Adam Retter

adam@evolvedbinary.com
 

Declarative Amsterdam @ Online
2020-10-09


@adamretter

Compiling XQuery
to
Native Machine Code

About Me

  • Director and CTO of Evolved Binary

    • XQuery / XSLT / Schema / RelaxNG

    • Scala / Java / C++ / Rust*

    • Concurrency and Scalability

  • Creator of FusionDB multi-model database (5 yrs.)

  • Contributor to Facebook's RocksDB (5 yrs.)

  • Core contributor to eXist-db XML Database (15 yrs.)

  • Founder of EXQuery, and creator of RESTXQ

  • Was a W3C XQuery WG Invited expert

  • Me: www.adamretter.org.uk

Framing the Problem

  • FusionDB inherited eXist-db's XQuery Engine

    • Several bugs/issues

    • EOL - Parser implemented in Antlr 2 (1997-2005)

      • Mixes concerns - Grammar with Java inlined in same file

      • Hard to optimise - lack of clean AST for transform/rewrite

      • Hard to maintain - lack of docs / support

    • We made some small incremental improvements

  • Performance Mismatch

    • XQuery Engine is high-level (Java)

      • Operates over an inefficient DOM abstraction

    • FDB Core Storage is low-level (C++)

Resolving the Problem

  • We need/want a new XQuery Engine for FusionDB

    • Must be able to support all XQuery Grammars

      • Query, Full-Text, Update, Scripting, etc.

    • Must offer excellent error reporting with context

      • Where did the error occur?

      • Why did it occur?

      • Should suggest possible fixes (e.g. Rust Compiler)

    • Must be able to exploit FDB storage performance

    • Should exploit modern techniques

    • Should be able to utilise modern hardware

    • Would like a clean AST without inline code

      • Enables AST -> AST transformations (Rewrite optimisation)

      • Enables Target Independence

Anatomy of an XQuery Engine

Parser Concerns

  • Many Different Algorithms

    • LR - LALR / GLR

    • LL - (Predictive) Recursive Descent / ALL(*) / PEG

    • Others - Earley / Leo-Earley / Pratt / Combinators / Marpa

  • Implementation trade-offs

    • Parser Generators, e.g. Flex/Bison

      • Pro - Speed of prototyping/impl.

      • Cons - Not just EBNF - Learn additional Grammar Syntax / Poor Error Messages

    • Parser Combinators, e.g. FastParse

      • Pro - Speed of prototyping/impl.

      • Con - Defined in code

    • Bespoke

      • Pro - Complete Control

      • Con - Huge amount of work

Our New XQuery Parser

  • Bespoke

    • We really care about error messages

    • We really care about performance

    • We wanted a clean AST

    • Can also produce a CST

  • Separate Lexer and Parser

  • Lexer is Intelligent

    • Data type conversion, e.g. IntegerLiteral -> native integer type

    • Recognises Sub-tokens, e.g. PredefinedEnityRef within StringLiteral

  • Predictive Recursive Descent

  • Implemented in Rust

Our New XQuery Compiler

  • Mostly it's just - LLVM!

    • Previously called: Low Level Virtual Machine

    • Open Source set of compiler and toolchain technologies

    • Front-end for language and Back-end for target

      • Separated by IR (Intermediate Representation) code

  • Our New XQuery Parser produces an AST

    • Our Language Front-end generates IR from the AST

Anatomy of our New XQuery Engine

declare function local:fib($num as xs:integer) as xs:integer {
    if ($num = 0) then
    	0
    else if ($num = 1) then
    	1
    else
    	local:fib($num - 1) + local:fib($num - 2)
};

local:fib(33)

Computing nth Fibonacci Number
  - XQuery Code

Computing nth Fibonacci Number
  - Lexer Output

DeclareWord(TokenContext(Extent(0, 7), Location(1, 1)))
FucntionWord(TokenContext(Extent(8, 16), Location(1, 9)))
QName(TokenContextWithSubTokens { token_context: TokenContext(Extent(17, 26), Location(1, 18)), sub_tokens: [NCName(TokenContext(Extent(17, 22), Location(1, 18))), NCName(TokenContext(Extent(23, 26), Location(1, 24)))] })
LeftParenChar(TokenContext(Extent(26, 27), Location(1, 27)))
DollarSignChar(TokenContext(Extent(27, 28), Location(1, 28)))
NCName(TokenContext(Extent(28, 31), Location(1, 29)))
NCName(TokenContext(Extent(32, 34), Location(1, 33)))
QName(TokenContextWithSubTokens { token_context: TokenContext(Extent(35, 45), Location(1, 36)), sub_tokens: [NCName(TokenContext(Extent(35, 37), Location(1, 36))), NCName(TokenContext(Extent(38, 45), Location(1, 39)))] })
RightParenChar(TokenContext(Extent(45, 46), Location(1, 46)))
NCName(TokenContext(Extent(47, 49), Location(1, 48)))
QName(TokenContextWithSubTokens { token_context: TokenContext(Extent(50, 60), Location(1, 51)), sub_tokens: [NCName(TokenContext(Extent(50, 52), Location(1, 51))), NCName(TokenContext(Extent(53, 60), Location(1, 54)))] })
LeftCurlyBracketChar(TokenContext(Extent(61, 62), Location(1, 62)))
IfWord(TokenContext(Extent(75, 77), Location(2, 13)))
LeftParenChar(TokenContext(Extent(78, 79), Location(2, 16)))
DollarSignChar(TokenContext(Extent(79, 80), Location(2, 17)))
NCName(TokenContext(Extent(80, 83), Location(2, 18)))
EqualsSignChar(TokenContext(Extent(84, 85), Location(2, 22)))
IntegerLiteral(TokenContextWithBigUint { token_context: TokenContext(Extent(86, 87), Location(2, 24)), value: BigUint { data: [] } })
RightParenChar(TokenContext(Extent(87, 88), Location(2, 25)))
ThenWork(TokenContext(Extent(89, 93), Location(2, 27)))
IntegerLiteral(TokenContextWithBigUint { token_context: TokenContext(Extent(110, 111), Location(3, 17)), value: BigUint { data: [] } })
ElseWord(TokenContext(Extent(124, 128), Location(4, 13)))
IfWord(TokenContext(Extent(129, 131), Location(4, 18)))
LeftParenChar(TokenContext(Extent(132, 133), Location(4, 21)))
DollarSignChar(TokenContext(Extent(133, 134), Location(4, 22)))
NCName(TokenContext(Extent(134, 137), Location(4, 23)))
EqualsSignChar(TokenContext(Extent(138, 139), Location(4, 27)))
IntegerLiteral(TokenContextWithBigUint { token_context: TokenContext(Extent(140, 141), Location(4, 29)), value: BigUint { data: [1] } })
RightParenChar(TokenContext(Extent(141, 142), Location(4, 30)))
ThenWork(TokenContext(Extent(143, 147), Location(4, 32)))
IntegerLiteral(TokenContextWithBigUint { token_context: TokenContext(Extent(164, 165), Location(5, 17)), value: BigUint { data: [1] } })
ElseWord(TokenContext(Extent(178, 182), Location(6, 13)))
QName(TokenContextWithSubTokens { token_context: TokenContext(Extent(199, 208), Location(7, 17)), sub_tokens: [NCName(TokenContext(Extent(199, 204), Location(7, 17))), NCName(TokenContext(Extent(205, 208), Location(7, 23)))] })
LeftParenChar(TokenContext(Extent(208, 209), Location(7, 26)))
DollarSignChar(TokenContext(Extent(209, 210), Location(7, 27)))
NCName(TokenContext(Extent(210, 213), Location(7, 28)))
HyphenMinusChar(TokenContext(Extent(214, 215), Location(7, 32)))
IntegerLiteral(TokenContextWithBigUint { token_context: TokenContext(Extent(216, 217), Location(7, 34)), value: BigUint { data: [1] } })
RightParenChar(TokenContext(Extent(217, 218), Location(7, 35)))
PlusSignChar(TokenContext(Extent(219, 220), Location(7, 37)))
QName(TokenContextWithSubTokens { token_context: TokenContext(Extent(221, 230), Location(7, 39)), sub_tokens: [NCName(TokenContext(Extent(221, 226), Location(7, 39))), NCName(TokenContext(Extent(227, 230), Location(7, 45)))] })
LeftParenChar(TokenContext(Extent(230, 231), Location(7, 48)))
DollarSignChar(TokenContext(Extent(231, 232), Location(7, 49)))
NCName(TokenContext(Extent(232, 235), Location(7, 50)))
HyphenMinusChar(TokenContext(Extent(236, 237), Location(7, 54)))
IntegerLiteral(TokenContextWithBigUint { token_context: TokenContext(Extent(238, 239), Location(7, 56)), value: BigUint { data: [2] } })
RightParenChar(TokenContext(Extent(239, 240), Location(7, 57)))
RightCurlyBracketChar(TokenContext(Extent(249, 250), Location(8, 9)))
SemicolonChar(TokenContext(Extent(250, 251), Location(8, 10)))
QName(TokenContextWithSubTokens { token_context: TokenContext(Extent(260, 269), Location(9, 9)), sub_tokens: [NCName(TokenContext(Extent(260, 265), Location(9, 9))), NCName(TokenContext(Extent(266, 269), Location(9, 15)))] })
LeftParenChar(TokenContext(Extent(269, 270), Location(9, 18)))
IntegerLiteral(TokenContextWithBigUint { token_context: TokenContext(Extent(270, 272), Location(9, 19)), value: BigUint { data: [33] } })
RightParenChar(TokenContext(Extent(272, 273), Location(9, 21)))

Computing nth Fibonacci Number
  - Parser Output
     (1 of 2)

Computing nth Fibonacci Number
  - Parser Output
     (2 of 2)

Generating IR for LLVM

  • We use a Visitor Pattern over our AST

  • LLVM provides an API (C++)

    • Various bindings for other languages

    • Starting point is the IRBuilder class

    • IRBuilder adds instructions to your Module class

    • The NamedValues class holds function parameters, variables, etc.

  • LLVM provides many possible Optimisations for IR

    • See: https://llvm.org/docs/Passes.html

    • Of particular interest (for XQuery):

      • Tail Call Elimination

      • Dead Code/Argument/Type/Global/Loop Elimination

Computing nth Fibonacci Number - IR

;; defines our fibonacci function as local_fib
;; t1 = $num
define i64 @"local_fib"(i64 %".1")
{
local_fib_body:
  ;; t3 = ($num = 0)
  %".3" = icmp eq i64 %".1", 0
  ;; if t3 then goto local_fib_body.ifbody1 else goto local_fib_body.ifelsebody1
  br i1 %".3", label %"local_fib_body.ifbody1", label %"local_fib_body.ifelsebody1"

local_fib_body.ifbody1:
  ;; return 0
  ret i64 0

local_entry.ifelsebody1:
  ;; t4 = ($num = 1)
  %".4" = icmp eq i64 %".1", 1
  ;; if t4 then goto local_fib_body.ifbody2 else goto local_fib_body.ifelsebody2
  br i1 %".4", label %"local_fib_body.ifbody2", label %"local_fib_body.ifelsebody2"

local_fib_body.ifbody2:
  ;; return 1
  ret i64 1

local_entry.ifelsebody2:
  ;; t6 = ($num - 1)
  %".6" = sub i64 %".1", 1
  ;; t7 = ($num - 2)
  %".7" = sub i64 %".1", 2
  ;; t8 = local_fib( t6 ) ; ie local_fib( n - 1 )
  %".8" = call i64 (i64) @"local_fib"(i64 %".6")
  ;; t9 = local_fib( t7 ) ; ie local_fib( n - 2 )
  %".9" = call i64 (i64) @"local_fib"(i64 %".7")
  ;; t10 = (t8 + t9)
  %".10" = add i64 %".8", %".9"
  ;; return t10
  ret i64 %".10"
}

Where we are at today...

  • Still early stage!

  • Lexer is OK

  • Parser is in-progress

  • We have experimental IR Generation and JIT

  • Lots still to do...

  • Thinking about the future...

    • LLVM can target WebAssembly

      • Then we can run/call XQuery from Web-browser and Node.js

    • LLVM can target the GPU

fusiondb.com

Compiling XQuery to Native Machine Code

By Adam Retter

Compiling XQuery to Native Machine Code

Presentation given for Declarative Amsterdam - 9 October 2020 - Online

  • 1,350