Subject: Re: [boost] [1.47.0] Permission to merge filesystem fix
From: Eric Niebler (eric_at_[hidden])
Date: 2011-07-05 11:31:47
Apologies for the delay. The US holiday distracted me.
On 7/3/2011 7:26 PM, Beman Dawes wrote:
> On Sun, Jul 3, 2011 at 9:24 PM, Eric Niebler <eric_at_[hidden]> wrote:
>> On 7/3/2011 2:48 PM, Beman Dawes wrote:
>>> This changeset:
>>> Has now cycled through many of the regression testers. Testing is
>>> still a little light on Windows, but I've tested locally on VC++ 8, 9,
>>> and 10, and late model mingw, too.
>>> This fixes a nasty POSIX case that was throwing before main() started
>>> if the environment LANG variable was whacked.
>>> OK to merge?
>> The changes in the .cpp just look like comments and formatting. Is that
>> right? And the extra path constructors ... are there cases where the new
>> behavior is different than the old?
> The two added constructors are the critical change. Since their
> argument is already of the required type and encoding, there is no
> need to access the codecvt facet. And since there is no need for the
> facet, there is also no need for constructing a std::locale("")
> The path implementation, in src/path.cpp, constructs two paths with
> the native value_type for internal use. Without the change, GCC thinks
> that the construction of these two paths can be done as static
> initialization, before main() runs. Thus if LANG is whacked, the
> resulting exception from trying to construct std::locale("") is thrown
> before mainj() starts. By adding the two constructors, that is
> avoided. It is also a small, but nice, optimization for the most
> common use case in code that runs on POSIX systems of narrow character
> paths. Not to mention avoiding the exception entirely if the program
> does no trafficking in wide character paths.
> If your eye are glazing over, don't worry. It took me three days to
> figure it out. Mostly learning to use Code::Blocks on Linux, and then
> an hour of debugging and testing to see exactly what was happening.
I'm sure it's fine if you believe there is no behavior change in any
supported use case.
Let me clarify why I was asking though, because I don't think I was
clear. The constructor arguments may already have the required type, but
you can't really know anything about their encoding because that's not
encoded in the type system. Previously, you were always running these
strings through a transformation. You're presuming (and you would know
best of course) that the transformation was always a no-op for arguments
of this type. I was asking if you were really 100% sure.
-- Eric Niebler BoostPro Computing http://www.boostpro.com
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk