Documentation

An {-# SCC #-} profiling annotation, either automatically added by the desugarer as a result of -auto-all, or added by the user.

A "tick" used by HPC to track the execution of each subexpression in the original source code.

A breakpoint for the GHCi debugger. This behaves like an HPC tick, but has a list of free variables which will be available for inspection in GHCi when the program stops at the breakpoint.

NB. we must take account of these Ids when (a) counting free variables, and (b) substituting (don't substitute for them)

A source note.

Source notes are pure annotations: Their presence should neither influence compilation nor execution. The semantics are given by causality: The presence of a source note means that a local change in the referenced source code span will possibly provoke the generated code to change. On the flip-side, the functionality of annotated code *must* be invariant against changes to all source code *except* the spans referenced in the source notes (see "Causality of optimized Haskell" paper for details).

Therefore extending the scope of any given source note is always valid. Note that it is still undesirable though, as this reduces their usefulness for debugging and profiling. Therefore we will generally try only to make use of this property where it is necessary to enable optimizations.

No scoping: The tick does not care about what code it covers. Transformations can freely move code inside as well as outside without any additional annotation obligations

Soft scoping: We want all code that is covered to stay covered. Note that this scope type does not forbid transformations from happening, as long as all results of the transformations are still covered by this tick or a copy of it. For example

let x = tick... (let y = foo in bar) in baz ===> let x = tick... bar; y = tick... foo in baz

Is a valid transformation as far as "bar" and "foo" is concerned, because both still are scoped over by the tick.

Note though that one might object to the "let" not being covered by the tick any more. However, we are generally lax with this - constant costs don't matter too much, and given that the "let" was effectively merged we can view it as having lost its identity anyway.

Also note that this scoping behaviour allows floating a tick "upwards" in pretty much any situation. For example:

case foo of x -> tick... bar ==> tick... case foo of x -> bar

While this is always legal, we want to make a best effort to only make us of this where it exposes transformation opportunities.

Cost centre scoping: We don't want any costs to move to other cost-centre stacks. This means we not only want no code or cost to get moved out of their cost centres, but we also object to code getting associated with new cost-centre ticks - or changing the order in which they get applied.

A rule of thumb is that we don't want any code to gain new annotations. However, there are notable exceptions, for example:

let f = y -> foo in tick... ... (f x) ... ==> tick... ... foo[x/y] ...

In-lining lambdas like this is always legal, because inlining a function does not change the cost-centre stack when the function is called.

Place ticks exactly on run-time expressions. We can still move the tick through pure compile-time constructs such as other ticks, casts or type lambdas. This is the most restrictive placement rule for ticks, as all tickishs have in common that they want to track runtime processes. The only legal placement rule for counting ticks. NB: We generally try to move these as close to the relevant runtime expression as possible. This means they get pushed through tyoe arguments. E.g. we create `(tick f) Bool` instead of `tick (f Bool)`.

As PlaceRuntime, but we float the tick through all lambdas. This makes sense where there is little difference between annotating the lambda and annotating the lambda's code.

In addition to floating through lambdas, cost-centre style tickishs can also be moved from constructors, non-function variables and literals. For example:

let x = scc... C (scc... y) (scc... 3) in ...

Neither the constructor application, the variable or the literal are likely to have any cost worth mentioning. And even if y names a thunk, the call would not care about the evaluation context. Therefore removing all annotations in the above example is safe.