llvm dev - Feb 2019 - funnel shift, select, and poison

On 2/25/2019 1:24 PM, John Regehr via llvm-dev wrote:
> 
> This is a sound transformation only if foo() returns poison when it is 
> called with poison as an argument.
Then how do you interpret "poison has to propagate through calls"?

A typical analysis of a function will either see a call or the inlined 
body. If "call(poison)" cannot be assumed to be a poison, then a call 
effectively stops the propagation of a poison.

-Krzysztof

> A typical analysis of a function will either see a call or the inlined 
> body. If "call(poison)" cannot be assumed to be a poison, then a
call
> effectively stops the propagation of a poison.
Sorry that I was unclear.

What I am trying to say is that calls and returns (like loads and 
stores) are transparent with respect to poison.

It is definitely unsound to assume call(poison) == poison, because not 
every function returns a value that is poison-dependent on the 
function's arguments.

John

On 2/25/2019 1:35 PM, John Regehr via llvm-dev wrote:
> It is definitely unsound to assume call(poison) == poison, because not 
> every function returns a value that is poison-dependent on the 
> function's arguments.
That's exactly the same as for a "select".

I'm in favor of a set of simple rules regarding the propagation of 
poison, so I lean towards the "one answer" you proposed in the first 
email, i.e.

 > One answer might be "only select and phi stop poison"

except augmented to include calls and invokes.


-Krzysztof