Re: branch bootstrap breakage in new DECL_RTL

[Robert Lipe <robertlipe at usa dot net>]

Mark Mitchell wrote:
> >>>>> "Richard" == Richard Kenner <kenner@vlsi1.ultra.nyu.edu> writes:
> 
>     Richard>     A test server would have tested, found the problem,
>     Richard> and it would have taken me only minutes to correct the
>     Richard> patch -- all without making poor Robert have to track it
>     Richard> down.
> 
>     Richard> A test server that knew to make this particular check,
>     Richard> you mean.
> 
> A test server that works as a user would work: 
> 
>   cvs checkout
>   configure
>   make bootstrap
> 
> on Robert's machine (and, in fact, on almost any machine without GNU

Since that's what we instruct users to do, that's essentially what my 
machine has slavishly done (twice; once for OpenServer, once for UDK)
each day for the last couple of years.   It's downright depressing how
often it doesn't work.   One of the two targets has seen exactly one 
successful bootstrap (December 6) since September 12.   It had only about
4,000 unexpected failures.  This same target worked "as well as Linux" 
prior to that.   I just haven't been able to keep up with the constant
breakage. 

As an example of variation from the above, I've learned to have it do a 
'CC=/evil/vendor/compiler .../configure' to prevent it from being sucked 
into the same "all the world is gcc on Linux" trap that trips up so many
commits.   At least I can sometimes raise a flag when we break things.

> CC as the host C compiler) would have caught this.  As it would have
> caught the `grep' problem on IRIX.  And lots and lots of others.

It does catch lots and lots.  It catches GNU C extensions that creep in.   
It catches wrong use of alloca.   It catches unguarded use of inline or
desginated initializers.  It's caught enough to earn me a reputation
as a whiner. :-)

RJL