Re: Performance patch for Win32

I was imagining that this would be a trap for linux developers
who saw nothing wrong with their code until it made it to the
build/test farm.  That's pretty far down the development
process.  Of course, it is also a trap in the other direction, for
Windows developers who use the pattern but do not include
anything equivalent for the non-Windows execution path.

On the whole, however, your argument in favor of tighter
patterns might be more convincing than my argument in favor
of catching bugs sooner.

I will start implementing your suggestion for patch v2.

________________________________
From: Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>
To: Mark Dilger <markdilger(at)yahoo(dot)com>
Cc: "pgsql-hackers(at)postgresql(dot)org" <pgsql-hackers(at)postgresql(dot)org>
Sent: Tuesday, May 29, 2012 3:42 PM
Subject: Re: [HACKERS] Performance patch for Win32

Mark Dilger <markdilger(at)yahoo(dot)com> writes:
> I am hesitant to write a function like AllocateDirWithFilePattern
> if the pattern is simply ignored on non-Windows.  In my patch,
> the pattern underspecified the files, and the ad-hoc matching code
> applied to all the returned files tightened that up.  But a person
> could just as well overspecify the pattern and then they would get
> different behavior on Windows vs. non-Windows, with fewer
> files returned by FindNextFile() than would have matched the
> ad-hoc pattern.

Well, if you're imagining that we wouldn't need to test carefully on
both Windows and non-Windows, I think that's a pipe dream.  As an
example, your proposal of AllocateDirWithFilePrefix would only work
consistently across platforms if the prefix didn't contain anything
that Windows thought was a pattern metacharacter.  (This might never
come up, but I'm not too sure what the metacharacters are on Windows.)

Having said that, I have nothing particularly against the idea of
specifying a prefix rather than an arbitrary pattern.  I'm just
saying it'll still need testing.  Also, I wonder how many of the
potential stat-equivalent operations we'll be unable to optimize
away with the more restricted definition.  Using a tighter pattern
on Windows seems basically free (modulo testing) if we accept that
it's Windows-only.

            regards, tom lane