Re: [path] PR 54900: store data race in if-conversion pass

[Andrew MacLeod <amacleod@redhat.com>]

On 10/15/2012 08:28 AM, Aldy Hernandez wrote:
>
> I am having a bit of a problem coming up with a generic testcase. 
> Perhaps Andrew or others have an idea.
>
> The attached testcase fails to trigger without the patch, because 
> AFAICT we have no way of testing an addition of zero to a memory 
> location:
>
>         cmpl    $1, flag(%rip)
>         setb    %al
>         addl    %eax, dont_write(%rip)
>
> In the simulate-thread harness I can test the environment before an 
> instruction, and after an instruction, but adding 0 to *dont_write 
> produces no measurable effects, particularly in a back-end independent 
> manner.  Ideas?

Hum. isn't that clever.   Well, the instruction is executed pretty much 
atomically... so a write of the same value becomes very difficult to 
detect, and impossible within the existing harness. And I dont think a 
hardware watch point can catch that...

The only way I can think of is if you put 'dont_write' into a section 
which will trap if it is written to...  I don't know the details of 
doing such a thing or how you monitor the trap within the harness...

Other than that I'm not sure we can detect this with our current set of 
tools, for the longer term we'd need a write detector.  I don't suppose 
something like systemtap can detect writes like this?

Andrew