dmd codegen improvements

[deadalnix via Digitalmars-d]

On Thursday, 3 September 2015 at 21:08:51 UTC, Ola Fosheim 
Grostad wrote:
> On Thursday, 3 September 2015 at 10:04:58 UTC, deadalnix wrote:
>> On Thursday, 3 September 2015 at 09:56:55 UTC, Ola Fosheim 
>> Grøstad wrote:
>>> On Thursday, 3 September 2015 at 06:18:54 UTC, deadalnix 
>>> wrote:
>>>> It is twice as slow as native. That's far from allowing 
>>>> generation of pure assembly.
>>>
>>> It is translatable to pure assembly, addressing is modulo 
>>> heap size. Performance is a different issue since it does not 
>>> provide SIMD yet.
>>
>> SIMD is not even remotely close to explaining the perf 
>> difference.
>
> What browser? Only FF supports it. Chrome just JIT it IIRC.

asm.js typically runs half the speed of natively compiled code. 
pNaCl run about 20% slower typically.

The gap is way to big for vectorization to be a reasonable 
explanation. In fact a large body of code just do not vectorize 
at all.

You seems to be fixated on that vectorization thing, when it is 
not even remotely close to the problem at hand.