HOTPATH VM. An Effective JIT Compiler for Resource-constrained Devices

Transcription

1 HOTPATH VM An Effective JIT Compiler for Resource-constrained Devices based on the paper by Andreas Gal, Christian W. Probst and Michael Franz, VEE 06

2 INTRODUCTION

3 INTRODUCTION Most important factor: speed

4 INTRODUCTION Most important factor: speed Overlooked: power consumption, memory footprint

5 INTRODUCTION Most important factor: speed Overlooked: power consumption, memory footprint Embedded JIT compilers (simpler algorithms)

6 JAMVM

7 JAMVM GNU classpath-based

8 JAMVM GNU classpath-based Own intermediate representation

9 JAMVM GNU classpath-based Own intermediate representation HotpathVM as add-on

10 JAMVM GNU classpath-based Own intermediate representation HotpathVM as add-on Integrated w/ 20 lines of code

11 HOTPATHVM

12 HOTPATHVM 150 kb memory footprint

13 HOTPATHVM 150 kb memory footprint Searching for hot code by tracing

14 HOTPATHVM 150 kb memory footprint Searching for hot code by tracing Cold code for interpreter

15 HOTPATHVM 150 kb memory footprint Searching for hot code by tracing Cold code for interpreter Optimizing and compiling traces

16 TRACE SELECTION

17 TRACE SELECTION Problem: bytecode is not purely sequential

18 TRACE SELECTION Problem: bytecode is not purely sequential Solution: Software Trace Scheduling

19 TRACE SELECTION Problem: bytecode is not purely sequential Solution: Software Trace Scheduling Limited to loops

20 IDENTIFYING LOOP HEADERS

21 IDENTIFYING LOOP HEADERS Backward jump record destination

22 IDENTIFYING LOOP HEADERS Backward jump record destination Assumption: bytecode is forward

23 IDENTIFYING LOOP HEADERS Backward jump record destination Assumption: bytecode is forward Counter for backward jumps

24 IDENTIFYING LOOP HEADERS Backward jump record destination Assumption: bytecode is forward Counter for backward jumps Saved in intermediate representation

25 IDENTIFYING LOOP HEADERS Backward jump record destination Assumption: bytecode is forward Counter for backward jumps Saved in intermediate representation Record after threshold is exceeded

26 TRACE EXAMPLE

27 TRACE EXAMPLE public static void main(string args[]) { int i, k = 0; for (i = 0; i < 1000, ++i) ++k; System.out.println(k); }

28 TRACE EXAMPLE public static void main(string args[]) { int i, k = 0; for (i = 0; i < 1000, ++i) ++k; System.out.println(k); } A: B: iconst_0 istore_2 iconst_0 istore_1 iload_1 sipush 1000 if_icmpge B iinc 2,1 iinc 1,1 goto A getstatic System.out iload_2 invokevirtual println(int) return

29 TRACE EXAMPLE public static void main(string args[]) { int i, k = 0; for (i = 0; i < 1000, ++i) ++k; hotspot System.out.println(k); } A: B: iconst_0 istore_2 iconst_0 istore_1 iload_1 sipush 1000 if_icmpge B iinc 2,1 iinc 1,1 goto A getstatic System.out iload_2 invokevirtual println(int) return

30 TRACE EXAMPLE public static void main(string args[]) { int i, k = 0; for (i = 0; i < 1000, ++i) ++k; hotspot System.out.println(k); } A: B: iconst_0 istore_2 iconst_0 istore_1 iload_1 sipush 1000 if_icmpge B iinc 2,1 hotpath iinc 1,1 goto A getstatic System.out iload_2 invokevirtual println(int) return

31 RECORDING TRACES

32 RECORDING TRACES Meta block after each instruction

33 RECORDING TRACES Meta block after each instruction Second stop-loss threshold

34 RECORDING TRACES Meta block after each instruction Second stop-loss threshold No overhead without tracer

35 RECORDING TRACES Meta block after each instruction Second stop-loss threshold No overhead without tracer Guards for branches

36 GUARDS

37 GUARDS Ensures following the hotpath

38 GUARDS Ensures following the hotpath Returns control to VM on failure

39 GUARDS Ensures following the hotpath Returns control to VM on failure Reset values from intermediate representation

40 BRANCHES

41 BRANCHES Multiple hot paths in loop

42 BRANCHES Multiple hot paths in loop Lazy recording

43 BRANCHES Multiple hot paths in loop Lazy recording Secondary traces (like Dynamo)

44 BRANCHES Multiple hot paths in loop Lazy recording Secondary traces (like Dynamo) Updates guard instruction

45 BRANCHES Multiple hot paths in loop Lazy recording Secondary traces (like Dynamo) Updates guard instruction

46 STOP CONDITIONS

47 STOP CONDITIONS Successful recording

48 STOP CONDITIONS Successful recording Aborted recording on rare events

49 STOP CONDITIONS Successful recording Aborted recording on rare events Exceptions

50 STOP CONDITIONS Successful recording Aborted recording on rare events Exceptions Native method invocation

51 COMPILING TRACES

52 COMPILING TRACES Compiling complete traces only

53 COMPILING TRACES Compiling complete traces only Assumption: trace is executed repeatedly

54 COMPILING TRACES Compiling complete traces only Assumption: trace is executed repeatedly

55 COMPILING TRACES Compiling complete traces only Assumption: trace is executed repeatedly Stack Deconstruction

56 COMPILING TRACES Compiling complete traces only Assumption: trace is executed repeatedly Stack Deconstruction Code Analysis

57 COMPILING TRACES Compiling complete traces only Assumption: trace is executed repeatedly Stack Deconstruction Code Analysis Code Generation

58 STACK DECONSTRUCTION

59 STACK DECONSTRUCTION Generates SSA-based IR

60 STACK DECONSTRUCTION Generates SSA-based IR φ pseudo instruction

61 STACK DECONSTRUCTION x = 5 Generates SSA-based IR if x > 5 φ pseudo instruction x = 5 x = 7 x = 7

62 STACK DECONSTRUCTION x1 x = 5 Generates SSA-based IR if x1 x > 5 φ pseudo instruction x2 x = 5 x3 x = 7 x4 = x φ(x2,x3) = 7

63 STACK DECONSTRUCTION x1 x = 5 Generates SSA-based IR if x1 x > 5 φ pseudo instruction Flags loop invariant code x2 x = 5 x3 x = 7 x4 = x φ(x2,x3) = 7

64 STACK DECONSTRUCTION x1 x = 5 Generates SSA-based IR if x1 x > 5 φ pseudo instruction Flags loop invariant code x2 x = 5 x3 x = 7 x4 = x φ(x2,x3) = 7

65 CODE ANALYSIS

66 CODE ANALYSIS Constant Propagation

67 CODE ANALYSIS Constant Propagation Loop Peeling

68 CODE ANALYSIS Constant Propagation Loop Peeling Common subexpression elimination

69 CODE ANALYSIS Constant Propagation Loop Peeling Common subexpression elimination Invariant code motion

70 CODE GENERATION

71 CODE GENERATION Generating code backwards

72 CODE GENERATION Generating code backwards Enables to emit special machine code

73 CODE GENERATION Generating code backwards Enables to emit special machine code Constant Folding

74 SIDE EXITS

75 SIDE EXITS Leaving the hot path

76 SIDE EXITS Leaving the hot path Generator generates compensation code

77 SIDE EXITS Leaving the hot path Generator generates compensation code Special case: inline methods

78 SIDE EXITS Leaving the hot path Generator generates compensation code Special case: inline methods Writing back current state

79 WALKTHROUGH

80 WALKTHROUGH Java Code

81 WALKTHROUGH compilation Java Code

82 WALKTHROUGH compilation Java Code Bytecode

83 WALKTHROUGH Bytecode

84 WALKTHROUGH stack deconstruction Bytecode

85 WALKTHROUGH stack deconstruction Bytecode SSA

86 WALKTHROUGH stack deconstruction Bytecode SSA Flag invariant code

87 WALKTHROUGH SSA

88 WALKTHROUGH analysis optimization SSA

89 WALKTHROUGH analysis optimization SSA SSA

90 WALKTHROUGH analysis optimization SSA SSA Constant Propagation Loop Peeling Common subexpression elimination Invariant code motion

91 WALKTHROUGH SSA

92 WALKTHROUGH generation SSA

93 WALKTHROUGH generation SSA Machine code

94 WALKTHROUGH generation SSA Machine code Constant Folding

95 BENCHMARKS

96 BENCHMARKS execution time in s LU SOR FFT SCR Linpack NumSort MonteCarlo Java VM JamVM HotpathVM Hotspot VM

101 BENCHMARKS

102 BENCHMARKS speedup factor LU SOR FFT SCR Linpack NumSort MonteCarlo Java VM JamVM HotpathVM Hotspot VM

107 OUTLOOK

108 OUTLOOK Trace Merging

109 OUTLOOK Trace Merging Support additional architectures (ARM)

110 OUTLOOK Trace Merging Support additional architectures (ARM) Use outside embedded environment

111 HOTPATH VM An Effective JIT Compiler for Resource-constrained Devices