Write sint32 values as varints of at most 5 bytes

[merlimat]

Problem

For a sint32 field, getSerializedSize() counts the zigzag-encoded value as an unsigned 32-bit varint, at most 5 bytes, as protobuf does. The writer, however, went through writeRawVarInt(), which is the int32 writer: it encodes a negative int as a 10-byte varint.

The zigzag value is negative for every v >= 2^30 and every v < -2^30. For each such value, writeTo() wrote 5 bytes more than it had reserved. This happens on freshly built messages, with no parsing involved:

• toByteArray() throws ArrayIndexOutOfBoundsException;
• writeTo() into a pooled heap buffer writes the extra bytes into the next buffer of the same chunk, because pooled heap buffers share one backing array.

Singular, repeated, packed and map sint32 fields are all affected.

Example

Numbers n = new Numbers().setXSint32(1 << 30);
n.getSerializedSize();   // 6
n.toByteArray();         // ArrayIndexOutOfBoundsException: 11 bytes written into 6

The current writer:

static int writeRawSignedVarInt(byte[] a, int i, int n) {
    return writeRawVarInt(a, i, encodeZigZag32(n));   // int32 writer: a negative n takes 10 bytes
}

Change

A new writeRawVarUInt() writes an unsigned 32-bit varint of at most 5 bytes, for both the byte[] and the NIO sink. writeRawSignedVarInt() uses it:

static int writeRawVarUInt(byte[] a, int i, int n) {
    while (true) {
        if ((n & ~0x7F) == 0) {
            a[i++] = (byte) n;
            return i;
        }
        a[i++] = (byte) ((n & 0x7F) | 0x80);
        n >>>= 7;
    }
}

static int writeRawSignedVarInt(byte[] a, int i, int n) {
    return writeRawVarUInt(a, i, encodeZigZag32(n));
}

Writes now match computeSignedVarIntSize(), and the output matches protobuf-java.

Testing

NumbersTest and RepeatedNumbersTest cover the values whose zigzag value has its top bit set: 2^30, -2^30 - 1, Integer.MAX_VALUE and Integer.MIN_VALUE. For singular, repeated and packed fields, sizes and bytes match protobuf-java. These tests fail on master. Full mvn verify passes.