493 lines
19 KiB
Factor
493 lines
19 KiB
Factor
! Copyright (C) 2008 Slava Pestov.
|
|
! See http://factorcode.org/license.txt for BSD license.
|
|
USING: slides help.markup math arrays hashtables namespaces
|
|
sequences kernel sequences parser memoize io.encodings.binary
|
|
locals kernel.private tools.vocabs.browser assocs quotations
|
|
urls peg.ebnf tools.vocabs tools.annotations tools.crossref
|
|
help.topics math.functions compiler.tree.optimizer
|
|
compiler.cfg.optimizer fry ;
|
|
IN: vpri-talk
|
|
|
|
: vpri-slides
|
|
{
|
|
{ $slide "Factor!"
|
|
{ $url "http://factorcode.org" }
|
|
"Development started in 2003"
|
|
"Open source (BSD license)"
|
|
"Influenced by Forth, Lisp, and Smalltalk"
|
|
"Blurs the line between language and library"
|
|
"Interactive development"
|
|
}
|
|
{ $slide "Programming is hard"
|
|
"Let's play tetris instead"
|
|
{ $vocab-link "tetris" }
|
|
"Tetris is hard too... let's cheat"
|
|
"Factor workflow: change code, F2, test, repeat"
|
|
}
|
|
{ $slide "Basics"
|
|
"Stack based, dynamically typed"
|
|
{ $code "{ 1 1 3 4 4 8 9 9 } dup duplicates diff ." }
|
|
"Words: named code snippets"
|
|
{ $code ": remove-duplicates ( seq -- seq' )" " dup duplicates diff ;" }
|
|
{ $code "{ 1 1 3 4 4 8 9 9 } remove-duplicates ." }
|
|
"Vocabularies: named sets of words"
|
|
{ $link "vocab-index" }
|
|
}
|
|
{ $slide "Quotations"
|
|
"Quotation: unnamed block of code"
|
|
"Combinators: words taking quotations"
|
|
{ $code "{ 1 1 3 4 4 8 9 9 }" "[ { 1 3 8 } member? ] filter ." }
|
|
{ $code "{ -1 1 -2 0 3 } [ 0 max ] map" }
|
|
"Partial application:"
|
|
{ $code ": clamp ( seq n -- seq' ) '[ _ max ] map" "{ -1 1 -2 0 3 } 0 clamp ;" }
|
|
}
|
|
{ $slide "Object system"
|
|
"CLOS with single dispatch"
|
|
"A tuple is a user-defined class which holds named values."
|
|
{ $code
|
|
"TUPLE: rectangle width height ;"
|
|
"TUPLE: circle radius ;"
|
|
}
|
|
}
|
|
{ $slide "Object system"
|
|
"Constructing instances:"
|
|
{ $code "rectangle new" }
|
|
{ $code "rectangle boa" }
|
|
"Let's encapsulate:"
|
|
{ $code
|
|
": <rectangle> ( w h -- r ) rectangle boa ;"
|
|
": <circle> ( r -- c ) circle boa ;"
|
|
}
|
|
}
|
|
{ $slide "Object system"
|
|
"Generic words and methods"
|
|
{ $code "GENERIC: area ( shape -- n )" }
|
|
"Two methods:"
|
|
{ $code
|
|
"USE: math.constants"
|
|
""
|
|
"M: rectangle area"
|
|
" [ width>> ] [ height>> ] bi * ;"
|
|
""
|
|
"M: circle area radius>> sq pi * ;"
|
|
}
|
|
}
|
|
{ $slide "Object system"
|
|
"We can compute areas now."
|
|
{ $code "100 20 <rectangle> area ." }
|
|
{ $code "3 <circle> area ." }
|
|
}
|
|
{ $slide "Object system"
|
|
"New operation, existing types:"
|
|
{ $code
|
|
"GENERIC: perimeter ( shape -- n )"
|
|
""
|
|
"M: rectangle perimeter"
|
|
" [ width>> ] [ height>> ] bi + 2 * ;"
|
|
""
|
|
"M: circle perimeter"
|
|
" radius>> 2 * pi * ;"
|
|
}
|
|
}
|
|
{ $slide "Object system"
|
|
"We can compute perimiters now."
|
|
{ $code "100 20 <rectangle> perimeter ." }
|
|
{ $code "3 <circle> perimeter ." }
|
|
}
|
|
{ $slide "Object system"
|
|
"New type, extending existing operations:"
|
|
{ $code
|
|
"TUPLE: triangle base height ;"
|
|
""
|
|
": <triangle> ( b h -- t ) triangle boa ;"
|
|
""
|
|
"M: triangle area"
|
|
" [ base>> ] [ height>> ] bi * 2 / ;"
|
|
}
|
|
}
|
|
{ $slide "Object system"
|
|
"New type, extending existing operations:"
|
|
{ $code
|
|
": hypotenuse ( x y -- z ) [ sq ] bi@ + sqrt ;"
|
|
""
|
|
"M: triangle perimeter"
|
|
" [ base>> ] [ height>> ] bi"
|
|
" [ + ] [ hypotenuse ] 2bi + ;"
|
|
}
|
|
}
|
|
{ $slide "Object system"
|
|
"Object system handles dynamic redefinition very well"
|
|
{ $code "TUPLE: person name age occupation ;" }
|
|
"Make an instance..."
|
|
}
|
|
{ $slide "Object system"
|
|
"Let's add a new slot:"
|
|
{ $code "TUPLE: person name age address occupation ;" }
|
|
"Fill it in with inspector..."
|
|
"Change the order:"
|
|
{ $code "TUPLE: person name occupation address ;" }
|
|
}
|
|
{ $slide "Object system"
|
|
"How does it work?"
|
|
"Objects are not hashtables; slot access is very fast"
|
|
"Redefinition walks the heap; expensive but rare"
|
|
}
|
|
{ $slide "Object system"
|
|
"Supports \"duck typing\""
|
|
"Two tuples can have a slot with the same name"
|
|
"Code that uses accessors will work on both"
|
|
"Accessors are auto-generated generic words"
|
|
}
|
|
{ $slide "Object system"
|
|
"More: inheritance, type declarations, read-only slots, predicate, intersection, singleton classes, reflection"
|
|
"Object system is entirely implemented in Factor"
|
|
{ { $vocab-link "generic" } ", " { $vocab-link "classes" } ", " { $vocab-link "slots" } }
|
|
}
|
|
{ $slide "The parser"
|
|
"All data types have a literal syntax"
|
|
"Literal hashtables and arrays are very useful in data-driven code"
|
|
"\"Code is data\" because quotations are objects (enables Lisp-style macros)"
|
|
{ $code "H{ { \"cookies\" 12 } { \"milk\" 10 } }" }
|
|
"Libraries can define new parsing words"
|
|
}
|
|
{ $slide "Example: float arrays"
|
|
{ $vocab-link "specialized-arrays.float" }
|
|
"Avoids boxing and unboxing overhead"
|
|
"Implemented with library code"
|
|
{ $code "float-array{ 3.14 7.6 10.3 }" }
|
|
}
|
|
{ $slide "Example: memoization"
|
|
{ "Memoization with " { $link POSTPONE: MEMO: } }
|
|
{ $code
|
|
": fib ( m -- n )"
|
|
" dup 1 > ["
|
|
" [ 1 - fib ] [ 2 - fib ] bi +"
|
|
" ] when ;"
|
|
}
|
|
"Very slow! Let's profile it..."
|
|
}
|
|
{ $slide "Example: memoization"
|
|
{ "Let's use " { $link POSTPONE: : } " instead of " { $link POSTPONE: MEMO: } }
|
|
{ $code
|
|
"MEMO: fib ( m -- n )"
|
|
" dup 1 > ["
|
|
" [ 1 - fib ] [ 2 - fib ] bi +"
|
|
" ] when ;"
|
|
}
|
|
"Much faster"
|
|
}
|
|
{ $slide "Meta-circularity"
|
|
{ { $link POSTPONE: MEMO: } " is just a library word" }
|
|
{ "But so is " { $link POSTPONE: : } }
|
|
"Factor's parser is written in Factor"
|
|
{ "All syntax is just parsing words: " { $link POSTPONE: [ } ", " { $link POSTPONE: " } }
|
|
}
|
|
{ $slide "Extensible syntax, DSLs"
|
|
"Most parsing words fall in one of two categories"
|
|
"First category: literal syntax for new data types"
|
|
"Second category: defining new types of words"
|
|
"Some parsing words are more complicated"
|
|
}
|
|
{ $slide "Example: printf"
|
|
{ { $link POSTPONE: EBNF: } ": a complex parsing word" }
|
|
"Implements a custom syntax for expressing parsers: like OMeta!"
|
|
{ "Example: " { $vocab-link "printf-example" } }
|
|
{ $code "\"vegan\" \"cheese\" \"%s is not %s\\n\" printf" }
|
|
{ $code "5 \"Factor\" \"%s is %d years old\\n\" printf" }
|
|
}
|
|
{ $slide "Example: simple web browser"
|
|
{ $vocab-link "webkit-demo" }
|
|
"Demonstrates Cocoa binding"
|
|
"Let's deploy a stand-alone binary with the deploy tool"
|
|
"Deploy tool generates binaries with no external dependencies"
|
|
}
|
|
{ $slide "Locals and lexical scope"
|
|
"Sometimes, there's no good stack solution to a problem"
|
|
"Or, you're porting existing code in a quick-and-dirty way"
|
|
"Our solution: implement named locals as a DSL in Factor"
|
|
"Influenced by Scheme and Lisp"
|
|
}
|
|
{ $slide "Locals and lexical scope"
|
|
{ "Define lambda words with " { $link POSTPONE: :: } }
|
|
{ "Establish bindings with " { $link POSTPONE: [let } " and " { $link POSTPONE: [let* } }
|
|
"Mutable bindings with correct semantics"
|
|
{ "Named inputs for quotations with " { $link POSTPONE: [| } }
|
|
"Full closures"
|
|
}
|
|
{ $slide "Locals and lexical scope"
|
|
"Combinator with 5 parameters!"
|
|
{ $code
|
|
":: branch ( a b neg zero pos -- )"
|
|
" a b = zero [ a b < neg pos if ] if ; inline"
|
|
}
|
|
"Unwieldy with the stack"
|
|
}
|
|
{ $slide "Locals and lexical scope"
|
|
{ $code
|
|
"ERROR: underage-exception ;"
|
|
""
|
|
": check-drinking-age ( age -- )"
|
|
" 21"
|
|
" [ underage-exception ]"
|
|
" [ \"Grats, you're now legal\" print ]"
|
|
" [ \"Go get hammered\" print ]"
|
|
" branch ;"
|
|
}
|
|
}
|
|
{ $slide "Locals and lexical scope"
|
|
"Locals are entirely implemented in Factor"
|
|
"Example of compile-time meta-programming"
|
|
"No performance penalty -vs- using the stack"
|
|
"In the base image, only 59 words out of 13,000 use locals"
|
|
}
|
|
{ $slide "More about partial application"
|
|
{ { $link POSTPONE: '[ } " is \"fry syntax\"" }
|
|
{ $code "'[ _ + ] == [ + ] curry" }
|
|
{ $code "'[ @ t ] == [ t ] compose" }
|
|
{ $code "'[ _ nth @ ] == [ [ nth ] curry ] dip compose" }
|
|
{ $code "'[ [ _ ] dip nth ] == [ [ ] curry dip nth ] curry" }
|
|
{ "Fry and locals desugar to " { $link curry } ", " { $link compose } }
|
|
}
|
|
{ $slide "More about partial application"
|
|
{ { $link call } " is fundamental" }
|
|
{ { $link quotation } ", " { $link curry } " and " { $link compose } " are classes" }
|
|
{ $code
|
|
"GENERIC: call ( quot -- )"
|
|
"M: curry call uncurry call ;"
|
|
"M: compose call uncompose slip call ;"
|
|
"M: quotation call (call) ;"
|
|
}
|
|
{ "So " { $link curry } ", " { $link compose } " are library features" }
|
|
}
|
|
{ $slide "Why stack-based?"
|
|
"Because nobody else is doing it"
|
|
"Interesting properties: concatenation is composition, chaining functions together, \"fluent\" interfaces, new combinators"
|
|
{ $vocab-link "smtp-example" }
|
|
{ $code
|
|
"{ \"chicken\" \"beef\" \"pork\" \"turkey\" }"
|
|
"[ 5 short head ] map ."
|
|
}
|
|
"To rattle people's cages"
|
|
}
|
|
{ $slide "Help system"
|
|
"Help markup is just literal data"
|
|
{ "Look at the help for " { $link T{ link f + } } }
|
|
"These slides are built with the help system and a custom style sheet"
|
|
{ $vocab-link "vpri-talk" }
|
|
}
|
|
{ $slide "Some line counts"
|
|
"VM: 12,000 lines of C"
|
|
"core: 9,000 lines of Factor"
|
|
"basis: 80,000 lines of Factor"
|
|
}
|
|
{ $slide "More line counts"
|
|
"Object system (core): 2184 lines"
|
|
"Dynamic variables (core): 40 lines"
|
|
"Deterministic scoped destructors (core): 56 lines"
|
|
"Optimizing compiler (basis): 12938 lines"
|
|
"Lexical variables and closures (basis): 477 lines"
|
|
"Fry (basis): 51 lines"
|
|
"Help system (basis): 1831 lines"
|
|
}
|
|
{ $slide "Implementation"
|
|
"VM: garbage collection, bignums, ..."
|
|
"Bootstrap image: parser, hashtables, object system, ..."
|
|
"Non-optimizing compiler"
|
|
"Stage 2 bootstrap: optimizing compiler, UI, ..."
|
|
"Full image contains machine code"
|
|
}
|
|
{ $slide "Compiler"
|
|
{ "Let's look at " { $vocab-link "benchmark.mandel" } }
|
|
"A naive implementation would be very slow"
|
|
"Combinators, currying, partial application"
|
|
"Boxed complex numbers"
|
|
"Boxed floats"
|
|
{ "Redundancy in " { $link absq } " and " { $link sq } }
|
|
}
|
|
{ $slide "Compiler: front-end"
|
|
"Builds high-level tree SSA IR"
|
|
"Stack code with uniquely-named values"
|
|
"Inlines combinators and calls to quotations"
|
|
{ $code "USING: compiler.tree.builder compiler.tree.debugger ;" "[ c pixel ] build-tree nodes>quot ." }
|
|
}
|
|
{ $slide "Compiler: high-level optimizer"
|
|
"12 optimization passes"
|
|
{ $link optimize-tree }
|
|
"Some passes collect information, others use the results of past analysis to rewrite the code"
|
|
}
|
|
{ $slide "Compiler: propagation pass"
|
|
"Propagation pass computes types with type function"
|
|
{ "Example: output type of " { $link + } " depends on the types of inputs" }
|
|
"Type: can be a class, a numeric interval, array with a certain length, tuple with certain type slots, literal value, ..."
|
|
"Mandelbrot: we infer that we're working on complex floats"
|
|
}
|
|
{ $slide "Compiler: propagation pass"
|
|
"Propagation also supports \"constraints\""
|
|
{ $code "[ dup array? [ first ] when ] optimized." }
|
|
{ $code "[ >fixnum dup 0 < [ 1 + ] when ] optimized." }
|
|
{ $code
|
|
"["
|
|
" >fixnum"
|
|
" dup [ -10 > ] [ 10 < ] bi and"
|
|
" [ 1 + ] when"
|
|
"] optimized."
|
|
}
|
|
}
|
|
{ $slide "Compiler: propagation pass"
|
|
"Eliminates method dispatch, inlines method bodies"
|
|
"Mandelbrot: we infer that integer indices are fixnums"
|
|
"Mandelbrot: we eliminate generic arithmetic"
|
|
}
|
|
{ $slide "Compiler: escape analysis"
|
|
"We identify allocations for tuples which are never returned or passed to other words (except slot access)"
|
|
{ "Partial application with " { $link curry } " and " { $link compose } }
|
|
"Complex numbers"
|
|
}
|
|
{ $slide "Compiler: escape analysis"
|
|
{ "Virtual sequences: " { $link <slice> } ", " { $link <reversed> } }
|
|
{ $code "[ <reversed> [ . ] each ] optimized." }
|
|
{ "Mandelbrot: we unbox " { $link curry } ", complex number allocations" }
|
|
}
|
|
{ $slide "Compiler: dead code elimination"
|
|
"Cleans up the mess from previous optimizations"
|
|
"After inlining and dispatch elimination, dead code comes up because of unused generality"
|
|
{ "No-ops like " { $snippet "0 +" } ", " { $snippet "1 *" } }
|
|
"Literals which are never used"
|
|
"Side-effect-free words whose outputs are dropped"
|
|
{ $code "[ c pixel ] optimized." }
|
|
}
|
|
{ $slide "Compiler: low level IR"
|
|
"Register-based SSA"
|
|
"Stack operations expand into low-level instructions"
|
|
{ $code "[ 5 ] test-mr mr." }
|
|
{ $code "[ swap ] test-mr mr." }
|
|
{ $code "[ append reverse ] test-mr mr." }
|
|
}
|
|
{ $slide "Compiler: low-level optimizer"
|
|
"5 optimization passes"
|
|
{ $link optimize-cfg }
|
|
"Gets rid of redundancy which is hidden in high-level stack code"
|
|
}
|
|
{ $slide "Compiler: optimize memory"
|
|
"First pass optimizes stack and memory operations"
|
|
{ "Example: " { $link 2array } }
|
|
{ { $link <array> } " fills array with initial value" }
|
|
"What if we immediately store new values into the array?"
|
|
{ $code "\\ 2array test-mr mr." }
|
|
"Mandelbrot: we optimize stack operations"
|
|
}
|
|
{ $slide "Compiler: value numbering"
|
|
"Identifies expressions which are computed more than once in a basic block"
|
|
"Simplifies expressions with various identities"
|
|
"Mandelbrot: redundant float boxing and unboxing, redundant arithmetic"
|
|
}
|
|
{ $slide "Compiler: dead code elimination"
|
|
"Dead code elimination for low-level IR"
|
|
"Again, cleans up results of prior optimizations"
|
|
}
|
|
{ $slide "Compiler: register allocation"
|
|
"IR assumes an infinite number of registers which are only assigned once"
|
|
"Real CPUs have a finite set of registers which can be assigned any number of times"
|
|
"\"Linear scan register allocation with second-chance binpacking\""
|
|
}
|
|
{ $slide "Compiler: register allocation"
|
|
"3 steps:"
|
|
"Compute live intervals"
|
|
"Allocate registers"
|
|
"Assign registers and insert spills"
|
|
}
|
|
{ $slide "Compiler: register allocation"
|
|
"Step 1: compute live intervals"
|
|
"We number all instructions consecutively"
|
|
"A live interval associates a virtual register with a list of usages"
|
|
}
|
|
{ $slide "Compiler: register allocation"
|
|
"Step 2: allocate registers"
|
|
"We scan through sorted live intervals"
|
|
"If a physical register is available, assign"
|
|
"Otherwise, find live interval with furthest away use, split it, look at both parts again"
|
|
}
|
|
{ $slide "Compiler: register allocation"
|
|
"Step 3: assign registers and insert spills"
|
|
"Simple IR rewrite step"
|
|
"After register allocation, one vreg may have several live intervals, and different physical registers at different points in time"
|
|
"Hence, \"second chance\""
|
|
{ "Mandelbrot: " { $code "[ c pixel ] test-mr mr." } }
|
|
}
|
|
{ $slide "Compiler: code generation"
|
|
"Iterate over list of instructions"
|
|
"Extract tuple slots and call hooks"
|
|
{ $vocab-link "cpu.architecture" }
|
|
"Finally, we hand the code to the VM"
|
|
{ $code "\\ 2array disassemble" }
|
|
}
|
|
{ $slide "Garbage collection"
|
|
"All roots are identified precisely"
|
|
"Generational copying for data"
|
|
"Mark sweep for native code"
|
|
}
|
|
{ $slide "History"
|
|
"Started in 2003, implemented in Java"
|
|
"Scripting language for a 2D shooter game"
|
|
"Interactive development is addictive"
|
|
"I wanted to write entire applications in Factor"
|
|
"Added JVM bytecode compiler pretty early on"
|
|
}
|
|
{ $slide "History"
|
|
"Wrote native C implementation, mid-2004"
|
|
"Added native compiler at some point"
|
|
"Added an FFI, SDL bindings, then UI"
|
|
"Switched UI to OpenGL and native APIs"
|
|
"Generational GC"
|
|
"Got rid of interpreter"
|
|
}
|
|
{ $slide "Project infrastructure"
|
|
{ $url "http://factorcode.org" }
|
|
{ $url "http://concatenative.org" }
|
|
{ $url "http://docs.factorcode.org" }
|
|
{ $url "http://planet.factorcode.org" }
|
|
"Uses our HTTP server, SSL, DB, Atom libraries..."
|
|
}
|
|
{ $slide "Project infrastructure"
|
|
"Build farm, written in Factor"
|
|
"12 platforms"
|
|
"Builds Factor and all libraries, runs tests, makes binaries"
|
|
"Saves us from the burden of making releases by hand"
|
|
"Maintains stability"
|
|
}
|
|
{ $slide "Community"
|
|
"#concatenative irc.freenode.net: 50-60 members"
|
|
"factor-talk@lists.sf.net: 180 subscribers"
|
|
"About 30 people have code in the Factor repository"
|
|
"Easy to get started: binaries, lots of docs, friendly community..."
|
|
}
|
|
{ $slide "Future direction: Factor 1.0"
|
|
"Continue doing what we're doing:"
|
|
"Polish off some language features"
|
|
"Stability"
|
|
"Performance"
|
|
"Documentation"
|
|
"Developer tools"
|
|
}
|
|
{ $slide "Future direction: Factor 2.0"
|
|
"Native threads"
|
|
"Syntax-aware Factor editor"
|
|
"Embedding Factor in C apps"
|
|
"Cross-compilation for smaller devices"
|
|
}
|
|
{ $slide "Research areas"
|
|
"Identify areas where stack languages are lacking, and try to find idioms, abstractions or DSLs to solve these problems"
|
|
"Factor is a good platform for DSLs (fry, locals, EBNF, help, ...); what about implementing a complete language on top?"
|
|
"Static typing, soft typing, for stack-based languages"
|
|
}
|
|
{ $slide "That's all, folks"
|
|
"It is hard to cover everything in a single talk"
|
|
"Factor has many cool things that I didn't talk about"
|
|
"Questions?"
|
|
}
|
|
} ;
|
|
|
|
: vpri-talk ( -- ) vpri-slides slides-window ;
|
|
|
|
MAIN: vpri-talk
|