[PATCH] erofs-utils: lib: converted division to shift in z_erofs_load_compact_lcluster

[Gao Xiang]

Hi Ashley,

On 2026/2/25 00:30, Ashley Lee wrote:
> perf on fsck.erofs reports that z_erofs_load_compact_lcluster was
> spending 20% of its time doing the div instruction. While the function
> itself is ~40% of the fsck.erofs runtime. In the source code, it seems
> that the compiler can't optimize the division by vcnt despite it only
> holding powers of two.
> 
> Running a benchmark on a lzma compressed freebsd source tree
> on x86 yields a ~3% increase in performance. The following
> test was run locally on an x86 machine.
> 
> $ hyperfine -w 10 -p "echo 3 > /proc/sys/vm/drop_caches; sleep 1" \
>    "./fsck.erofs ./bsd.erofs.lzma"
> 
> With shift optimization
> Time (mean ± σ):     360.0 ms ±  12.0 ms    \
>    [User: 236.3 ms, System: 120.6 ms]
> Range (min … max):   342.3 ms … 379.8 ms    10 runs
> 
> Original Dev Branch
> Time (mean ± σ):     371.1 ms ±  16.1 ms    \
>    [User: 254.8 ms, System: 115.0 ms]
> Range (min … max):   354.8 ms … 404.4 ms    10 runs

Which compiler are you using? Does both gcc and clang
behave the same?

I wonder if it's possible to refactor this a bit instead
of introducing another `vdiv`, also if the kernel
codebase behaves the same, we should improve the kernel too.

Thanks,
Gao Xiang

> 
> Signed-off-by: Ashley Lee <yester1324 at gmail.com>
> 
> ---
>   lib/zmap.c | 15 +++++++++------
>   1 file changed, 9 insertions(+), 6 deletions(-)
> 
> diff --git a/lib/zmap.c b/lib/zmap.c
> index baec278..1ba52b5 100644
> --- a/lib/zmap.c
> +++ b/lib/zmap.c
> @@ -112,7 +112,7 @@ static int z_erofs_load_compact_lcluster(struct z_erofs_maprecorder *m,
>   	const unsigned int lclusterbits = vi->z_lclusterbits;
>   	const unsigned int totalidx = BLK_ROUND_UP(sbi, vi->i_size);
>   	unsigned int compacted_4b_initial, compacted_2b, amortizedshift;
> -	unsigned int vcnt, lo, lobits, encodebits, nblk, bytes;
> +	unsigned int vcnt, vdiv, lo, lobits, encodebits, nblk, bytes;
>   	bool big_pcluster = vi->z_advise & Z_EROFS_ADVISE_BIG_PCLUSTER_1;
>   	erofs_off_t pos;
>   	u8 *in, type;
> @@ -144,13 +144,16 @@ static int z_erofs_load_compact_lcluster(struct z_erofs_maprecorder *m,
>   	pos += lcn * (1 << amortizedshift);
>   
>   	/* figure out the lcluster count in this pack */
> -	if (1 << amortizedshift == 4 && lclusterbits <= 14)
> +	if (1 << amortizedshift == 4 && lclusterbits <= 14) {
>   		vcnt = 2;
> -	else if (1 << amortizedshift == 2 && lclusterbits <= 12)
> +		vdiv = 1;
> +	} else if (1 << amortizedshift == 2 && lclusterbits <= 12) {
>   		vcnt = 16;
> -	else
> +		vdiv = 4;
> +	} else {
>   		return -EOPNOTSUPP;
> -
> +	}
> +
>   	in = erofs_read_metabuf(&m->map->buf, sbi, pos,
>   				erofs_inode_in_metabox(vi));
>   	if (IS_ERR(in))
> @@ -160,7 +163,7 @@ static int z_erofs_load_compact_lcluster(struct z_erofs_maprecorder *m,
>   	m->nextpackoff = round_down(pos, vcnt << amortizedshift) +
>   			 (vcnt << amortizedshift);
>   	lobits = max(lclusterbits, ilog2(Z_EROFS_LI_D0_CBLKCNT) + 1U);
> -	encodebits = ((vcnt << amortizedshift) - sizeof(__le32)) * 8 / vcnt;
> +	encodebits = (((vcnt << amortizedshift) - sizeof(__le32)) * 8) >> vdiv;
>   	bytes = pos & ((vcnt << amortizedshift) - 1);
>   	in -= bytes;
>   	i = bytes >> amortizedshift;