[PATCH] Branch folding causes different code generation at "-O2 -g" and "-O2"

[Robinson, Paul]

> Are there really no other tests for debug info changing codegen in the
> x86 tree?

The LLVM test infrastructure doesn't support doing that in a general way.
We've been finding these things in a test suite where we can make the
test driver compile the same file two ways and compare the generated code.
It's easy to do that when you start with C/C++ and just add -g but not
so easy when you're starting with IR.

Personally I think inserting IR instructions to carry debug annotations
is a poor choice, but replacing it with something that optimizations
are less inclined to screw up would be a really big project.
--paulr