Comments on: S-expressions and variant

By: Joel de Guzman

Joel de Guzman — Tue, 28 Feb 2012 00:32:22 +0000

That is the common perception alright, but i’m not so sure about that for this case. memset has to deal with off sizes. Anyway, if I really wanted blazing speed, I’d unroll the loop (since the size is static anyway), but that is not the main intent.

By: Matthias vallentin

Matthias vallentin — Tue, 28 Feb 2012 00:25:23 +0000

Thanks, I understand (1) and now (2a). For (2b), why didn’t you use memset? My understanding is that is the most efficient way to initialize memory with a constant value.

By: Joel de Guzman

Joel de Guzman — Mon, 27 Feb 2012 23:45:59 +0000

Mathias, its purpose is to make sure that 1) we initialize the string to zero 2) that we have a size that’s divisible by sizeof(long). The latter has the advantage of a) optimizing the buffer usage since machine architectures tend to align the structs/unions anyway, and b) having to zero the string in size/sizeof(long) operations instead of size operations.

By: Matthias vallentin

Matthias vallentin — Mon, 27 Feb 2012 20:00:26 +0000

Would you mind explaining why you need the lbuff field in the fast_string union? I see that it has the size buff_size / (sizeof(long) / sizeof(char)); why that size specifically?

You seem to use during utree initialization before assigning a value. My first thought was that its purpose is to zero out the memory, but then why would you need an extra union field?

By: Joel de Guzman

Joel de Guzman — Mon, 30 May 2011 06:44:10 +0000

1) size constraints. std::list is 16 bytes last I checked (32 bit compiler) 2) I needed to keep the elements in the variant PODs to simplify the implementation. It’s a lot easier this way than having to rewrite variant for our very tuned and specific application.

By: Olaf

Olaf — Mon, 30 May 2011 06:17:00 +0000

Only for interest: Why do you wrote your own list implementation and don’t use std::list? Why is there no allocator interface like the std::list has which would allow to use pools, e.g. boost’s fast_pool_allocator? Would be any benefit from them?

By: Joel de Guzman

Joel de Guzman — Tue, 17 May 2011 09:16:26 +0000

Optionals are just T | nil
utree is inherently recursive

By: Olaf

Olaf — Tue, 17 May 2011 05:32:23 +0000

How can I handle optional, even rekursive??

By: Xavi Gratal

Xavi Gratal — Fri, 22 Oct 2010 18:32:43 +0000

That looks good. I will certainly have a use for this.
But I have a concern… isn’t storing members of a union as one type and then reading them as another forbidden by the standard. IIRC it breaks the strict-aliasing rule, and it can cause problems with certain compilers that do optimization taking advantage of the aliasing rules.
It would probably be better to store it as an unsigned char array and then cast as needed (casting to and from char pointers is allowed by the standard).

By: Dainis

Dainis — Thu, 16 Sep 2010 01:33:32 +0000

Thanks Hartmut, that works, just now I understand that if
utree is constructed as list , only makes a sense to construct it as list of
of utrees