[LLVMdev] Dependence Analysis [was: Flow-Sensitive AA]
dag at cray.com
Tue Aug 26 15:29:30 PDT 2008
On Friday 22 August 2008 17:07, Vikram S. Adve wrote:
> On Aug 22, 2008, at 4:27 PM, David Greene wrote:
> > Exactly right. One of my first jobs here was to fix a bunch of
> > overflow
> > problems exposed by optimization. It's now become a tradition to
> > give this task to any new optimizer employee. :)
> But that sounds like Cray is being fairly conservative in treating
> overflow problems as errors during optimization. Is that correct?
> That's more or less the opposite of what Dan Berlin was suggesting.
All the problems I fixed involved Cray being too aggressive. We needed to
add checks to handle unsigned int overflow. Customer codes were breaking.
For unsigned int in C compilers *must* be conservative because that's what
the standard says. This pops up in places you might not expect, right now
deep into the guts of expression reordering and canonicalization. Because
of this, various higher-level optimizations that depend on expressions being
in a certain form also tend to get shut down because the canonicalizer can't
Unsigned arithmetic is just evil.
> This is not just idle curiosity -- I'm trying to get an understanding
> of what real-world compilers are actually doing on this issue.
Absolutely. Lots of nasty stuff starts to happend with unsigned arithmetic.
As I understand it, -fwrapv is for those codes that violate the standard --
they assume signed arithmetic wraps. I've not seen any real-world codes
that depend on that but the gcc guys certainly see a lot more C than we do.
More information about the llvm-dev