Re: [RFC] Fixing gdb.base/completion.exp (PR testsuite/12649)

[Jan Kratochvil <jan.kratochvil@redhat.com>]

On Thu, 28 Apr 2011 17:14:31 +0200, Pedro Alves wrote:
>  set oldtimeout1 $timeout
> -set timeout 30
> +set timeout 10

10 is too low for parallel runs where machine can be in 20+ load.  I do not
see this test needs to excercise $timeout, I would even remove this whole
override.


> @@ -114,7 +113,6 @@ gdb_expect  {
>  #exp_internal 0
>  
>  send_gdb "show output\t"
> -sleep 1
>  gdb_expect  {
>          -re "^show output-radix $"\
>              { send_gdb "\n"
###              gdb_expect {
###                      -re "Default output radix for printing of values is 10\\..*$gdb_prompt $"\
###                                        { pass "complete 'show output'"}
###                      -re ".*$gdb_prompt $" { fail "complete 'show output'"}
> @@ -125,16 +123,6 @@ gdb_expect  {
>                        timeout           {fail "(timeout) complete 'show output'"}
>                       }
>              }
> -        -re "^show output$"\
> -            { send_gdb "\n"
> -               gdb_expect {
> -                      -re "Default output radix for printing of values is 10\\..*$gdb_prompt $"\
> -                                        { fail "complete 'show output'"}
> -                      -re ".*$gdb_prompt $" { fail "complete 'show output'"}
> -                      timeout           { fail "(timeout) complete 'show output'"}
> -                     }
> -
> -             }
>  
>          -re ".*$gdb_prompt $"       { fail "complete 'show output'" }
>          timeout         { fail "(timeout) complete 'show output'" }


The problem with this proposed intermediate step is that it in fact brings a
testsuite regression.  Original "sleep 1" was there to ensure all the output
has been caught.  This was racy but in most cases it worked.

Now it will false PASS with regressing GDB where the current FSF GDB HEAD
testcase would correctly FAIL.  If GDB outputs "show output-radix " first and
after 0.5sec it yet outputs "foobar" the original testcase correctly FAILed
while the current testcase will falsely PASS.

The "complete" command appraoch does introduce this new kind of race.

But the patch can be commited in two parts if it is preferred although
reviewing these racy send_gdb-gdb_expect cases for the intermediate step is
tricky and it gets dropped immediately afterwards.


> @@ -410,7 +365,7 @@ gdb_expect  {
>  		    timeout           {fail "(timeout) complete 'p \"break1.'"}
>  		}
>  	    }
> -	-re "^p \"break1\\..*$"
> +	-re "^p \"break1\\...*$"
>  	    {	send_gdb "\n"
>  		gdb_expect {
>  		    -re ".*$gdb_prompt $" { fail "complete 'p \"break1.'"}

I do not see this change as valid/relevant.


Thanks,
Jan