val);
>         } else {
>             st_memop(haddr, op, val);
>         }

I asked the same question on v2 and Richard explained that - for
whatever reason -  the compiler will not properly propagate the constant
in the "op ^ MO_BSWAP" case.

> 
> and the like.
> 
> Paolo


-- 

Thanks,

David / dhildenb