mirror of
https://github.com/libevent/libevent.git
synced 2025-01-31 09:12:55 +08:00
1af8017648
The sizeof `struct event` can reduced on both 32 bit and 64 bit systems
by moving the 4 bytes that make up `ev_events` and `ev_res` below `ev_fd`,
before `struct event_base * ev_base;` since our compiler wouldn't dare do
such a thing (it instead will pad twice, whereas it only needs to be padded
once)
```C
struct event {
/* OFFS | SZ Bytes | Total Bytes | START - END */
struct event_callback ev_evcallback; /* 0x0 | 40 | 40 | 0x0 - 0x28 */
union { /* 0x28 | ----------- | ----------- | ------------ */
TAILQ_ENTRY(event) ev_next_with_common_timeout; /* | ((16)) | | */
int min_heap_idx; /* | ((04)) | | */
} ev_timeout_pos; /* | 16 | 56 | 0x28 - 0x38 */
int ev_fd; /* 0x38 | 04 | 60 | 0x38 - 0x3c */
```
Since the next field is 8 bytes in length, and we are up to 60 bytes, `ev_fd` ends up being padded (4 more bytes on 64b).
```C
/* --- 1 byte gap HERE ---> 1 | <61> */
/* --- 1 byte gap HERE ---> 1 | <62> */
/* --- 1 byte gap HERE ---> 1 | <63> */
/* --- 1 byte gap HERE ---> 1 | <64> */
struct event_base * ev_base; /* 0x3c | 8 | 68 | 0x3c - 0x40 */
union { /* 0x40 | ------------ | ---------- | ------------ */
struct { /* | ------------ | | */
LIST_ENTRY (event) ev_io_next; /* | ((16+ | | */
struct timeval ev_timeout; /* | 16)) | | */
} ev_io; /* | ((32)) | | */
struct { /* | ------------ | | */
LIST_ENTRY (event) ev_signal_next; /* | ((16+ | | */
short ev_ncalls; /* | 02+ | | */
short * ev_pncalls; /* | 08)) | | */
} ev_signal; /* | ((26)) | | */
} ev_; /* 0x60 | 32 | 100 | 0x40 - 0x60 */
short ev_events; /* 0x60 | 2 | 102 | 0x60 - 0x62 */
short ev_res; /* 0x62 | 2 | 104 | 0x62 - 0x64 */
```
We now hit another line, `struct timeval` is 16 bytes on 64b arch, so we have 4 more bytes
of padding on `ev_res`.
```C
/* --- 1 byte gap HERE --- */
/* --- 1 byte gap HERE --- */
/* --- 1 byte gap HERE --- */
/* --- 1 byte gap HERE --- */
struct timeval ev_timeout; /* 0x64 | 16 | 120 | 0x64 - 0x74 */
};
```
After moving `ev_events` and `ev_res` below `ev_fd` we have something
a bit more optimal:
```C
struct event2 {
/* OFFS | SZ / Bytes | RSUM Bytes | START - END */
struct event_callback ev_evcallback; /* 0x0 | 40 | 40 | 0x0 - 0x28 */
union { /* 0x28 | ------------ | ---------- | ------------ */
TAILQ_ENTRY(event) ev_next_with_common_timeout; /* | ((16)) | | */
int min_heap_idx; /* | ((04)) | | */
} ev_timeout_pos; /* | 16 | 56 | 0x28 - 0x38 */
int ev_fd; /* 0x38 | 4 | 60 | 0x38 - 0x3c */
short ev_events; /* 0x3c | 2 | 62 | 0x3c - 0x3e */
short ev_res; /* 0x3e | 2 | 64 | 0x3e - 0x40 */
struct event_base * ev_base; /* 0x40 | 8 | 74 | 0x40 - 0x48 */
union { /* 0x48 | ------------ | ---------- | ------------ */
struct { /* | ------------ | | */
LIST_ENTRY (event) ev_io_next; /* | ((16+ | | */
struct timeval ev_timeout; /* | 16)) | | */
} ev_io; /* | ((32)) | | */
struct { /* | ------------ | | */
LIST_ENTRY (event) ev_signal_next; /* | ((16+ | | */
short ev_ncalls; /* | 02+ | | */
short * ev_pncalls; /* | 08)) | | */
} ev_signal; /* | ((26)) | | */
} ev_; /* | 32 | 106 | 0x48 - 0x68 */
struct timeval ev_timeout; /* 0x68 | 16 | 120 | 0x68 - 0x78 */
};
```
We still have a gap here, but the first was removed.
Again, we can save 8 bytes on both 32 and 64 word sizes (32/64 byte cacheline).
Below are the results for testing v2.1.6 -> master -> master + this patch (Release/-O3)
Code:
```C
#include <event2/event.h>
int
main(int argc, char ** argv) {
printf("%zu\n", event_get_struct_event_size());
return 0;
}
```
Branch: `master` (2.2.x)
```
$ gcc -O3 -Wall -Wl,-R/usr/local/lib bleh.c -L/usr/local/lib -o bleh -levent
$ ldd bleh
linux-vdso.so.1 => (0x00007ffc3df50000)
libevent.so.2.2.0 => /usr/local/lib/libevent.so.2.2.0 (0x00007f91fd781000)
libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f91fd3a1000)
libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f91fd182000)
/lib64/ld-linux-x86-64.so.2 (0x00007f91fdbcc000)
$ ./bleh
128
```
Release: `2.1.6`
```
$ gcc -O3 bleh.c -o bleh -levent
$ ldd bleh
linux-vdso.so.1 => (0x00007ffd43773000)
libevent-2.1.so.6 => /usr/lib/x86_64-linux-gnu/libevent-2.1.so.6 (0x00007feb3add6000)
libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007feb3a9f6000)
libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007feb3a7d7000)
/lib64/ld-linux-x86-64.so.2 (0x00007feb3b22a000)
$ ./bleh
128
```
Branch: `this one`
```
$ gcc -O3 -Wl,-R./lib bleh.c -o bleh -L./lib -levent
$ ldd bleh
linux-vdso.so.1 => (0x00007ffff55f7000)
libevent.so.2.2.0 => ./lib/libevent.so.2.2.0 (0x00007ff8e5c82000)
libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007ff8e58a2000)
libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007ff8e5683000)
/lib64/ld-linux-x86-64.so.2 (0x00007ff8e60cd000)
$ ./bleh
120
```