[cmake-developers] Custom commands with Ninja on Windows

[Brad King]

On 05/20/2016 03:41 AM, Martin Ankerl wrote:
> From what I have understood it seems more safe to use .bat and not
> .cmd, because the behavior of the errorlevel is different. From
> http://waynes-world-it.blogspot.co.at/2008/08/difference-between-bat-and-cmd.html:
> 
>> The differences between .CMD and .BAT as far as CMD.EXE is concerned are:
>> With extensions enabled, PATH/APPEND/PROMPT/SET/ASSOC in .CMD files will set
>> ERRORLEVEL regardless of error. .BAT sets ERRORLEVEL only on errors.

I read that to mean the opposite: cmd is better than bat because it will
always set ERRORLEVEL even if it is to zero, so one can reliably determine
whether a command worked (Git's wrapper on windows is git.cmd and not git.bat).

> I've created a first attempt of this implementation, please see the attached
> patch.

Good start.

> +  // TODO fail if command is too long and no file specified

Several of the call sites of BuildCommandLine are constructing
commands using ninja's $VAR reference/placeholder syntax.  For
these call sites it is never safe to use a separate cmdFile.
Fortunately they tend to be for compiler/linker invocations
that can use the normal RSP_FILE approach.

Instead of adding a cmdFile argument to BuildCommandLine, add
a boolean (or enum) to indicate whether the commands hold any
placeholders.  Only when no placeholders are used is it safe to
activate the new behavior.  When activated, I think it is
cleaner to use a hash of the command line as the cmdFile name.
That way when the command changes Ninja will be aware of the
change and re-run it.

Thanks,
-Brad