The Highs and Lows of Macros in a Modern Language Laurence Tratt - - PowerPoint PPT Presentation

▶
the highs and lows of macros in a modern language
SMART_READER_LITE
LIVE PREVIEW

The Highs and Lows of Macros in a Modern Language Laurence Tratt - - PowerPoint PPT Presentation

Jan 04, 2023 216 likes •1.76k views

The Highs and Lows of Macros in a Modern Language Laurence Tratt Software Development Team 2016-08-09 1 / 59 http://soft-dev.org/ Background Background ABCgjy 2 / 59 http://soft-dev.org/ Background Background A perfect programming

slide-1
SLIDE 1

The Highs and Lows of Macros in a Modern Language

Laurence Tratt

Software Development Team 2016-08-09

1 / 59 http://soft-dev.org/

slide-2
SLIDE 2

Background Background

ABCgjy

2 / 59 http://soft-dev.org/

slide-3
SLIDE 3

Background Background

A perfect programming language

2 / 59 http://soft-dev.org/

slide-4
SLIDE 4

Background Background

Solution ABCgjy

3 / 59 http://soft-dev.org/

slide-5
SLIDE 5

Background Background

Solution A new programming language

3 / 59 http://soft-dev.org/

slide-6
SLIDE 6

Background Background

Reality ABCgjy

3 / 59 http://soft-dev.org/

slide-7
SLIDE 7

Background Background

Reality Another imperfect programming language

3 / 59 http://soft-dev.org/

slide-8
SLIDE 8

What to expect from this talk What to expect from this talk

4 / 59 http://soft-dev.org/

slide-9
SLIDE 9

What to expect from this talk What to expect from this talk

1 What happens when you put macros

into a modern programming language?

4 / 59 http://soft-dev.org/

slide-10
SLIDE 10

What to expect from this talk What to expect from this talk

1 What happens when you put macros

into a modern programming language?

2 If it doesn’t work out well, is there an

alternative?

4 / 59 http://soft-dev.org/

slide-11
SLIDE 11

Part I Defining the area

5 / 59 http://soft-dev.org/

slide-12
SLIDE 12

What is a macro? What is a macro?

6 / 59 http://soft-dev.org/

slide-13
SLIDE 13

What is a macro? What is a macro?

It’s complicated...

6 / 59 http://soft-dev.org/

slide-14
SLIDE 14

What is a macro? What is a macro?

It’s complicated... Let’s simplify to “a calculation that happens at compile-time”.

6 / 59 http://soft-dev.org/

slide-15
SLIDE 15

Text substitution Text substitution

This C fragment:

#define sq2(y) ((y) * (y)) int main() { printf("%d\n", sq2(3)); }

is preprocessed to:

int main() { printf("%d\n", ((3) * (3))); }

and then compiled.

7 / 59 http://soft-dev.org/

slide-16
SLIDE 16

Text substitution: the good Text substitution: the good

Some clever (and useful) things are possible e.g.:

#define TRY { \ jmp_buf _env; \ if (setjmp(_env) == 0) { \ add_exception_frame(_env); #define CATCH(v) \ remove_exception_frame(); \ } \ else { \ (v) = read_and_reset_exception(); #define TRY_END } }

8 / 59 http://soft-dev.org/

slide-17
SLIDE 17

Text substitution: the good Text substitution: the good

Some clever (and useful) things are possible e.g.:

#define TRY { \ jmp_buf _env; \ if (setjmp(_env) == 0) { \ add_exception_frame(_env); #define CATCH(v) \ remove_exception_frame(); \ } \ else { \ (v) = read_and_reset_exception(); #define TRY_END } }

can be used – fairly naturally – for:

Exception *e; TRY { ... } CATCH (e) { ... } TRY_END

8 / 59 http://soft-dev.org/

slide-18
SLIDE 18

Text substitution: the bad Text substitution: the bad

What does the following print out?

#define sq2(y) (y * y) int main() { printf("%d\n", sq2(3)); printf("%d\n", sq2(1+2)); }

9 / 59 http://soft-dev.org/

slide-19
SLIDE 19

Text substitution: the bad Text substitution: the bad

What does the following print out?

#define sq2(y) (y * y) int main() { printf("%d\n", sq2(3)); printf("%d\n", sq2(1+2)); }

Obviously 9, 5?!

9 / 59 http://soft-dev.org/

slide-20
SLIDE 20

Text substitution: the bad Text substitution: the bad

What does the following print out?

#define sq2(y) (y * y) int main() { printf("%d\n", sq2(3)); printf("%d\n", sq2(1+2)); }

Obviously 9, 5?! What about:

#define sq2(y) ((y) * (y)) typedef struct { int y; } C; int main() { C x; x.y = 3; printf("%d\n", sq2(++x.y)); }

9 / 59 http://soft-dev.org/

slide-21
SLIDE 21

Text substitution: the bad Text substitution: the bad

What does the following print out?

#define sq2(y) (y * y) int main() { printf("%d\n", sq2(3)); printf("%d\n", sq2(1+2)); }

Obviously 9, 5?! What about:

#define sq2(y) ((y) * (y)) typedef struct { int y; } C; int main() { C x; x.y = 3; printf("%d\n", sq2(++x.y)); }

Obviously 20?!

9 / 59 http://soft-dev.org/

slide-22
SLIDE 22

Text substitution: the bad Text substitution: the bad

What does the following print out?

#define sq2(y) (y * y) int main() { printf("%d\n", sq2(3)); printf("%d\n", sq2(1+2)); }

Obviously 9, 5?! What about:

#define sq2(y) ((y) * (y)) typedef struct { int y; } C; int main() { C x; x.y = 3; printf("%d\n", sq2(++x.y)); }

Obviously 20?! There are other problems too, but you get the idea...

9 / 59 http://soft-dev.org/

slide-23
SLIDE 23

Hetergeneous vs. homogeneous Hetergeneous vs. homogeneous

Hetergeneous: where the meta-programming language/system (e.g. the C preprocessor) is different than the main language/system (e.g. C). Homogeneous: where the two are the same. Crudely: hetergeneous is powerful, but difficult to use, and unsafe; homogeneous is safe(r) and easier to use.

[See Sheard 2003 ‘Accomplishments and Research Challenges in Meta-programming’]

10 / 59 http://soft-dev.org/

slide-24
SLIDE 24

Lisp Lisp

The ‘Lisp’ family is huge.

11 / 59 http://soft-dev.org/

slide-25
SLIDE 25

Lisp Lisp

The ‘Lisp’ family is huge. In a typical-ish Lisp, one might do:

(defmacro sq2 (e) (list ’* e e)) (print (macroexpand ’(* (+ 1 2) (+ 1 2)))) (print (macroexpand ’(sq2 (+ 1 2))))

which will print:

(* (+ 1 2) (+ 1 2)) (* (+ 1 2) (+ 1 2))

Note: everything is done on trees.

11 / 59 http://soft-dev.org/

slide-26
SLIDE 26

The tumbleweed years The tumbleweed years

For decades, macro research was Lisp. Why?

12 / 59 http://soft-dev.org/

slide-27
SLIDE 27

The tumbleweed years The tumbleweed years

For decades, macro research was Lisp. Why? Brackets (maybe); homoiconicity (definitely).

12 / 59 http://soft-dev.org/

slide-28
SLIDE 28

The tumbleweed years The tumbleweed years

For decades, macro research was Lisp. Why? Brackets (maybe); homoiconicity (definitely). Until MetaML and successors, including Template Haskell.

12 / 59 http://soft-dev.org/

slide-29
SLIDE 29

Part II What happens when you put macros into a modern programming language?

13 / 59 http://soft-dev.org/

slide-30
SLIDE 30

Converge Converge

14 / 59 http://soft-dev.org/

slide-31
SLIDE 31

Converge Converge

Summary: Python + TH-esque macros

import Sys func main(): Sys::println("hello world")

14 / 59 http://soft-dev.org/

slide-32
SLIDE 32

Compile-time Meta-programming / Macros Compile-time Meta-programming / Macros

Code (as trees, not text) is programatically generated.

15 / 59 http://soft-dev.org/

slide-33
SLIDE 33

Compile-time Meta-programming / Macros Compile-time Meta-programming / Macros

Code (as trees, not text) is programatically generated. Expression

2 + 3

evaluates to 5 (as one expects). Splice

$<x>

evaluates

x

at compile-time; the AST returned overwrites the splice. Quasi-quote [| 2 + 3 |] evaluates to a hygienic AST repre- senting 2 + 3. Insertion

[| 2 + ${x} |] ‘inserts’ the AST x into the AST be-

ing created by the quasi-quotes.

15 / 59 http://soft-dev.org/

slide-34
SLIDE 34

When do things execute? When do things execute?

When are x and y evaluated?

$<x> func main(): y

16 / 59 http://soft-dev.org/

slide-35
SLIDE 35

The power Function The power Function

We want:

power3 := $<mk_power(3)>

to be compiled to:

power3 := func (x): return x * x * x * 1

How to do it?

17 / 59 http://soft-dev.org/

slide-36
SLIDE 36

The printf Function The printf Function

18 / 59 http://soft-dev.org/

slide-37
SLIDE 37

What can we use this stuff for? What can we use this stuff for?

19 / 59 http://soft-dev.org/

slide-38
SLIDE 38

What can we use this stuff for? What can we use this stuff for? IMHO, macros are useful if your language has:

1 very little syntax

19 / 59 http://soft-dev.org/

slide-39
SLIDE 39

What can we use this stuff for? What can we use this stuff for? IMHO, macros are useful if your language has:

1 very little syntax and/or 2 a (restrictive) static type system.

19 / 59 http://soft-dev.org/

slide-40
SLIDE 40

What can we use this stuff for? What can we use this stuff for? IMHO, macros are useful if your language has:

1 very little syntax and/or 2 a (restrictive) static type system.

Neither is true in modern dynamically typed languages.

19 / 59 http://soft-dev.org/

slide-41
SLIDE 41

What can we use this stuff for? What can we use this stuff for? IMHO, macros are useful if your language has:

1 very little syntax and/or 2 a (restrictive) static type system.

Neither is true in modern dynamically typed languages. Do macros have uses?

19 / 59 http://soft-dev.org/

slide-42
SLIDE 42

Embedding DSLs Embedding DSLs

Splice

$<x>

evaluates

x

at compile-time; the AST returned overwrites the splice. Quasi-quote [| 2 + 3 |] evaluates to a hygienic AST repre- senting 2 + 3. Insertion

[| 2 + ${x} |] ‘inserts’ the AST x into the AST be-

ing created by the quasi-quotes.

20 / 59 http://soft-dev.org/

slide-43
SLIDE 43

Embedding DSLs Embedding DSLs

Splice

$<x>

evaluates

x

at compile-time; the AST returned overwrites the splice. Quasi-quote [| 2 + 3 |] evaluates to a hygienic AST repre- senting 2 + 3. Insertion

[| 2 + ${x} |] ‘inserts’ the AST x into the AST be-

ing created by the quasi-quotes. DSL blocks

$<<x>>: y

pass the string y to the function x at compile-time.

20 / 59 http://soft-dev.org/

slide-44
SLIDE 44

Building a DSL Building a DSL

21 / 59 http://soft-dev.org/

slide-45
SLIDE 45

DSL debugging DSL debugging

We normally assume that compilers are perfect

22 / 59 http://soft-dev.org/

slide-46
SLIDE 46

DSL debugging DSL debugging

We normally assume that compilers are perfect DSL compilers are probably imperfect

22 / 59 http://soft-dev.org/

slide-47
SLIDE 47

DSL debugging DSL debugging

We normally assume that compilers are perfect DSL compilers are probably imperfect Are errors due to the user or the compiler?

22 / 59 http://soft-dev.org/

slide-48
SLIDE 48

Static error reporting Static error reporting

23 / 59 http://soft-dev.org/

slide-49
SLIDE 49

Run-time error reporting Run-time error reporting Src infos are a triple: (file ID, char offset, char span) Threaded throughout the compiler:

1 Each token/lexeme has one src info 2 Each parse tree has more than one src info 3 Each bytecode has more than one src info

24 / 59 http://soft-dev.org/

slide-50
SLIDE 50

Fun with names Fun with names

Dynamic scoping is dangerous.

25 / 59 http://soft-dev.org/

slide-51
SLIDE 51

Fun with names Fun with names

Dynamic scoping is dangerous. Can it be made safe?

25 / 59 http://soft-dev.org/

slide-52
SLIDE 52

Meta-levels Meta-levels Three relative meta-levels describe everything: Meta-level Description

26 / 59 http://soft-dev.org/

slide-53
SLIDE 53

Meta-levels Meta-levels Three relative meta-levels describe everything: Meta-level Description Normal compilation

26 / 59 http://soft-dev.org/

slide-54
SLIDE 54

Meta-levels Meta-levels Three relative meta-levels describe everything: Meta-level Description

  • 1

Splicing ($<...>) Normal compilation

26 / 59 http://soft-dev.org/

slide-55
SLIDE 55

Meta-levels Meta-levels Three relative meta-levels describe everything: Meta-level Description

  • 1

Splicing ($<...>) Normal compilation +1 Quasi-quoting ([| ...|])

26 / 59 http://soft-dev.org/

slide-56
SLIDE 56

What works well? What works well?

1 src infos make debugging possible.

27 / 59 http://soft-dev.org/

slide-57
SLIDE 57

What works well? What works well?

1 src infos make debugging possible. 2 rename enables building huge, name-safe trees.

27 / 59 http://soft-dev.org/

slide-58
SLIDE 58

What works well? What works well?

1 src infos make debugging possible. 2 rename enables building huge, name-safe trees. 3 DSL layers work and are useful.

27 / 59 http://soft-dev.org/

slide-59
SLIDE 59

What works well? What works well?

1 src infos make debugging possible. 2 rename enables building huge, name-safe trees. 3 DSL layers work and are useful. 4 The compiler is surprisingly simple

27 / 59 http://soft-dev.org/

slide-60
SLIDE 60

What works well? What works well?

1 src infos make debugging possible. 2 rename enables building huge, name-safe trees. 3 DSL layers work and are useful. 4 The compiler is surprisingly simple (though

calculations with names make my head hurt).

27 / 59 http://soft-dev.org/

slide-61
SLIDE 61

What doesn’t work? What doesn’t work?

1 Delimiters are far too ugly for repeated use.

28 / 59 http://soft-dev.org/

slide-62
SLIDE 62

What doesn’t work? What doesn’t work?

1 Delimiters are far too ugly for repeated use. 2 Macro evaluation is top-to-bottom. DSLs can’t

validate e.g.:

$<<SQL>><SELECT c1 FROM t> $<<SQL>><CREATE TABLE t ( c2 STR )>

28 / 59 http://soft-dev.org/

slide-63
SLIDE 63

What doesn’t work? What doesn’t work?

1 Delimiters are far too ugly for repeated use. 2 Macro evaluation is top-to-bottom. DSLs can’t

validate e.g.:

$<<SQL>><SELECT c1 FROM t> $<<SQL>><CREATE TABLE t ( c2 STR )>

3 Syntax composition is nearly impossible.

28 / 59 http://soft-dev.org/

slide-64
SLIDE 64

What doesn’t work? What doesn’t work?

1 Delimiters are far too ugly for repeated use. 2 Macro evaluation is top-to-bottom. DSLs can’t

validate e.g.:

$<<SQL>><SELECT c1 FROM t> $<<SQL>><CREATE TABLE t ( c2 STR )>

3 Syntax composition is nearly impossible. 4 Performance for mildly complex DSLs is poor.

28 / 59 http://soft-dev.org/

slide-65
SLIDE 65

Where do we go from here?

29 / 59 http://soft-dev.org/

slide-66
SLIDE 66

Part III A different way

30 / 59 http://soft-dev.org/

slide-67
SLIDE 67

Language composition: two levels of challenge Language composition: two levels of challenge

31 / 59 http://soft-dev.org/

slide-68
SLIDE 68

Language composition: two levels of challenge Language composition: two levels of challenge

Tooling

31 / 59 http://soft-dev.org/

slide-69
SLIDE 69

Language composition: two levels of challenge Language composition: two levels of challenge

Tooling Language friction

31 / 59 http://soft-dev.org/

slide-70
SLIDE 70

Tooling challenges Tooling challenges

32 / 59 http://soft-dev.org/

slide-71
SLIDE 71

Tooling challenges Tooling challenges

PyHyp Python PHP

32 / 59 http://soft-dev.org/

slide-72
SLIDE 72

Tooling challenges Tooling challenges

PyHyp runtime syntax Python syntax runtime PHP syntax runtime

32 / 59 http://soft-dev.org/

slide-73
SLIDE 73

Tooling challenges Tooling challenges

Language boxes PyHyp runtime syntax Python syntax runtime PHP syntax runtime

32 / 59 http://soft-dev.org/

slide-74
SLIDE 74

Tooling challenges Tooling challenges

Composed meta-tracing VMs Language boxes PyHyp runtime syntax Python syntax runtime PHP syntax runtime

32 / 59 http://soft-dev.org/

slide-75
SLIDE 75

Syntax composition Syntax composition

PL X

<grammar>

expr::= ... term::= ... | ... | ... func ::= ...

PL Y

<program>

for (j : js) { doStuff(); } . . . 33 / 59 http://soft-dev.org/

slide-76
SLIDE 76

Syntax composition Syntax composition

PL X

<grammar>

expr::= ... term::= ... | ... | ... func ::= ...

PL Y

<program>

for (j : js) { doStuff(); } . . .

Parser

33 / 59 http://soft-dev.org/

slide-77
SLIDE 77

Syntax composition Syntax composition

PL X

<grammar>

expr::= ... term::= ... | ... | ... func ::= ...

PL Y

<program>

for (j : js) { doStuff(); } . . .

Parser

Parse Tree

33 / 59 http://soft-dev.org/

slide-78
SLIDE 78

Syntax composition Syntax composition

PL X

<grammar>

expr::= ... term::= ... | ... | ... func ::= ...

PL Y

<program>

for (j : js) { doStuff(); } . . .

LR

Parse Tree

33 / 59 http://soft-dev.org/

slide-79
SLIDE 79

Syntax composition Syntax composition

PL X

<grammar>

expr::= ... term::= ... | ... | ... func ::= ...

PL Y

<program>

for (j : js) { doStuff(); } . . .

LR

Parse Tree

U n d e f i n e d

33 / 59 http://soft-dev.org/

slide-80
SLIDE 80

Syntax composition Syntax composition

PL X

<grammar>

expr::= ... term::= ... | ... | ... func ::= ...

PL Y

<program>

for (j : js) { doStuff(); } . . .

Generalised

Parse Tree

33 / 59 http://soft-dev.org/

slide-81
SLIDE 81

Syntax composition Syntax composition

PL X

<grammar>

expr::= ... term::= ... | ... | ... func ::= ...

PL Y

<program>

for (j : js) { doStuff(); } . . .

Generalised

Parse Tree

A m b i g u

  • u

s

33 / 59 http://soft-dev.org/

slide-82
SLIDE 82

Syntax composition Syntax composition

PL X

<grammar>

expr::= ... term::= ... | ... | ... func ::= ...

PL Y

<program>

for (j : js) { doStuff(); } . . .

PEG

Parse Tree

33 / 59 http://soft-dev.org/

slide-83
SLIDE 83

Syntax composition Syntax composition

PL X

<grammar>

expr::= ... term::= ... | ... | ... func ::= ...

PL Y

<program>

for (j : js) { doStuff(); } . . .

PEG

Parse Tree

S h a d

  • w

s

33 / 59 http://soft-dev.org/

slide-84
SLIDE 84

The only choice? The only choice?

34 / 59 http://soft-dev.org/

slide-85
SLIDE 85

The only choice? The only choice?

SDE

34 / 59 http://soft-dev.org/

slide-86
SLIDE 86

The challenge The challenge

Challenge: SDE’s power + a text editor feel?

35 / 59 http://soft-dev.org/

slide-87
SLIDE 87

Eco demo Eco demo

36 / 59 http://soft-dev.org/

slide-88
SLIDE 88

Runtime composition Runtime composition

37 / 59 http://soft-dev.org/

slide-89
SLIDE 89

Runtime composition Runtime composition

PL X PL Y C/C++

Interpreter Interpreter

37 / 59 http://soft-dev.org/

slide-90
SLIDE 90

Runtime composition Runtime composition

PL X PL Y C/C++

Interpreter Interpreter

Too slow

37 / 59 http://soft-dev.org/

slide-91
SLIDE 91

Runtime composition Runtime composition

PL X PL Y C/C++

Interpreter Interpreter

JIT Compiler JIT Compiler

37 / 59 http://soft-dev.org/

slide-92
SLIDE 92

Runtime composition Runtime composition

PL X PL Y C/C++

Interpreter Interpreter

Too much engineering

JIT Compiler JIT Compiler

37 / 59 http://soft-dev.org/

slide-93
SLIDE 93

Runtime composition Runtime composition

PL X PL Y JVM/CLR

Interpreter Interpreter

JIT Compiler

37 / 59 http://soft-dev.org/

slide-94
SLIDE 94

Runtime composition Runtime composition

PL X PL Y JVM/CLR

Interpreter

JIT Compiler

Interpreter

Semantic mismatch

37 / 59 http://soft-dev.org/

slide-95
SLIDE 95

Runtime composition Runtime composition

37 / 59 http://soft-dev.org/

slide-96
SLIDE 96

Runtime composition Runtime composition

RPython

Interpreter Tracing JIT

PL Y PL X

Interpreters

Glue

PL Z

Meta-tracing

37 / 59 http://soft-dev.org/

slide-97
SLIDE 97

Meta-tracing translation with RPython Meta-tracing translation with RPython

Interpreter

38 / 59 http://soft-dev.org/

slide-98
SLIDE 98

Meta-tracing translation with RPython Meta-tracing translation with RPython

Interpreter

RPython translator

38 / 59 http://soft-dev.org/

slide-99
SLIDE 99

Meta-tracing translation with RPython Meta-tracing translation with RPython

Interpreter Optimised Interpreter JIT

RPython translator

38 / 59 http://soft-dev.org/

slide-100
SLIDE 100

Meta-tracing translation with RPython Meta-tracing translation with RPython

Interpreter Optimised Interpreter JIT

RPython translator

38 / 59 http://soft-dev.org/

slide-101
SLIDE 101

Meta-tracing translation with RPython Meta-tracing translation with RPython

Interpreter Optimised Interpreter JIT

RPython translator

38 / 59 http://soft-dev.org/

slide-102
SLIDE 102

Adding a JIT compiler to an RPython interpreter Adding a JIT compiler to an RPython interpreter

... pc := 0 while 1: instr := load_next_instruction(pc) if instr == POP: stack.pop() pc += 1 elif instr == BRANCH:

  • ff = load_branch_jump(pc)

pc += off elif ...: ...

Observation: interpreters are big loops.

39 / 59 http://soft-dev.org/

slide-103
SLIDE 103

Adding a JIT compiler to an RPython interpreter Adding a JIT compiler to an RPython interpreter

... pc := 0 while 1: jit_merge_point(pc) instr := load_next_instruction(pc) if instr == POP: stack.pop() pc += 1 elif instr == BRANCH:

  • ff = load_branch_jump(pc)

if off < 0: can_enter_jit(pc) pc += off elif ...: ...

Observation: interpreters are big loops.

39 / 59 http://soft-dev.org/

slide-104
SLIDE 104

RPython translation RPython translation

Interpreter Optimised Interpreter JIT

RPython translator

40 / 59 http://soft-dev.org/

slide-105
SLIDE 105

RPython translation RPython translation

Interpreter Language Interpreter Trace Interpreter

RPython translator

40 / 59 http://soft-dev.org/

slide-106
SLIDE 106

Tracing JITs Tracing JITs

User program (lang FL)

if x < 0: x = x + 1 else: x = x + 2 x = x + 3

41 / 59 http://soft-dev.org/

slide-107
SLIDE 107

Tracing JITs Tracing JITs

User program (lang FL) Trace when x is set to 6

if x < 0: x = x + 1 else: x = x + 2 x = x + 3 guard_type(x, int) guard_not_less_than(x, 0) guard_type(x, int) x = int_add(x, 2) guard_type(x, int) x = int_add(x, 3)

41 / 59 http://soft-dev.org/

slide-108
SLIDE 108

Tracing JITs Tracing JITs

User program (lang FL) Optimised trace

if x < 0: x = x + 1 else: x = x + 2 x = x + 3 guard_type(x, int) guard_not_less_than(x, 0) x = int_add(x, 5)

41 / 59 http://soft-dev.org/

slide-109
SLIDE 109

Meta-tracing VM components Meta-tracing VM components

Language Interpreter Trace Interpreter

  • 1. Start
  • 2. Detect hot loop
  • 3. Execute and trace
  • 4. Convert trace to

machine code

42 / 59 http://soft-dev.org/

slide-110
SLIDE 110

Meta-tracing JITs Meta-tracing JITs

FL Interpreter

program_counter = 0; stack = [] vars = {...} while True: jit_merge_point(program_counter) instr = load_instruction(program_counter) if instr == INSTR_VAR_GET: stack.push( vars[read_var_name_from_instruction()]) program_counter += 1 elif instr == INSTR_VAR_SET: vars[read_var_name_from_instruction()] = stack.pop() program_counter += 1 elif instr == INSTR_INT: stack.push(read_int_from_instruction()) program_counter += 1 elif instr == INSTR_LESS_THAN: rhs = stack.pop() lhs = stack.pop() if isinstance(lhs, int) and isinstance(rhs, int): if lhs < rhs: stack.push(True) else: stack.push(False) else: ... program_counter += 1 elif instr == INSTR_IF: result = stack.pop() if result == True: program_counter += 1 else: program_counter += read_jump_if_instruction() elif instr == INSTR_ADD: lhs = stack.pop() rhs = stack.pop() if isinstance(lhs, int) and isinstance(rhs, int): stack.push(lhs + rhs) else: ... program_counter += 1 43 / 59 http://soft-dev.org/

slide-111
SLIDE 111

Meta-tracing JITs Meta-tracing JITs

FL Interpreter

program_counter = 0; stack = [] vars = {...} while True: jit_merge_point(program_counter) instr = load_instruction(program_counter) if instr == INSTR_VAR_GET: stack.push( vars[read_var_name_from_instruction()]) program_counter += 1 elif instr == INSTR_VAR_SET: vars[read_var_name_from_instruction()] = stack.pop() program_counter += 1 elif instr == INSTR_INT: stack.push(read_int_from_instruction()) program_counter += 1 elif instr == INSTR_LESS_THAN: rhs = stack.pop() lhs = stack.pop() if isinstance(lhs, int) and isinstance(rhs, int): if lhs < rhs: stack.push(True) else: stack.push(False) else: ... program_counter += 1 43 / 59 http://soft-dev.org/

slide-112
SLIDE 112

Meta-tracing JITs Meta-tracing JITs

FL Interpreter User program (lang FL)

program_counter = 0; stack = [] vars = {...} while True: jit_merge_point(program_counter) instr = load_instruction(program_counter) if instr == INSTR_VAR_GET: stack.push( vars[read_var_name_from_instruction()]) program_counter += 1 elif instr == INSTR_VAR_SET: vars[read_var_name_from_instruction()] = stack.pop() program_counter += 1 elif instr == INSTR_INT: stack.push(read_int_from_instruction()) program_counter += 1 elif instr == INSTR_LESS_THAN: rhs = stack.pop() lhs = stack.pop() if isinstance(lhs, int) and isinstance(rhs, int): if lhs < rhs: stack.push(True) else: stack.push(False) else: ... program_counter += 1 if x < 0: x = x + 1 else: x = x + 2 x = x + 3 43 / 59 http://soft-dev.org/

slide-113
SLIDE 113

Meta-tracing JITs Meta-tracing JITs

FL Interpreter Initial trace

program_counter = 0; stack = [] vars = {...} while True: jit_merge_point(program_counter) instr = load_instruction(program_counter) if instr == INSTR_VAR_GET: stack.push( vars[read_var_name_from_instruction()]) program_counter += 1 elif instr == INSTR_VAR_SET: vars[read_var_name_from_instruction()] = stack.pop() program_counter += 1 elif instr == INSTR_INT: stack.push(read_int_from_instruction()) program_counter += 1 elif instr == INSTR_LESS_THAN: rhs = stack.pop() lhs = stack.pop() if isinstance(lhs, int) and isinstance(rhs, int): if lhs < rhs: stack.push(True) else: stack.push(False) else: ... program_counter += 1 v0 = <program_counter> v1 = <stack> v2 = <vars> v3 = load_instruction(v0) guard_eq(v3, INSTR_VAR_GET) v4 = dict_get(v2, "x") list_append(v1, v4) v5 = add(v0, 1) v6 = load_instruction(v5) guard_eq(v6, INSTR_INT) list_append(v1, 0) v7 = add(v5, 1) v8 = load_instruction(v7) guard_eq(v8, INSTR_LESS_THAN) v9 = list_pop(v1) v10 = list_pop(v1) guard_type(v9, int) guard_type(v10, int) guard_not_less_than(v9, v10) list_append(v1, False) v11 = add(v7, 1) v12 = load_instruction(v11) guard_eq(v12, INSTR_IF) v13 = list_pop(v1) guard_false(v13) ... 43 / 59 http://soft-dev.org/

slide-114
SLIDE 114

Meta-tracing JITs Meta-tracing JITs

Initial trace in full

v0 = <program_counter> v1 = <stack> v2 = <vars> v3 = load_instruction(v0) guard_eq(v3, INSTR_VAR_GET) v4 = dict_get(v2, "x") list_append(v1, v4) v5 = add(v0, 1) v6 = load_instruction(v5) guard_eq(v6, INSTR_INT) list_append(v1, 0) v7 = add(v5, 1) v8 = load_instruction(v7) guard_eq(v8, INSTR_LESS_THAN) v9 = list_pop(v1) v10 = list_pop(v1) guard_type(v9, int) guard_type(v10, int) guard_not_less_than(v9, v10) list_append(v1, False) v11 = add(v7, 1) v12 = load_instruction(v11) guard_eq(v12, INSTR_IF) v13 = list_pop(v1) guard_false(v13) v14 = add(v11, 2) v15 = load_instruction(v14) guard_eq(v15, INSTR_VAR_GET) v16 = dict_get(v2, "x") list_append(v1, v16) v17 = add(v14, 1) v18 = load_instruction(v17) guard_eq(v18, INSTR_INT) list_append(v1, 2) v19 = add(v17, 1) v20 = load_instruction(v19) guard_eq(v20, INSTR_ADD) v21 = list_pop(v1) v22 = list_pop(v1) guard_type(v21, int) guard_type(v22, int) v23 = add(v22, v21) list_append(v1, v23) v24 = add(v19, 1) v25 = load_instruction(v24) guard_eq(v25, INSTR_VAR_SET) v26 = list_pop(v1) dict_set(v2, "x", v26) v27 = add(v24, 1) v28 = load_instruction(v27) guard_eq(v28, INSTR_VAR_GET) v29 = dict_get(v2, "x") list_append(v1, v29) v30 = add(v27, 1) v31 = load_instruction(v30) guard_eq(v31, INSTR_INT) list_append(v1, 3) v32 = add(v30, 1) v33 = load_instruction(v32) guard_eq(v33, INSTR_ADD) v34 = list_pop(v1) v35 = list_pop(v1) guard_type(v34, int) guard_type(v35, int) v36 = add(v35, v34) list_append(v1, v36) v37 = add(v32, 1) v38 = load_instruction(v37) guard_eq(v38, INSTR_VAR_SET) v39 = list_pop(v1) dict_set(v2, "x", v39) v40 = add(v37, 1) 43 / 59 http://soft-dev.org/

slide-115
SLIDE 115

Trace optimisation (1) Trace optimisation (1)

Removing constants (from jit_merge_point)

v1 = <stack> v2 = <vars> v4 = dict_get(v2, "x") list_append(v1, v4) list_append(v1, 0) v9 = list_pop(v1) v10 = list_pop(v1) guard_type(v9, int) guard_type(v10, int) guard_not_less_than(v9, v10) list_append(v1, False) v13 = list_pop(v1) guard_false(v13) v16 = dict_get(v2, "x") list_append(v1, v16) list_append(v1, 2) v21 = list_pop(v1) v22 = list_pop(v1) guard_type(v21, int) guard_type(v22, int) v23 = add(v22, v21) list_append(v1, v23) v26 = list_pop(v1) dict_set(v2, "x", v26) v29 = dict_get(v2, "x") list_append(v1, v29) list_append(v1, 3) v34 = list_pop(v1) v35 = list_pop(v1) guard_type(v34, int) guard_type(v35, int) v36 = add(v35, v34) list_append(v1, v36) v39 = list_pop(v1) dict_set(v2, "x", v39) 44 / 59 http://soft-dev.org/

slide-116
SLIDE 116

Optimisation #2 & #3 Optimisation #2 & #3

List folded trace

v1 = <stack> v2 = <vars> v4 = dict_get(v2, "x") guard_type(v4, int) guard_not_less_than(v4, 0) v16 = dict_get(v2, "x") guard_type(v16, int) v23 = add(v16, 2) dict_set(v2, "x", v23) v29 = dict_get(v2, "x") guard_type(v29, int) v36 = add(v29, 3) dict_set(v2, "x", v36) 45 / 59 http://soft-dev.org/

slide-117
SLIDE 117

Optimisation #2 & #3 Optimisation #2 & #3

List folded trace Dict folded trace

v1 = <stack> v2 = <vars> v4 = dict_get(v2, "x") guard_type(v4, int) guard_not_less_than(v4, 0) v16 = dict_get(v2, "x") guard_type(v16, int) v23 = add(v16, 2) dict_set(v2, "x", v23) v29 = dict_get(v2, "x") guard_type(v29, int) v36 = add(v29, 3) dict_set(v2, "x", v36) v1 = <stack> v2 = <vars> v4 = dict_get(v2, "x") guard_type(v4, int) guard_not_less_than(v4, 0) v23 = add(v4, 2) guard_type(v23, int) v36 = add(v23, 3) dict_set(v2, "x", v36) 45 / 59 http://soft-dev.org/

slide-118
SLIDE 118

Optimisation #4 & #5 Optimisation #4 & #5

Type folded trace

v1 = <stack> v2 = <vars> v4 = dict_get(v2, "x") guard_type(v4, int) guard_not_less_than(v4, 0) v23 = add(v4, 2) v36 = add(v23, 3) dict_set(v2, "x", v36) 46 / 59 http://soft-dev.org/

slide-119
SLIDE 119

Optimisation #4 & #5 Optimisation #4 & #5

Type folded trace Arithmetic folded trace

v1 = <stack> v2 = <vars> v4 = dict_get(v2, "x") guard_type(v4, int) guard_not_less_than(v4, 0) v23 = add(v4, 2) v36 = add(v23, 3) dict_set(v2, "x", v36) v1 = <stack> v2 = <vars> v4 = dict_get(v2, "x") guard_type(v4, int) guard_not_less_than(v4, 0) v23 = add(v4, 5) dict_set(v2, "x", v23) 46 / 59 http://soft-dev.org/

slide-120
SLIDE 120

Optimisation #4 & #5 Optimisation #4 & #5

Type folded trace Arithmetic folded trace

v1 = <stack> v2 = <vars> v4 = dict_get(v2, "x") guard_type(v4, int) guard_not_less_than(v4, 0) v23 = add(v4, 2) v36 = add(v23, 3) dict_set(v2, "x", v36) v1 = <stack> v2 = <vars> v4 = dict_get(v2, "x") guard_type(v4, int) guard_not_less_than(v4, 0) v23 = add(v4, 5) dict_set(v2, "x", v23)

Trace optimisation: from 72 trace elements to 7.

46 / 59 http://soft-dev.org/

slide-121
SLIDE 121

Runtime composition recap Runtime composition recap

47 / 59 http://soft-dev.org/

slide-122
SLIDE 122

Runtime composition recap Runtime composition recap

RPython

Interpreter Tracing JIT

PL Y PL X

Interpreters

Glue

PL Z

Meta-tracing

47 / 59 http://soft-dev.org/

slide-123
SLIDE 123

Runtime composition recap Runtime composition recap

RPython

Interpreter Tracing JIT

Hippy PyPy

Interpreters

Glue

PyHyp

Meta-tracing

47 / 59 http://soft-dev.org/

slide-124
SLIDE 124

Composed Richards vs. other VMs Composed Richards vs. other VMs

Type VM CPython 2.7.7 9.475 ± 0.0127 HHVM 3.4.0 4.264 ± 0.0386 Mono HippyVM 0.250 ± 0.0008 PyPy 2.4.0 0.178 ± 0.0006 Composed Zend 5.5.13 9.070 ± 0.0361

48 / 59 http://soft-dev.org/

slide-125
SLIDE 125

Composed Richards vs. other VMs Composed Richards vs. other VMs

Type VM CPython 2.7.7 9.475 ± 0.0127 HHVM 3.4.0 4.264 ± 0.0386 Mono HippyVM 0.250 ± 0.0008 PyPy 2.4.0 0.178 ± 0.0006 Composed Zend 5.5.13 9.070 ± 0.0361 Composed PyHyp 0.335 ± 0.0012

48 / 59 http://soft-dev.org/

slide-126
SLIDE 126

Datatype conversion Datatype conversion

PHPObject PHPRoot PHPInt PHPFunc

49 / 59 http://soft-dev.org/

slide-127
SLIDE 127

Datatype conversion Datatype conversion

PHPObject PHPRoot PHPInt PHPFunc PyObject PyRoot PyInt PyFunc

49 / 59 http://soft-dev.org/

slide-128
SLIDE 128

Datatype conversion: primitive types Datatype conversion: primitive types

PHP Python

50 / 59 http://soft-dev.org/

slide-129
SLIDE 129

Datatype conversion: primitive types Datatype conversion: primitive types

2 : PHPInt PHP Python

50 / 59 http://soft-dev.org/

slide-130
SLIDE 130

Datatype conversion: primitive types Datatype conversion: primitive types

2 : PHPInt PHP Python 2 : PyInt

50 / 59 http://soft-dev.org/

slide-131
SLIDE 131

Datatype conversion: user types Datatype conversion: user types

PHP Python

51 / 59 http://soft-dev.org/

slide-132
SLIDE 132

Datatype conversion: user types Datatype conversion: user types

  • : PHPObject

PHP Python

51 / 59 http://soft-dev.org/

slide-133
SLIDE 133

Datatype conversion: user types Datatype conversion: user types

PyObject PyRoot PyInt PyFunc

51 / 59 http://soft-dev.org/

slide-134
SLIDE 134

Datatype conversion: user types Datatype conversion: user types

PyObject PyRoot PyInt PyFunc PyPHPAdapter

51 / 59 http://soft-dev.org/

slide-135
SLIDE 135

Datatype conversion: user types Datatype conversion: user types

PyObject PyRoot PyInt PyFunc PyPHPAdapter

php_obj : PHPObject

51 / 59 http://soft-dev.org/

slide-136
SLIDE 136

Datatype conversion: user types Datatype conversion: user types

  • : PHPObject

PHP Python

51 / 59 http://soft-dev.org/

slide-137
SLIDE 137

Datatype conversion: user types Datatype conversion: user types

  • : PHPObject

PHP Python :PyPHPAdapter

51 / 59 http://soft-dev.org/

slide-138
SLIDE 138

Datatype conversion: user types Datatype conversion: user types

  • : PHPObject

PHP Python :PyPHPAdapter

php_obj

51 / 59 http://soft-dev.org/

slide-139
SLIDE 139

Datatype conversion: user types Datatype conversion: user types

  • : PHPObject

PHP Python :PyPHPAdapter

php_obj

Immutable field

51 / 59 http://soft-dev.org/

slide-140
SLIDE 140

Friction Friction

A good composition needs to reduce friction.

52 / 59 http://soft-dev.org/

slide-141
SLIDE 141

Friction Friction

A good composition needs to reduce friction. Some examples:

  • Lexical scoping (or lack thereof) in PHP and Python (semantic

friction)

52 / 59 http://soft-dev.org/

slide-142
SLIDE 142

Friction Friction

A good composition needs to reduce friction. Some examples:

  • Lexical scoping (or lack thereof) in PHP and Python (semantic

friction)

  • PHP datatypes are immutable except for references and objects;

Python’s are largely mutable (semantic and performance friction)

52 / 59 http://soft-dev.org/

slide-143
SLIDE 143

Friction Friction

A good composition needs to reduce friction. Some examples:

  • Lexical scoping (or lack thereof) in PHP and Python (semantic

friction)

  • PHP datatypes are immutable except for references and objects;

Python’s are largely mutable (semantic and performance friction)

  • PHP has only dictionaries; Python has lists and dictionaries

(semantic friction)

52 / 59 http://soft-dev.org/

slide-144
SLIDE 144

Unipycation Unipycation

RPython

Interpreter Tracing JIT

Hippy PyPy

Interpreters

Glue

PyHyp

Meta-tracing

53 / 59 http://soft-dev.org/

slide-145
SLIDE 145

Unipycation Unipycation

RPython

Interpreter Tracing JIT

Pyrolog PyPy

Interpreters

Glue

Unipycation

Meta-tracing

53 / 59 http://soft-dev.org/

slide-146
SLIDE 146

Unipycation demo Unipycation demo

54 / 59 http://soft-dev.org/

slide-147
SLIDE 147

Absolute timing comparison Absolute timing comparison

VM Benchmark Python Prolog Python → Prolog CPython-SWI SmallFunc 0.125s

±0.007

0.257s

±0.002

28.893s

±0.227

L1A0R 2.924s

±0.284

7.352s

±0.048

9.310s

±0.084

L1A1R 4.184s

±0.038

18.890s

±0.111

20.865s

±0.067

NdL1A1R 7.531s

±0.080

18.643s

±0.197

667.682s

±6.895

TCons 264.415s

±2.250

48.819s

±0.252

2185.150s

±18.225

Lists 9.374s

±0.059

25.148s

±0.221

2207.304s

±16.073

Unipycation SmallFunc 0.001s

±0.000

0.006s

±0.001

0.001s

±0.000

L1A0R 0.085s

±0.000

0.086s

±0.000

0.087s

±0.000

L1A1R 0.112s

±0.000

0.114s

±0.000

0.115s

±0.000

NdL1A1R 0.500s

±0.003

0.548s

±0.085

2.674s

±0.012

TCons 6.053s

±0.288

2.444s

±0.003

36.069s

±0.225

Lists 0.845s

±0.002

1.416s

±0.003

5.056s

±0.035

Jython-tuProlog SmallFunc 0.088s

±0.003

3.050s

±0.053

52.294s

±0.475

L1A0R 1.078s

±0.009

206.590s

±3.846

199.963s

±2.476

L1A1R 2.145s

±0.232

293.311s

±5.691

294.781s

±6.193

NdL1A1R 7.939s

±0.457

1857.687s

±5.169

1990.985s

±15.071

TCons 543.347s

±3.289

8014.477s

±17.710

8202.362s

±24.904

Lists 5.661s

±0.046

6981.873s

±18.795

5577.322s

±15.754 55 / 59 http://soft-dev.org/

slide-148
SLIDE 148

Relative timing comparison Relative timing comparison

VM Benchmark Python→Prolog Python Python→Prolog Prolog Python→Prolog Unipycation CPython-SWI SmallFunc 231.770×

±13.136

112.567×

±1.242

27821.079×

±2331.665

L1A0R 3.184×

±0.300

1.266×

±0.014

107.591×

±0.995

L1A1R 4.987×

±0.049

1.105×

±0.007

181.899×

±0.590

NdL1A1R 88.654×

±1.368

35.814×

±0.554

249.737×

±2.922

TCons 8.264×

±0.101

44.760×

±0.453

60.583×

±0.637

Lists 235.459×

±2.314

87.772×

±1.017

436.609×

±4.415

Unipycation SmallFunc 1.295×

±0.105

0.182×

±0.054

1.000× L1A0R 1.020×

±0.002

1.012×

±0.002

1.000× L1A1R 1.025×

±0.002

1.002×

±0.003

1.000× NdL1A1R 5.349×

±0.045

4.879×

±0.924

1.000× TCons 5.959×

±0.282

14.756×

±0.092

1.000× Lists 5.982×

±0.045

3.569×

±0.026

1.000× Jython-tuProlog SmallFunc 592.904×

±19.517

17.143×

±0.338

50354.204×

±4341.413

L1A0R 185.460×

±2.818

0.968×

±0.021

2310.844×

±28.093

L1A1R 137.427×

±14.537

1.005×

±0.028

2569.873×

±52.847

NdL1A1R 250.776×

±14.666

1.072×

±0.009

744.699×

±6.726

TCons 15.096×

±0.106

1.023×

±0.004

227.409×

±1.592

Lists 985.149×

±8.674

0.799×

±0.003

1103.206×

±8.338 56 / 59 http://soft-dev.org/

slide-149
SLIDE 149

What can we use this for? What can we use this for?

57 / 59 http://soft-dev.org/

slide-150
SLIDE 150

What can we use this for? What can we use this for?

First-class languages

57 / 59 http://soft-dev.org/

slide-151
SLIDE 151

What can we use this for? What can we use this for?

First-class languages Language migration

57 / 59 http://soft-dev.org/

slide-152
SLIDE 152

Thanks to our funders Thanks to our funders

  • EPSRC: COOLER and Lecture.
  • Oracle: various.

58 / 59 http://soft-dev.org/

slide-153
SLIDE 153

Thanks for listening Thanks for listening

  • : PHPObject

PHP Python :PyPHPAdapter

php_obj

Immutable field RPython

Interpreter Tracing JIT

Hippy PyPy

Interpreters Glue

PyHyp

Meta-tracing

http://soft-dev.org/

59 / 59 http://soft-dev.org/