PR/32004, tree-ssa caused in/out asm constraints to often need reloads

[Paolo Bonzini]

> Just a latent bug showed up because the trunk and the branches
> have large differences in the dataflow.

I don't need any kind of defense because I *did* make a mistake in the 
testing and because this is *not* a latent bug.  The backport was not as 
obvious as I had thought, so in retrospect what I should have done would 
have been to first submit a patch for 4.3 only, and then a patch for 
4.2/4.1.

I remember I had made the same mistake (not looking at the wrong logs, 
O:-) rather assuming that the same fix applied to two wildly different 
codebases) with a performance PR affecting 4.0 and 3.4 (I don't remember 
the PR number); in that case Eric also had suggested reverting on 3.4 
and I did that exactly.

In this case however this is a much worse bug (it is P1 for a reason, 
even if the testcase was not reported by a user), so I'm committing the 
additional fix which has been approved by Richard.

Paolo