Add rule for rand(Bool) instead of rand() < 0.5

[iuliadmtru]

After this Slack thread.

[iuliadmtru]

The fix does not remove the assignment in cases like:

x = rand()
if x < 0.5
    ...
end

This fixes to:

x = rand()
if rand(Bool)
    ...
end

[imciner2]

The fix does not remove the assignment in cases like:

x = rand()
if x < 0.5
    ...
end

This fixes to:

x = rand()
if rand(Bool)
    ...
end

This is different code semantically. If x was being used anywhere inside the if statement, then this can cause different execution behavior, since now the entrance to the if statement is guarded by a different random draw. The only way this transformation is valid is if x is not used anywhere else.

Take this section of code for instance:

x = rand()
if x < 0.5
    y = sqrt(0.5 - x)
end

Under this transformation, it is now possible to end up with a domain error in the square root when it wasn't before - because if this is transformed to

x = rand()
if rand(Bool)
    y = sqrt(0.5 - x)
end

then there is no longer a guarantee that x is less than 0.5 allowing 0.5 - x to be always positive. It could be that rand(Bool) returns true when x=rand() has returned 0.75 for instance.

[iuliadmtru]

I hadn't thought of that! I would say then either just remove the autofix, or remove altogether the pattern that checks for assigning rand() to a variable and using that variable (and might as well keep the autofix in that case). Not sure which is better.

[iuliadmtru]

Other ideas:

• keep this as a partial fix, needing checking by whoever runs the runs (not nice)
• write two separate rules, one for rand() < 0.5 with autofix, one for x = rand(); x < 0.5 without autofix

[aviks]

write two separate rules, one for rand() < 0.5 with autofix, one for x = rand(); x < 0.5 without autofix

maybe this is the best way forward?

[iuliadmtru]

I split it in two rules. The test files are the same, only the rule names differ (and the tests they pass/fail for of course). The name for the second rule is not very inspired...

[aviks]

Looks good. Can we add a description of some sort in the rule definition files so that I'm not confused when I see this in a few months?