> On Feb 25, 2026, at 21:10, Daniel Gustafsson <[email protected]> wrote:
> 
>> On 25 Feb 2026, at 13:41, Chao Li <[email protected]> wrote:
>> 
>>> On Feb 25, 2026, at 18:21, Daniel Gustafsson <[email protected]> wrote:
>>> 
>>>> On 25 Feb 2026, at 07:31, yangyz <[email protected]> wrote:
>>> 
>>>> 2.Performance Overhead
>>>> In each iteration, the entire buffer of size LZ4_CHUNK_SZ (potentially 
>>>> several megabytes) is zero-initialized. Since these memory blocks are 
>>>> immediately overwritten by decompressed data, this zeroing operation 
>>>> constitutes an unnecessary consumption of CPU resources.
>>> 
>>> When proposing a performance improvement it's important to provide some 
>>> level
>>> of benchmarks to show the improvement. Is removing this memset noticeable?
>> 
>> I don’t think this patch is about performance. Although removing the memset 
>> might save a few CPU cycles, the real benefit seems to be cleanup and 
>> consistency. The memset appears unnecessary, and similar functions don’t use 
>> it, so I think this change mainly improves maintainability.
> 
> I would argue the opposite, clearing a buffer before passing it to an external
> library function writing to it seems the right thing to do unless it can be
> proven to regress performance too much.  Also, "appears unnecessary" doesn't
> instill enough confidence to perform a change IMO.
> 
> --
> Daniel Gustafsson


As I pointed out earlier, ReadDataFromArchiveLZ4() has a very similar loop that 
doesn’t zero out  the output buffer:
```
while (readp < readend)
{
    size_t out_size = DEFAULT_IO_BUFFER_SIZE;
    size_t read_size = readend - readp;

    status = LZ4F_decompress(ctx, outbuf, &out_size,
                                                 readp, &read_size, &dec_opt);
    if (LZ4F_isError(status))
        pg_fatal("could not decompress: %s",
              LZ4F_getErrorName(status));

    ahwrite(outbuf, 1, out_size, AH);
    readp += read_size;
}
```

Do you think we should add a memset there? There are a couple of more callers 
of LZ4F_decompress that don’t zero out the output buffer.

Best regards,
--
Chao Li (Evan)
HighGo Software Co., Ltd.
https://www.highgo.com/






Reply via email to