Variable references in build inputs don't take into account the build's bindings

[Cole Faust]

rule touch
    command = touch $out

rule copy
    command = cp $in $out

build out/a: copy ${my_dep}
    my_dep = out/b

build out/b: touch

default out/a

This build file works in ninja, but not n2. Android has this case. I've only found one occurance that was trivial to remove so far, but I haven't gotten a full build working yet. If possible, maybe it's better not to fix it if it allows faster parsing.

[Evan Martin]

This is #39 and it is dumb that I didn't make it work, but I was greedy.

The reason for my greed is that n2 currently expands everything right while it parses, so when it sees ${my_dep} it wants to expand it to the final value without waiting to parse down through the definition of my_dep = . This means it never needs to buffer any of these intermediate strings. But it also means it's just plain incorrect relative to Ninja.

Now that I look at your example, I wonder if I could do some trick around assuming things will go the fast way and then if I see a variable reference later going back and re-expanding. Or if I should just buffer everything anyway. It would probably help for me to actually profile before I micro-optimize in this manner.

[Evan Martin]

Relevant expansion inline code, I think maybe that ought to consult more scopes than it currently does too...

[Cole Faust]

Oh, feel free to close as a duplicate then. Interestingly, this issue also causes parse errors:

rule touch
    command = touch $out

rule copy
    command = cp $in $out

my_var = out/b

build out/a: copy out/b | ${my_dep} out/d
    my_dep = out/c

build out/b: touch
build out/c: touch
build out/d: touch

default out/a

Produces:

n2: error: parse error: expected '\n', got ' '
build.ninja:9: build out/a: copy out/b | ${my_dep} out/...
                                                  ^

I actually think this will need to be properly fixed for the android build. Not all of these search results, but many of them, suffer from this issue.

[Evan Martin]

(Sorry for not responding yet -- your other patches were obvious merges, but this one is more subtle so I want to give it proper attention and I have some life stuff going on the next week or so.)

[Cole Faust]

Thanks for the update, I'll hold off on new PRs for now then.