Add a known bounds optimization to ASan (Clang).

This patch is currently required to make the ASan known bounds optimization work.

[kcc2, 2012-10-18 07:31:42]

https://codereview.appspot.com/6635055/diff/1/lib/CodeGen/BackendUtil.cpp
File lib/CodeGen/BackendUtil.cpp (right):

https://codereview.appspot.com/6635055/diff/1/lib/CodeGen/BackendUtil.cpp#new...
lib/CodeGen/BackendUtil.cpp:133: PM.add(new ScalarEvolution());
I wonder if we can move this somewhere into LLVM code and under the
-asan-opt-known-bounds flag

[otinn, 2012-10-18 10:24:41]

On 2012/10/18 07:31:42, kcc2 wrote:
> https://codereview.appspot.com/6635055/diff/1/lib/CodeGen/BackendUtil.cpp
> File lib/CodeGen/BackendUtil.cpp (right):
> 
>
https://codereview.appspot.com/6635055/diff/1/lib/CodeGen/BackendUtil.cpp#new...
> lib/CodeGen/BackendUtil.cpp:133: PM.add(new ScalarEvolution());
> I wonder if we can move this somewhere into LLVM code and under the
> -asan-opt-known-bounds flag

In theory we could use the analysis only if the flag is set and the analysis
happens to be available (getAnalysisIfAvailable) but I am not aware of a way of
doing this that would always work when we want it to.

[samsonov, 2012-10-18 11:54:14]

https://codereview.appspot.com/6635055/diff/1/lib/CodeGen/BackendUtil.cpp
File lib/CodeGen/BackendUtil.cpp (right):

https://codereview.appspot.com/6635055/diff/1/lib/CodeGen/BackendUtil.cpp#new...
lib/CodeGen/BackendUtil.cpp:133: PM.add(new ScalarEvolution());
On 2012/10/18 07:31:43, kcc2 wrote:
> I wonder if we can move this somewhere into LLVM code and under the
> -asan-opt-known-bounds flag

+1

[samsonov, 2012-10-18 11:55:23]

On 2012/10/18 11:54:14, samsonov wrote:
> https://codereview.appspot.com/6635055/diff/1/lib/CodeGen/BackendUtil.cpp
> File lib/CodeGen/BackendUtil.cpp (right):
> 
>
https://codereview.appspot.com/6635055/diff/1/lib/CodeGen/BackendUtil.cpp#new...
> lib/CodeGen/BackendUtil.cpp:133: PM.add(new ScalarEvolution());
> On 2012/10/18 07:31:43, kcc2 wrote:
> > I wonder if we can move this somewhere into LLVM code and under the
> > -asan-opt-known-bounds flag
> 
> +1

Ah, disregard this, saw your response too late.

[kcc2, 2012-10-25 09:14:03]

I can live w/o ability to disable ScalarEvolution in asan pass. 

But I still don't like the fact that we are changing clang to get llvm-only
effect. 
In llvm, I see quite a few uses of INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)

Doesn't it work here?

[otinn, 2012-10-26 07:35:15]

On 2012/10/25 09:14:03, kcc2 wrote:
> I can live w/o ability to disable ScalarEvolution in asan pass. 
> 
> But I still don't like the fact that we are changing clang to get llvm-only
> effect. 
> In llvm, I see quite a few uses of INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
> 
> Doesn't it work here?
It seemed to make no difference as clang still needed the explicit pass
creation.
It worked automatically using the opt tool either way.