Age | Commit message (Collapse) | Author |
|
This is a large change. In order to remove myself from libc's arcane
interface, I implemented an independent runtime layer. It is based on
musl's wonderful implementation mostly. Critically, if libc is linked to
the program, then we cooperate. Namely, we call start main and let libc
do all initialization. If not, then we have a noop defined in rt3.a.
The general structure of the file is:
1. sys/$os/$arch contains all architecture dependent code
2. sys/$os/port contains all code that depends on the os, but is
portable
3. rt/$arch contains all the runtime architecture dependent code
4. rt/* contains the portable runtime code.
Obviously testing is needed. Specifically, while code is checked in for
the most popular architectures, it only has been tested on one computer!
Overall this is exciting and as been educational.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
This is very much a work in progress. Still ruminating on the structure
of the library. It feels right but I want a more "social" presence -
namely the ability to link to a libc seemlessly. The solution is most
likely weak aliasing that musl uses - but musl itself weak aliases
global symbols, e.g malloc. We would have to define weak internal
symbols that musl defines as strong links but this knows too much about
the internals of musl...
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Updated our assumptions of readline to handle valid unicode input.
This required integrating with an explicit library to handle unicode
knowledge.
|
|
|
|
|
|
|
|
|
|
The readline functionality operated on the assumption that 1 byte = 1
character. This is obviously wrong if you input a non-ascii character.
This commit temporarily removes a lot of functionality but parses input
bytes in a unicode-aware manner.
The outstanding problem now is 1 unicode rune != 1 column. There are
double wide characters, as well as zero width runes, that further break
our assumption that 1 rune = 1 character = 1 column. This is the next
iteration.
|
|
I was hiding too many important constants. This commit moves them to the
main exported header.
|
|
Additionally, decode can now apply backwards on a byte string.
|
|
|
|
Refactored code to pull out utf8 functions from base into a standalone
library. Also left the required function inside arg.c so that code that
calls ARG_BEGIN doesn't have to link to libunicode.
|
|
|
|
Prototypes for loops sketched. This required recognizing keywords and
returning from yylex. Debugging/testing will be required.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Now correctly prints out argv[1]
|
|
|
|
|
|
|
|
|
|
One quick solution to the lack of tracking deep into the command line is
to note that the pattern of code emitted for an async is:
Xasync
|__ child (command)
|__ parent (continues)
The child creates a process group, as described before. If the child is
a simple command, we will now "exec" as it will exit immediately after
the command. This gives us the correct behavior, at least for simple
cases.
This also fixed pipes. However, if child has to be forked, i.e. can't be
immediately execed, then I don't think this process works...
|
|
Hit a bit of a stopping point. Specifically, the way XAsync runs
currently is by forking the execution context and having the child run
the async code while the parent runs the remainder. The problem with
this architecture is it doesn't interact well with job control. When we
fork, we create a new process group. Thus the Xasync fork becomes the
new leader.
In short, our traversal of the parse tree as to be less "preorder" and
more "in order", i.e. from the leaves up. The "left" command of the
pipeline should be the "leader" of the process group.
|
|
Slowly chipping away at a decent feature list.
Subshell commands are executed by @{ ... }.
|
|
|