Обсуждение: semi-variable length type

I want to make a float(n) type that emulates the mssql float type. The storage requirements are documented as 4 bytes
for1 <= n <=24, and 8 bytes for 25 <= n <= 53. If I understand correctly, my options for emulating this in postgres
are:

1. declare as variable length. Storage is then 8 bytes (4 byte length + 4 byte storage), or 12 bytes (4 byte length + 8
bytestorage). 

2. declare as fixed 8 byte length.

In the absence of some magical 3rd option, I assume #2 is the best way forward... can anyone confirm?

Thanks

James

James Harper <james.harper@bendigoit.com.au> writes:
> I want to make a float(n) type that emulates the mssql float type. The
> storage requirements are documented as 4 bytes for 1 <= n <=24, and 8
> bytes for 25 <= n <= 53.

Haven't we got that already?

regression=# create table t1 (f1 float(5), f2 float(30));
CREATE TABLE
regression=# \d t1
           Table "public.t1"
 Column |       Type       | Modifiers
--------+------------------+-----------
 f1     | real             |
 f2     | double precision |

Other than the fact that we don't remember whether you asked for 5 bits
or 24, I think this meets the spec requirements.

> If I understand correctly, my options for
> emulating this in postgres are: 1. declare as variable length. Storage
> is then 8 bytes (4 byte length + 4 byte storage), or 12 bytes (4 byte
> length + 8 byte storage).

Well, you could use a short varlena header (1 byte).  The type doesn't
need to be int-aligned, either, though whether that buys anything will
depend on context --- and you'll need to spend cycles realigning it,
if you want the code to work on non-Intel architectures.

            regards, tom lane

>
> James Harper <james.harper@bendigoit.com.au> writes:
> > I want to make a float(n) type that emulates the mssql float type. The
> > storage requirements are documented as 4 bytes for 1 <= n <=24, and 8
> > bytes for 25 <= n <= 53.
>
> Haven't we got that already?
>
> regression=# create table t1 (f1 float(5), f2 float(30));
> CREATE TABLE
> regression=# \d t1
>            Table "public.t1"
>  Column |       Type       | Modifiers
> --------+------------------+-----------
>  f1     | real             |
>  f2     | double precision |
>
> Other than the fact that we don't remember whether you asked for 5 bits
> or 24, I think this meets the spec requirements.
>

Is the 4 byte or 8 byte decision based on (n) implemented in the parser? I can't see a 'float' type in pg_type.

Thanks

James

James Harper <james.harper@bendigoit.com.au> writes:
>> regression=# create table t1 (f1 float(5), f2 float(30));
>> CREATE TABLE
>> regression=# \d t1
>> Table "public.t1"
>> Column |       Type       | Modifiers
>> --------+------------------+-----------
>> f1     | real             |
>> f2     | double precision |
>>
>> Other than the fact that we don't remember whether you asked for 5 bits
>> or 24, I think this meets the spec requirements.

> Is the 4 byte or 8 byte decision based on (n) implemented in the parser? I can't see a 'float' type in pg_type.

Yeah, gram.y is hard-wired to produce either "float4" or "float8" from
float(N) (see the opt_float production).

Ordinarily I'd think that that was a pretty sucky implementation
technique, but since this is implementing behavior called for in
the SQL standard, it's probably not worth trying to make it
any more flexible.

            regards, tom lane