CGAL users discussion list

[Marc Glisse <>]

On Fri, 15 Mar 2013, marcosscriven wrote:

> After an awful lot of experimenting with LLVM opt settings, I found the cause
> of this issue to be the 'early-cse' optimisation pass.
>
> I can't find a straight forward explanation of what the pass does however.

It would be more interesting to reduce the size of the code. cse usually 
looks for redundant computations and eliminates one of them. A typical 
issue is when it sees a+b (rounded up) and a+b (rounded down) and since it 
knows nothing about rounding it assumes they are the same and computes 
only one. However, sensitive code in CGAL is protected against this (using 
volatile or inline asm so llvm doesn't know it is the same 'a' in both 
operations). Many different kinds of issues (unrelated to rounding) could 
be revealed by cse, and unless you know which piece of code is 
misoptimized, it is hard to say anything.

--
Marc Glisse