SCT: fix use of liveness information in for loops

[vbgl]

Fixes #1218.

[eponier]

I don't understand the fix, but I trust you.

[vbgl]

You should not: I have also introduced the bug…

[eponier]

Then can you explain the problem and the fix, please?

[vbgl]

The problem is explained en the bug report: when checking a for loop and looking for a fixed point, the environments are restricted to a set of “live” variables. The set currently being used is the set of variables that are live before the for loop, which is wrong: in the example, variable a dies after evaluating the guard hence is no longer bound in the environment after analyzing the body.

The fix consists in using the correct liveness data: the set of variables live at the beginning of the loop body, i.e., after the condition.

This bug also suggests that it might be useful to store this live set in the Cfor instruction, as is done with while loops.

[eponier]

Thanks for the explanation!