Re: [boost] Proposed templated integer_sort

Subject: Re: [boost] Proposed templated integer_sort
From: Scott McMurray (me22.ca+boost_at_[hidden])
Date: 2008-12-16 10:34:39

On Tue, Dec 16, 2008 at 10:19, Giovanni Piero Deretta
<gpderetta_at_[hidden]> wrote:
> On Tue, Dec 16, 2008 at 3:03 PM, Pyry Jahkola <pyry.jahkola_at_[hidden]> wrote:
>>I'd
>> suggest using a union instead, for the purpose of treating a float as
>> integer:
>>
>>     template <typename To, typename From>
>>     inline To bitwise_cast(From const & from) {
>>         BOOST_STATIC_ASSERT(sizeof(To) == sizeof(From));
>>         union { To to; From from; } u;
>>         u.from = from;
>>         return u.to;
>>     }
>>
>
> This is also undefined behavior even if it is explicitly supported by
> most compilers as an extension. It is not necessarily better than a
> memcpy and might actually be slower.
>

Empyrically, putting the following into the LLVM online demo
(http://llvm.org/demo/) shows that, in llvm-g++ at least, the
(intermediate) code generated from the memcpy is optimal:

#include <string.h>
int foo(float const y) {
    int x;
    memcpy(&x,&y,sizeof(int));
    return x;
}

define i32 @_Z3foof(float %y) nounwind readnone {
entry:
        %0 = bitcast float %y to i32		; <i32> [#uses=1]
        ret i32 %0
}

So since the memcpy is always safe and doesn't break the aliasing
rules, it certainly seems the best approach.