From: | Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> |
---|---|
To: | pgsql-hackers(at)lists(dot)postgresql(dot)org |
Cc: | Andres Freund <andres(at)anarazel(dot)de> |
Subject: | Is DLIST_STATIC_INIT() a net loss? |
Date: | 2018-12-13 17:35:15 |
Message-ID: | 30244.1544722515@sss.pgh.pa.us |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-hackers |
I happened to notice today that the initializer macro for dlist_head
variables is
#define DLIST_STATIC_INIT(name) {{&(name).head, &(name).head}}
However, all the functions that work with dlists are prepared to handle
a dlist_head that starts out as zeroes, so that this could also be
#define DLIST_STATIC_INIT(name) {{NULL, NULL}}
I submit that we'd be better off with the latter. The number of cycles
that the linker and loader expend on getting those non-constant values
correctly set up (especially in PIE builds) probably dwarf what it
costs for the first dlist access to initialize them. It's especially
obviously a loss in processes that never touch the particular dlist
at all.
Another thought is that maybe we should deprecate the use of the
[DS]LIST_STATIC_INIT macros altogether, and just write
static dlist_header myheader;
leaving the compiler to drop such variables into a BSS area instead
of an initialized-data area. I'm not very sure how much that saves,
but I bet it saves something.
regards, tom lane
From | Date | Subject | |
---|---|---|---|
Next Message | David Steele | 2018-12-13 17:48:36 | Re: Add timeline to partial WAL segments |
Previous Message | Andrey Borodin | 2018-12-13 17:06:27 | Re: Connections hang indefinitely while taking a gin index's LWLock buffer_content lock |