Minor problem with multiple targets and 'restat'

[maximuska]

Minor problem with multiple targets and 'restat' (anyone uses that?) which I've noticed while working on #608.

rule true
  command = [ -e auto.h ] || touch auto.h  # Simulate 'restat'
  restat = 1 
rule cat 
  command = touch $out
build auto.h auto.c: true
build out: cat auto.c | auto.h

There is a bug in the 'true' command above, it generates only one of the outputs and returns '0'.
On the first build:

ubuntu: ninja> ./ninja -d explain -v -f tst2.ninja 
ninja explain: command line not found in log for auto.h
ninja explain: auto.c is dirty
ninja explain: auto.h is dirty
[1/2] [ -e auto.h ] || touch auto.h  # Simulate 'restat'
ninja explain: command line not found in log for out
[2/2] touch out

that's ok.

On the 2nd:

ubuntu: ninja> ./ninja -d explain -v -f tst2.ninja 
ninja explain: output auto.c doesn't exist
ninja explain: auto.c is dirty
ninja explain: auto.h is dirty
[1/2] [ -e auto.h ] || touch auto.h  # Simulate 'restat'

That I would expect is that 'out' would be rebuilt also (as auto.c doesn't exist and restat was not supposed to clean 'out'), but it is not.

ubuntu: ninja> tail -n1 .ninja_log
0 2 0 auto.c 97a40aff32c1d21a

I think the problem is that '0' stored as 'restat' value of auto.c is also used as 'file doesn't exists' indicator in 'Node'.

[nico]

Pull request #634 fixes another issue that is caused by mtime_ == 0 meaning "doesn't exist" – maybe it fixes this issue too?

[jhasse]

This still happens. I'm wondering why 'out' should be rebuild? auto.c still doesn't exist.

What do you mean by

restat was not supposed to clean 'out'

? It didn't clean 'out', didn't it?