bug#20095: sed conditional branching

[Erik Popp]

A couple of days ago, I sent a bug report to address@hidden, but it was late enough at night that I did a bad job of explaining what was wrong.  I also forgot to include the word "sed" in the subject line.

Here's take 2:

• I'm using sed to filter a log, before passing it to the next part of a shell script
• It appears that sed's "t" command doesn't work properly under some situations
  
• The log file's format is almost identical to JSON:
• {
  property: "value",
  property2: "value",
  ...
  lastproperty: "value"
  }
  
  {
  property_of_next_entry: "value",
  ...
  
• The sed script does 2 things:
• Block any invalid entries
• Transform the log's data into a format that's easy for a shell script to process:
  property_name,value
  
• sed options:
  sed -nrf log-filter.sed
• sed version:
  GNU sed 4.2.2
• uname -a
  Linux yum 3.13.0-40-generic #69-Ubuntu SMP Thu Nov 13 17:56:26 UTC 2014 i686 athlon i686 GNU/Linux
  
• There are 4 commands that are relevant:
• Trim white space from both sides
  

  s/^[ 	]*|[ 	]*$//g

• If the line is a curly brace, jump to success handler
  

  /^[{}]$/ b success

• If the line is in the correct format for a JSON property: "value" pair, rearrange the property and value into a format that's easy for a shell script to process
  

  s/^([A-Za-z_\-]+):[ 	]*"([A-Za-z0-9 _\-\.\:\/]*)",?$/\1,\2/

• If the above substitution succeeds, jump to success handler
  

  t success
  

• The script is structured so that it should print nothing unless the script branches to a success handler.  After the above-mentioned commands, there's a "d" command.  The success handler is below that.
• The first 3 above-mentioned components work properly individually.  Of particular relevance is that if I comment out the white space remover, and also the curly brace handler, so that the first command that runs is the JSON property verifier, the "t" command works properly.  Input that matches it causes the "t" command to jump to the success handler, while input that doesn't causes it to fall through to the "d" command, as designed.
• If I run these 4 commands in sequence, the "t" command always succeeds, even if the substitution right above it fails.  It's as if on the back end of the "t" command, sed checks if any substitution in the whole script has succeed on this line, and branches if any did.  Since the white space remover line "succeeds" pretty much all the time, this would explain the symptom.
  

log-filter.sed
Description: Text document

filesystem-events.log
Description: Text document