"w" alone also sets O_CREAT | O_TRUNC.

"+" shouldn't clear O_CREAT | O_TRUNC for "w+."

I'd suggest copying a subset of stdio/flags.c:

int m, flags;

...

flags = 0;
if (mode[0] == 'r')
    m = O_RDONLY;
else {
    m = O_WRONLY;
    flags = O_CREAT | O_TRUNC;
}
if (mode[1] == '+')
    m = O_RDWR;

fd = open(filename, m | flags);
...

Ouch :-)

const void *ptr

Should error check this

How is the API for io.write() defined? Is there a spec / could you write a short spec comment?

should we not still pushinteger(L, 0) ?

Why are we de-consting the result?

args is leaked

Why do we immediately const and de-const? fwrite() should take a const void *.

Lua strings can contain embedded nuls. There should be an API to get the length as well as the pointer from the stack argument.

The problem is we don't actually have O_CREAT | O_TRUNC support in stand. We really just need to be able to translate some of this stuff to what we've got (O_RDONLY, O_WRONLY, O_RWDR -- that's it) so that we can implement at least a subset of the functionality io wants.

Ditto above comment- we don't have these in stand, as far as I know.

This is actually just copying a whole bunch of stand stupid-ness, that I'll describe in a minute.

I've not found much of anything worthwhile, other than "it takes a file and any number of string arguments" and every example ignores its return value.

Right

Because of the above-mentioned write stupidity. write() in stand-land takes a non-const buffer. Correcting this requires another mountain of fun, because that error is cascaded through to a bunch of filesystems that would also need corrected.

Whoops =)

Right, I'll fix that. The const and de-const was copy-pasta, we should really just deconst it and be done (because the way this write stuff works here is dumb)

We could at least stub it out? Idk.

Then take a const pointer and __DECONST in fwrite(), to at least contain it there. Instead of pushing it further up the stack.

libsa isn't sacrosanct. If we need to do so, we can add something :)

We should fix the const issue, but not with this code review...

Not sure what the "Ouch" part is- this is in a section I didn't touch? =p

Ouch is the casting the ssize_t return value of read to size_t and totally ignoring possible errors. Yes, it's not something you touched. I only saw it because it was nearby.

I think we can/should just avoid these memory allocations entirely.

Just do one pass where we check that each argument is convertible to a string. Then do a second pass where we get the pointer+size and write it. (lua_tostring() actually converts the stack object in-place, so this shouldn't be "expensive" even if invoked with a lot of non-string objects.)

Sorry, lost a semicolon here in diff fungling process.

Yep

This seems to have snuck back in

wrsz appears unused

semicolon

I was surprised, but this looks correct. What a weird little language:

int lua_isstring(...) {
...
return (t == LUA_TSTRING || t == LUA_TNUMBER);

We could store the filename in our userdata for lua file objects to enhance this error message. That's probably out of scope for this patch, though.