https://github.com/akkartik/mu/blob/master/031transforms.cc
 1 //: Ordering transforms is a well-known hard problem when building compilers.
 2 //: In our case we also have the additional notion of layers. The ordering of
 3 //: layers can have nothing in common with the ordering of transforms when
 4 //: SubX is tangled and run. This can be confusing for readers, particularly
 5 //: if later layers start inserting transforms at arbitrary points between
 6 //: transforms introduced earlier. Over time adding transforms can get harder
 7 //: and harder, having to meet the constraints of everything that's come
 8 //: before. It's worth thinking about organization up-front so the ordering is
 9 //: easy to hold in our heads, and it's obvious where to add a new transform.
10 //: Some constraints:
11 //:
12 //:   1. Layers force us to build SubX bottom-up; since we want to be able to
13 //:   build and run SubX after stopping loading at any layer, the overall
14 //:   organization has to be to introduce primitives before we start using
15 //:   them.
16 //:
17 //:   2. Transforms usually need to be run top-down, converting high-level
18 //:   representations to low-level ones so that low-level layers can be
19 //:   oblivious to them.
20 //:
21 //:   3. When running we'd often like new representations to be checked before
22 //:   they are transformed away. The whole reason for new representations is
23 //:   often to add new kinds of automatic checking for our machine code
24 //:   programs.
25 //:
26 //: Putting these constraints together, we'll use the following broad
27 //: organization:
28 //:
29 //:   a) We'll divide up our transforms into "levels", each level consisting
30 //:   of multiple transforms, and dealing in some new set of representational
31 //:   ideas. Levels will be added in reverse order to the one their transforms
32 //:   will be run in.
33 //:
34 //:     To run all transforms:
35 //:       Load transforms for level n
36 //:       Load transforms for level n-1
37 //:       ...
38 //:       Load transforms for level 2
39 //:       Run code at level 1
40 //:
41 //:   b) *Within* a level we'll usually introduce transforms in the order
42 //:   they're run in.
43 //:
44 //:     To run transforms for level n:
45 //:       Perform transform of layer l
46 //:       Perform transform of layer l+1
47 //:       ...
48 //:
49 //:   c) Within a level it's often most natural to introduce a new
50 //:   representation by showing how it's transformed to the level below. To
51 //:   make such exceptions more obvious checks usually won't be first-class
52 //:   transforms; instead code that keeps the program unmodified will run
53 //:   within transforms before they mutate the program. As an example:
54 //:
55 //:     Layer l introduces a transform
56 //:     Layer l+1 adds precondition checks for the transform
57 //:
58 //: This may all seem abstract, but will hopefully make sense over time. The
59 //: goals are basically to always have a working program after any layer, to
60 //: have the order of layers make narrative sense, and to order transforms
61 //: correctly at runtime.
62 
63 :(before "End One-time Setup")
64 // Begin Transforms
65 // End Transforms