On 2015-12-17 15:51, Andrey Semashev wrote: > On 2015-12-17 15:33, Pali Rohár wrote: >> On Thursday 17 December 2015 15:10:12 Andrey Semashev wrote: >>> On 2015-12-17 14:50, Andrey Semashev wrote: >>>> On 2015-12-17 02:44, Kamil Rytarowski wrote: >>>>> >>>>> On 16.12.2015 11:02, Pali Rohár wrote: >>>>>> >>>>>> Hi! There is probably no portable way for checking CPU type. But >>>>>> you do not need to know CPU type here. You need to know if CPU has >>>>>> fast 64bit numeric operations (and use 64bit speed optimization), >>>>>> right? >>>>>> >>>>>> And this could be possible by checking size of int_fast32_t type >>>>>> (defined in stdint.h). If is same as size of int_fast64_t type then >>>>>> you could use 64bit optimization otherwise not (= 32bit integers >>>>>> are faster then 64bit). >>>>>> >>>>>> But I will let this suggestion for reviewing of other people. >>>>> >>>>> This sounds like a good way to go. >>>> >>>> I don't think it does. int_fast32_t does not have to be 64-bit if >>>> 32-bit >>>> integers are fast. >> >> Yes. I did not wrote that int_fast32_t must be 64-bit. I wrote to >> compare size of int_fast32_t and int_fast64_t. > > And they will be different on x32 which, following your logic, means > that on x32 32-bit integers are faster than 64-bit. That's not true. > >>>> It _may_ be larger than 32-bit if 32-bit integers are >>>> slow (most likely - do not have the target hardware support). I don't >>>> think this is the case for x32 or x86_64. >> >> On 64bi x86 sizeof(int_fast32_t) is 8 and sizeof(int_fast64_t) also 8. On >> 32bit x86 sizeof(int_fast32_t) is only 4. > > On x86_64, yes: > > gcc -m64 -dM -E - < /dev/null | grep FAST32 > #define __INT_FAST32_MAX__ 0x7fffffffffffffffL > #define __INT_FAST32_TYPE__ long int > #define __UINT_FAST32_MAX__ 0xffffffffffffffffUL > #define __UINT_FAST32_TYPE__ long unsigned int > > On x32, no: > > gcc -mx32 -dM -E - < /dev/null | grep FAST32 > #define __INT_FAST32_MAX__ 0x7fffffff > #define __INT_FAST32_TYPE__ int > #define __UINT_FAST32_MAX__ 0xffffffffU > #define __UINT_FAST32_TYPE__ unsigned int > >>>> If integer performance is what this preprocessor check influences, I'd >>>> suggest checking the size of size_t. Or testing for particular >>>> architectures explicitly, with 64-bit being the default and 32-bit only >>>> for the tested architectures (this would probably make the code more >>>> easily maintainable in the future). >>> >>> I've sent this privately in error. >>> >>> Also, it seems size_t is 32-bit on x32, so disregard that suggestion. >> >> Important question is: what you want to distinguish? It is not 32 vs 64 >> bit processor right? It is if architecture has 64bit long registers? Or >> something else? > > The bitness of the architecture is a rather fuzzy term, but yes, it is > important to determine what for this check is needed. From the > discussion I gathered that the goal is to determine the largest fast > integer type for math. Damn, I keep pressing Reply instead of Reply to all... Sorry again.
