Ignoring white space in regression tests really a good idea?

pg_regress compares expected and actual output using "diff -w"
(a/k/a --ignore-all-space). We have always done this, and I think
the idea was to avoid getting a lot of useless diff noise when the
only real difference is that one column value in a tabular display is
wider than expected. I'm wondering if it's such a good idea though.
I just noticed that -w was masking a couple of recent changes in the
regression outputs, notably this one in create_cast:

  SELECT 1234::int4::casttesttype; -- No cast yet, should fail
  ERROR:  cannot cast type integer to casttesttype
  LINE 1: SELECT 1234::int4::casttesttype;
!                ^
  CREATE CAST (int4 AS casttesttype) WITH INOUT;
  SELECT 1234::int4::casttesttype; -- Should work now
   casttesttype 
--- 53,59 ----
  SELECT 1234::int4::casttesttype; -- No cast yet, should fail
  ERROR:  cannot cast type integer to casttesttype
  LINE 1: SELECT 1234::int4::casttesttype;
!                          ^
  CREATE CAST (int4 AS casttesttype) WITH INOUT;
  SELECT 1234::int4::casttesttype; -- Should work now
   casttesttype 

If one supposes that one part of what this test is doing is checking
which part of the construct is pointed to by the error cursor, then
using -w in the check renders it useless.

I'm thinking maybe we should remove -w. Comments?

regards, tom lane