Better zz_t

[skirpichev]

• hide details like alloc field?
• use signed size (for sign)?
• flags for static allocation

Keep in mind also #3.

[skirpichev]

Probably we want something like

typedef struct {
    int64_t _reserved;
    zz_size_t size;
    zz_digit_t *digits;
} zz_private_t;

where private memory-compatible structure will be either:

typedef struct {
    zz_size_t alloc;
    zz_size_t size;
    zz_digit_t *digits;
} zz_private_t;    

or

typedef struct {
    struct {
        uint8_t negative : 1;
        int64_t alloc : 63;
    };
    zz_size_t size;
    zz_digit_t *digits;
} zz_private_t;

The first version looks like the mpz_t, but sign will be extracted from the alloc field. Second version seems to be faster, but still slower than the current structure with negative boolean field.

For bench/collatz.py:

Benchmark	ref	signed	bit-fields
collatz0(97)	83.5 us	89.3 us: 1.07x slower	84.9 us: 1.02x slower
collatz0(871)	125 us	134 us: 1.07x slower	127 us: 1.02x slower
collatz0((1<<128)+31)	681 us	716 us: 1.05x slower	not significant
collatz1(97)	98.5 us	102 us: 1.03x slower	101 us: 1.02x slower
collatz1(871)	148 us	152 us: 1.03x slower	153 us: 1.03x slower
collatz1((1<<128)+31)	795 us	812 us: 1.02x slower	817 us: 1.03x slower
Geometric mean	(ref)	1.04x slower	1.02x slower

We could also consider to use int32_t for size and alloc fields (like GMP/FLINT). With this, a number with the maximal magnitude will require ~16GiB memory.