[PATCH] Fix CSE bogus RTL creation

[Paolo Bonzini]

On 06/15/2010 04:08 PM, Mark Mitchell wrote:
> Of course, I am all for superior patches, good infrastructure,
> clean-ups, and so forth.  But, it seemed clear to me that (a)
> validate_change wasn't in the business of making canonical RTL, and
> (b) simplify_rtx was

For that matter, canonicalize_change_group also is, and is more directly
related to validate_change.

> and (c) this patch made the compiler better from a user-observable
> point of view, in that it eliminated crashes,

Without knowing the real bug, you cannot know if there are other
identical crashes waiting to bite you.

> and (d) there had been some previous discussion suggesting this
> approach

Where?

> I think that should be good enough.  Now, of course, if you have an
> idea about how to do it better, I think that's great!  A more
> general, more elegant solution would be terrific.

I gave a patch and Steven outlined another idea that might be even
better.  Again, I cannot say if Steven's idea would work without knowing
what the real bug is.

> But, I don't want perfect to be the enemy of good.

But "papering over the real bug" _is_ the enemy of "good".  That is
something clearly undesirable, not just suboptimal, especially at the
beginning of stage1.

I'm not opposing the patch at all costs, but definitely the bug hasn't
been analyzed thoroughly enough.

Paolo