GitHub user LouisLou2 closed a discussion: Doubts in FuryJava's
MemoryBuffer::readBytesAsInt64
```java
/** Read {@code len} bytes into a long using little-endian order. */
public long readBytesAsInt64(int len) {
int readerIdx = readerIndex;
// use subtract to avoid overflow
int remaining = size - readerIdx;
if (remaining >= 8) {
readerIndex = readerIdx + len;
long v =
UNSAFE.getLong(heapMemory, address + readerIdx)
& (0xffffffffffffffffL >>> ((8 - len) * 8));
return LITTLE_ENDIAN ? v : Long.reverseBytes(v);
}
return slowReadBytesAsInt64(remaining, len);
}
```
I have some doubts about this method above when applying big endian.
### Problem Analysis
1. **Incorrect Mask Calculation**
- The issue arises when `len < 8`. The current mask calculation:
```java
0xffffffffffffffffL >>> ((8 - len) * 8)
```
keeps the **lower `len * 8` bits**, but in the **Big Endian** scenario,
the valid data may actually reside in the **higher `len * 8` bits** (the upper
bits).
**Example:**
When `len = 3`, the mask calculation becomes:
```java
0xffffffffffffffffL >>> 40
```
which results in the mask `0x0000000000ffffff`, effectively keeping the
**lower 24 bits**. However, if the data is stored in **Big Endian**, the valid
data would actually reside in the **upper 24 bits** (the first 3 bytes). As a
result, this incorrect mask clears the valid data.
2. **Incorrect Byte Order Handling**
- The byte order reversal operation should happen **before** applying the
mask. Currently, the reversal happens after the mask is applied:
```java
return LITTLE_ENDIAN ? v : Long.reverseBytes(v);
```
This is problematic because, after the byte order is reversed, the
high-order bits may get erroneously cleared due to the mask being applied in
the wrong order. Reversing the byte order after masking results in the
potential loss of the valid high-order bits.
GitHub link: https://github.com/apache/fory/discussions/2067
----
This is an automatically sent email for [email protected].
To unsubscribe, please send an email to: [email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
