Fix broadcast() corrupting upper 64 bits for IPv4 addresses

[rustyconover]

Summary

• The Broadcast() method uses ~Netmask().address (bitwise NOT on 128-bit uhugeint_t) to compute the host mask. For IPv4, the netmask has upper=0, so ~0 sets the upper 64 bits to 0xFFFFFFFFFFFFFFFF. This garbage propagates into the broadcast result.
• Observable effects:
  • broadcast('192.168.1.0/24') = '192.168.1.255/24' → false (should be true)
  • broadcast('192.168.1.0/24') > '0.0.0.1' → false (should be true)
  • broadcast('10.0.0.0/24') + 1 → throws spurious overflow error
  • Raw stored address is negative instead of the correct positive value
• The VARCHAR display happened to be correct since ToStringIPv4 only reads the lower 32 bits, masking the bug from casual observation.
• Fix: compute the host mask as full_mask ^ netmask instead of ~netmask, where full_mask is the address-space mask (0xFFFFFFFF for IPv4, all-ones for IPv6). For IPv6 this is equivalent since all_ones ^ x == ~x.

Reproduction

LOAD inet;

-- Equality is broken
SELECT broadcast(INET '192.168.1.5/24') = INET '192.168.1.255/24';
-- Expected: true
-- Actual:   false

-- Ordering is broken
SELECT broadcast(INET '192.168.1.5/24') > INET '0.0.0.1';
-- Expected: true
-- Actual:   false

-- Raw address shows the corruption
SELECT broadcast('192.168.1.0/24'::INET)::STRUCT(ip_type UTINYINT, address HUGEINT, mask USMALLINT);
-- Shows address = -1062731265 (should be 3232236031)

Test plan

• Added IPv4 broadcast equality test: broadcast(... /24) = '192.168.1.255/24'
• Added IPv4 broadcast ordering test: broadcast(... /24) > '0.0.0.1'
• Added IPv4 broadcast /0 test
• Added IPv6 broadcast equality test (regression check)
• Existing broadcast display tests still pass

🤖 Generated with Claude Code