source: branches/1.1.x/doc/design/pathnames/merging-defaults.markdown

# ISSUE MERGE-PATHNAMES with specialization of JAR-PATHNAME 

We wish to resolve the following issues.

## UC0 Loading jna.jar for CFFI via Quicklisp

Currently we cannount load systems via ASDF recursively in jars, most
importantly for those packaged in 'abcl-contrib.jar', which prevents
CFFI from loading the necessary JNA code.

## UC1.1
If *DEFAULT-PATHNAME-DEFAULTS* is a JAR-PATHNAME, commands that would
normally be expected to work such as 

    CL-USER> (probe-file #p"/") 
    
will fail.  

### UC1.2

If *DEFAULT-PATHNAME-DEFAULTS* is a JAR-PATHNAME, COMPILE-FILE
operations specifying an OUTPUT-FILE with a NIL DEVICE will fail, as
COMPILE-FILE-PATHNAME is required to merge its arguments with the
defaults.

## CLHS Citations

Some especially relevant portions of the CLHS for consideration.

###  19.2.3 Merging Pathnames
http://www.lispworks.com/documentation/HyperSpec/Body/19_bc.htm

"Except as explicitly specified otherwise, for functions that
manipulate or inquire about files in the file system, the pathname
argument to such a function is merged with *default-pathname-defaults*
before accessing the file system (as if by merge-pathnames)."

Note that this implies that the arguments to PARSE-NAMESTRING--which
is what the SHARSIGN-P reader macro correpsponds to--should not be
merged.


## 19.2.2.2.3 :UNSPECIFIC as a Component Value
http://www.lispworks.com/documentation/HyperSpec/Body/19_bbbc.htm

"If :unspecific is the value of a pathname component, the component is
considered to be ``absent'' or to ``have no meaning'' in the filename
being represented by the pathname."

Having an :UNSPECIFIC DEVICE when a PATHNAME refers to a file would
address problems when merging when the defaults contains a JAR-PATHNAME.

### MERGE-PATHNAMES 
http://www.lispworks.com/documentation/HyperSpec/Body/f_merge_.htm

"If pathname explicitly specifies a host and not a device, and if the
host component of default-pathname matches the host component of
pathname, then the device is taken from the default-pathname;
otherwise the device will be the default file device for that host. If
pathname does not specify a host, device, directory, name, or type,
each such component is copied from default-pathname."

This suggests that the contents HOST should be considered as an
additional axis of type for PATHNAME, so that when PATHNAMES of
differing types get merged, a DEVICE which has no meaning for a given
type does not get inserted. 

## Other implementations

A survey of the the "default" host for #p"/" on startup.

### SBCL

A host nonce which appears in the reader as
#<SB-IMPL::UNIX-HOST {1000290443}>.  (Is there a different one under
Windows?)

### CLISP

HOST is NIL.

### CCL

HOST is :UNSPECIFIC. 

### ECL

HOST is NIL.

## Implementation

Since Windows systems do have a default DEVICE for a normal file
PATHNAME, namely the current "drive letter" of the process, the
implementation changes will be mostly wrapped in runtime conditionals
for non-Windows systems.

### TRUENAME sets DEVICE to :UNSPECIFIC

TRUENAME sets DEVICE to :UNSPECIFIC running on non-Windows when
resolving a path to a plain file.

### Use an implicit type for merging 

In the case for which a merge occurs when DEFAULT-PATHNAME
is a JAR-PATHNAME and the following conditions hold:

1.  HOST and DEVICE of the PATHNAME are NIL

2.  The DIRECTORY of the PATHNAME represents an absolute path. 

3.  We are not on Windows.

we set the DEVICE to be :UNSPECIFIC.

### COLOPHON

Mark <evenson@panix.com>
Created:  01-SEP-2012
Revised:  09-OCT-2012