SixtyPical Fallthru =================== This is a test suite, written in [Falderal][] format, for SixtyPical's ability to detect which routines make tail calls to other routines, and thus can be re-arranged to simply "fall through" to them. The theory is as follows. SixtyPical supports a `goto`, but it can only appear in tail position. If a routine r1 ends with a unique `goto` to a fixed routine r2 it is said to *potentially fall through* to r2. A *unique* `goto` means that there are not multiple different `goto`s in tail position (which can happen if, for example, an `if` is the last thing in a routine, and each branch of that `if` ends with a different `goto`.) A *fixed* routine means, a routine which is known at compile time, not a `goto` through a vector. Consider the set R of all available routines in the program. Every routine either potentially falls through to a single other routine or it does not potentially fall through to any routine. More formally, we can say > fall : R → R ∪ {nil}, fall(r) ≠ r where `nil` is an atom that represents no routine. Now consider an operation chain() vaguely similar to a transitive closure on fall(). Starting with r, we construct a list of r, fall(r), fall(fall(r)), ... with the following restrictions: - we stop when we reach `nil` (because fall(`nil`) is not defined) - we stop when we see an element that is not in R. - we stop when we see an element that we have already added to the list (this is to prevent infinite lists due to cycles.) With these definitions, our algorithm is something like this. Treat R as a mutable set and start with an empty list of lists L. Then, - For all r ∈ R, find all chain(r). - Pick a longest such chain. Call it C. - Append C to L. - Remove all elements occurring in C, from R. - Repeat until R is empty. When times comes to generate code, generate it in the order given by L. In addition, each sublist in L represents a number of routines to generate; all except the final routine in such a sublist need not have any jump instruction generated for its final `goto`. The tests in this document test against the list L. Note that this optimization is a feature of the SixtyPical's reference compiler, not the language. So an implementation is not required to pass these tests to be considered an implementation of SixtyPical. [Falderal]: http://catseye.tc/node/Falderal -> Functionality "Dump fallthru info for SixtyPical program" is implemented by -> shell command "bin/sixtypical --optimize-fallthru --dump-fallthru-info --analyze-only --traceback %(test-body-file)" -> Functionality "Compile SixtyPical program with fallthru optimization" is implemented by -> shell command "bin/sixtypical --prelude=c64 --optimize-fallthru --traceback %(test-body-file) >/tmp/foo && tests/appliances/bin/dcc6502-adapter Tests for functionality "Dump fallthru info for SixtyPical program" A single routine, obviously, falls through to nothing and has nothing fall through to it. | define main routine | { | } = [ = [ = "main" = ] = ] If `main` does a `goto foo`, then it can fall through to `foo`. | define foo routine trashes a, z, n | { | ld a, 0 | } | | define main routine trashes a, z, n | { | goto foo | } = [ = [ = "main", = "foo" = ] = ] More than one routine can fall through to a routine. We pick one of them to fall through, when selecting the order of routines. | define foo routine trashes a, z, n | { | ld a, 0 | } | | define bar routine trashes a, z, n | { | ld a, 0 | goto foo | } | | define main routine trashes a, z, n | { | goto foo | } = [ = [ = "main", = "foo" = ], = [ = "bar" = ] = ] Because `main` is always serialized first (so that the entry point of the entire program appears at the beginning of the code), nothing ever falls through to `main`. | define foo routine trashes a, z, n | { | ld a, 0 | goto main | } | | define main routine trashes a, z, n | { | ld a, 1 | } = [ = [ = "main" = ], = [ = "foo" = ] = ] There is nothing stopping two routines from tail-calling each other, but we will only be able to make one of them, at most, fall through to the other. | define foo routine trashes a, z, n | { | ld a, 0 | goto bar | } | | define bar routine trashes a, z, n | { | ld a, 0 | goto foo | } | | define main routine trashes a, z, n | { | } = [ = [ = "main" = ], = [ = "bar", = "foo" = ] = ] If a routine does two tail calls (which is possible because they can be in different branches of an `if`) it cannot fall through to another routine. | define foo routine trashes a, z, n | { | ld a, 0 | } | | define bar routine trashes a, z, n | { | ld a, 0 | } | | define main routine inputs z trashes a, z, n | { | if z { | goto foo | } else { | goto bar | } | } = [ = [ = "main" = ], = [ = "bar" = ], = [ = "foo" = ] = ] If, however, they are the same goto, one can be optimized away. | define foo routine trashes a, z, n | { | ld a, 0 | if z { | ld a, 1 | goto bar | } else { | ld a, 2 | goto bar | } | } | | define bar routine trashes a, z, n | { | ld a, 255 | } | | define main routine trashes a, z, n | { | } = [ = [ = "main" = ], = [ = "foo", = "bar" = ] = ] Similarly, a tail call to a vector can't be turned into a fallthru, because we don't necessarily know what actual routine the vector contains. | vector routine trashes a, z, n | vec | | define foo routine trashes a, z, n | { | ld a, 0 | } | | define bar routine trashes a, z, n | { | ld a, 0 | } | | define main routine outputs vec trashes a, z, n | { | copy bar, vec | goto vec | } = [ = [ = "main" = ], = [ = "bar" = ], = [ = "foo" = ] = ] Our algorithm might not be strictly optimal, but it does a good job. | define r1 routine trashes a, z, n | { | ld a, 0 | goto r2 | } | | define r2 routine trashes a, z, n | { | ld a, 0 | goto r3 | } | | define r3 routine trashes a, z, n | { | ld a, 0 | goto r4 | } | | define r4 routine trashes a, z, n | { | ld a, 0 | } | | define r5 routine trashes a, z, n | { | ld a, 0 | goto r6 | } | | define r6 routine trashes a, z, n | { | ld a, 0 | goto r3 | } | | define main routine trashes a, z, n | { | goto r1 | } = [ = [ = "main", = "r1", = "r2", = "r3", = "r4" = ], = [ = "r5", = "r6" = ] = ] -> Tests for functionality "Compile SixtyPical program with fallthru optimization" Basic test for actually applying this optimization when compiling SixtyPical programs. | define foo routine trashes a, z, n | { | ld a, 0 | } | | define bar routine trashes a, z, n | { | ld a, 255 | goto foo | } | | define main routine trashes a, z, n | { | goto foo | } = $080D LDA #$00 = $080F RTS = $0810 LDA #$FF = $0812 JMP $080D It can optimize out one of the `goto`s if they are the same. | define foo routine trashes a, z, n | { | ld a, 0 | if z { | ld a, 1 | goto bar | } else { | ld a, 2 | goto bar | } | } | | define bar routine trashes a, z, n | { | ld a, 255 | } | | define main routine trashes a, z, n | { | } = $080D RTS = $080E LDA #$00 = $0810 BNE $0817 = $0812 LDA #$01 = $0814 JMP $0819 = $0817 LDA #$02 = $0819 LDA #$FF = $081B RTS It cannot optimize out the `goto`s if they are different. Note, this currently produces unfortunately unoptimized code, because generating code for the "true" branch of an `if` always generates a jump out of the `if`, even if the last instruction in the "true" branch is a `goto`. | define foo routine trashes a, z, n | { | ld a, 0 | if z { | ld a, 1 | goto bar | } else { | ld a, 2 | goto main | } | } | | define bar routine trashes a, z, n | { | ld a, 255 | } | | define main routine trashes a, z, n | { | } = $080D RTS = $080E LDA #$FF = $0810 RTS = $0811 LDA #$00 = $0813 BNE $081D = $0815 LDA #$01 = $0817 JMP $080E = $081A JMP $0822 = $081D LDA #$02 = $081F JMP $080D