On 08/10/2011 08:27 PM, Mark Kettenis wrote:
>> Date: Tue, 09 Aug 2011 23:25:06 +0800
>> From: Yao Qi <yao@codesourcery.com>
>>
>> On 08/09/2011 11:15 PM, Mark Kettenis wrote:
>>> Why depend on this NOMMU-magic?  Just install the signal handler for
>>> bth SIGSEGV and SIGILL, try a store to (or perhaps a read from)
>>> address 0, and then fall through to executing an illegal instruction.
>>
>> Because I want to reduce the scope of using invalid instruction.  We
>> only need to know the invalid instruction for HAS_NOMMU arch, and emit
>> error if we forget to define an invalid instruction for a new NOMMU
>> port.  Do it make sense?
> 
> I don't really think that's an issue.  If the test is run on an
> mmu-less machine for which no illegal instruction is defined, the test
> will fail.  That should prompt someone to look at the test and add the
> missing instruction.
> 

If you don't think that is an issue, I am OK with it.  Updated my patch
for it.

> I really just want to avoid the #ifdef maze you're creating which
> makes the code more complicated and the test less generic.  I think
> you're too much focussed on reducing the number of FAILs for your
> particular target to zero instead of improving the tests such that
> they become more generally useful.  For example:
> 

I don't know how did you get such impression, but I really want to
convert/refactor test cases as general as they can be.  gdb-patches@
archive can show my recent work is about this.

>> diff --git a/gdb/testsuite/gdb.base/savedregs.exp b/gdb/testsuite/gdb.base/savedregs.exp
>> index eeee0ff..4408137 100644
>> --- a/gdb/testsuite/gdb.base/savedregs.exp
>> +++ b/gdb/testsuite/gdb.base/savedregs.exp
>> @@ -84,6 +84,14 @@ proc process_saved_regs { current inner outer } {
>>  		# Sigtramp frames don't yet print <signal trampoline>.
>>  		set pat "Stack frame at .* Saved registers:.*"
>>  	    }
>> +	    thrower {
>> +		if { [istarget tic6x-*-*] } {
>> +		    # On tic6x, there is no register saved in function thrower.
>> +		    set pat "Stack frame at .* in $func .*"
>> +		} else {
>> +		    set pat "Stack frame at .* in $func .* Saved registers:.*"
>> +		}
> 
> Why are you special-casing tic6x here?  Is the architecture really
> that special that there are no saved registers?  I suspect it isn't
> and that this can happen on other architectures as well, depending on
> how much optimization the compiler is doing.

Leave tic6x alone at first, it is the test case's problem here to expect
"save registers in a frame", because it is possible there is no register
saved on a certain frame.  IMO, tic6x port exposes such problem, and my
fix in this patch is to make tests "more generally useful".  If we see
"no registers saved" on other ports, we can put these targets together
in this condition checking, like,

if { [istarget tic6x-*-*] || [istarget foo-*-*] || [istarget bar-*-*]} {
    # On tic6x/foo/bar, there is no register saved in function thrower.
    set pat "Stack frame at .* in $func .*"
} else {
    set pat "Stack frame at .* in $func .* Saved registers:.*"
}

Is it OK to you?

-- 
Yao (齐尧)