llvm dev - Nov 2017 - How and where optimizing of undefined behavior happens

Dear community,

I was wondering how exactly clang/llvm handles undefined behavior (UB).
Because of this experiment: https://godbolt.org/g/cEmGHa,
I tend to think that clang is light-weight in this subject and does
not detect UB.
And all the optimizations related to UB are done in llvm via some
optimization pass (passes).

Is it so? If so than what is the name of that pass? And do different
optimization levels detect different UB cases, i.e. will situation
change with O1, O2, O3?

-- 
Best regards,
Denis.

By default, Clang has little or nothing to do with UB, its job is to do 
a straightforward translation to IR.  UB-related optimizations are 
spread throughout the passes.  You can get a bit of an idea of what's 
going on by running this command from an LLVM source tree:

   find ./lib -name '*.cpp' | xargs grep -i undef

But also, a lot of sanitizer code lives in Clang.

John


On 11/18/2017 11:35 AM, Denis Bakhvalov via llvm-dev
wrote:
> Dear community,
> 
> I was wondering how exactly clang/llvm handles undefined behavior (UB).
> Because of this experiment: https://godbolt.org/g/cEmGHa,
> I tend to think that clang is light-weight in this subject and does
> not detect UB.
> And all the optimizations related to UB are done in llvm via some
> optimization pass (passes).
> 
> Is it so? If so than what is the name of that pass? And do different
> optimization levels detect different UB cases, i.e. will situation
> change with O1, O2, O3?
>

On 18 Nov 2017, at 20:55, John Regehr via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> 
> By default, Clang has little or nothing to do with UB, its job is to do a
straightforward translation to IR.
That’s not quite true, because the mapping to IR helps define what is UB in the
source language.  For example, if you don’t enable the undefined behaviour
sanitiser or pass -ftrapv / -fwrapv, clang will set the nsw (no signed wrap)
flag on IR arithmetic operations on types that are signed in C, which then
permits the optimisers to treat wrapping behaviour as undefined.

David