Introduction to JVM JIT Optimizations

Transcript

1. Introduction to JVM Introduction to JVM JIT Optimizations JIT Optimizations Diego Parra | @dpsoft | Core Team @ Kamon
2. Agenda Agenda What is JIT? Tiered Compilation Optimizations Conclusion
3. JIT JIT Just-In-Time compilation Compilation work happens during application execution JIT'ing requires Profiling for more efficient optimization Because you don't want JIT everything Aggressively optimize based on profile information Tiered compilation Runtime Overhead?
4. JIT JIT Just-In-Time compilation Compilation work happens during application execution JIT'ing requires Profiling for more efficient optimization Because you don't want JIT everything Aggressively optimize based on profile information Tiered compilation Runtime Overhead?
5. JIT Lifecycle JIT Lifecycle
6. JIT Compilers JIT Compilers C1 Client C2 Server VS Fast compiler Invocation count > 1500 Produces Compilations quickly Generated code runs Slow Profile is about counters: https://github.com/dmlloyd/openjdk/blob/jdk8u/jdk8u/hotspot/src/share/vm/oops/methodCounters.hpp Smart compiler Invocation count > 10000 Produces Compilations slowly Generated code runs Fast 10x faster than Interpreter 2x faster than C1
7. JIT Compilers JIT Compilers C1 Client C2 Server VS Fast compiler Invocation count > 1500 Produces Compilations quickly Generated code runs Slow Profile is about counters: https://github.com/dmlloyd/openjdk/blob/jdk8u/jdk8u/hotspot/src/share/vm/oops/methodCounters.hpp Smart compiler Invocation count > 10000 Produces Compilations slowly Generated code runs Fast Profiled Guided Speculative 10x faster than Interpreter 2x faster than C1
8. Tiered Compilation Tiered Compilation Available in Java 7 Default in Java 8 Client Startup Server Performance Tiered : https://github.com/dmlloyd/openjdk/blob/jdk8u/jdk8u/hotspot/src/share/vm/runtime/advancedThresholdPolicy.hpp#L35-L158
9. Tiered Compilation Tiered Compilation Available in Java 7 Default in Java 8 Client Startup Server Performance Best of Both Worlds - C1 + C2 Tiered : https://github.com/dmlloyd/openjdk/blob/jdk8u/jdk8u/hotspot/src/share/vm/runtime/advancedThresholdPolicy.hpp#L35-L158
10. Common Transitions Common Transitions
11. Common Transitions Common Transitions Common Transitions Patterns : https://www.slideshare.net/maddocig/tiered Notes: https://github.com/dmlloyd/openjdk/blob/jdk8u/jdk8u/hotspot/src/share/vm/runtime/advancedThresholdPolicy.cpp
12. JIT Optimizations JIT Optimizations Compiler tactics Compiler tactics delayed compilation tiered compilation on-stack replacement delayed reoptimization program dependence graph representation static single assignment representation Speculative (profile-based) Speculative (profile-based) optimistic nullness assertions optimistic type assertions optimistic type strengthening optimistic array length strengthening untaken branch pruning optimistic N-morphic inlining branch frequency prediction call frequency prediction Proof-based techniques Proof-based techniques exact type inference memory value inference memory value tracking constant folding reassociation operator strength reduction null check elimination type test strength reduction type test elimination algebraic simplification common subexpression elimination integer range typing Flow-sensitive rewrites Flow-sensitive rewrites conditional constant propagation dominating test detection flow-carried type narrowing dead code elimination Text and much more..
13. Null check Elimination Null check Elimination
14. Null check Elimination Null check Elimination In equivalent machine code
15. Null check Elimination Null check Elimination In equivalent machine code
16. Null check Elimination Null check Elimination In equivalent machine code
17. Null check Elimination Null check Elimination Into the rabbit hole -XX:+PrintCompilation
18. Null check Elimination Null check Elimination Into the rabbit hole -XX:+PrintCompilation Compilation Level
19. Null check Elimination Null check Elimination Into the rabbit hole -XX:+PrintCompilation Compilation Level
20. Null check Elimination Null check Elimination Into the rabbit hole -XX:+PrintCompilation Compilation Level
21. Null check Elimination Null check Elimination Into the rabbit hole -XX:+PrintCompilation Compilation Level
22. Null check Elimination Null check Elimination Into the rabbit hole -XX:+PrintCompilation Compilation Level Deoptimizations
23. Null check Elimination Null check Elimination Into the rabbit hole -XX:+UnlockDiagnosticVMOptions -XX:+PrintAssembly -XX:CompileCommand=print,NullCheck::isPowerOfTwo
24. Null check Elimination Null check Elimination Into the rabbit hole -XX:+UnlockDiagnosticVMOptions -XX:+PrintAssembly -XX:CompileCommand=print,NullCheck::isPowerOfTwo Someone said Assembly?
25. Null check Elimination Null check Elimination Into the rabbit hole -XX:+UnlockDiagnosticVMOptions -XX:+PrintAssembly -XX:CompileCommand=print,NullCheck::isPowerOfTwo Someone said Assembly?
26. Null check Elimination Null check Elimination Into the rabbit hole
27. Null check Elimination Null check Elimination Into the rabbit hole
28. Null check Elimination Null check Elimination Into the rabbit hole implicit exception: dispatches to 0x00007fdff5318274
29. Null check Elimination Null check Elimination JVM uses JVM uses SIGSEGV SIGSEGV as control flow as control flow Into the rabbit hole Handle Signals: https://docs.oracle.com/javase/8/docs/technotes/guides/troubleshoot/signals001.html#CIHBBDED Speculative Optimization Speculative Optimization Faster Faster IF IF not null are encountered (Uncommon Trap) not null are encountered (Uncommon Trap)
30. Null check Elimination Null check Elimination JVM uses JVM uses SIGSEGV SIGSEGV as control flow as control flow Into the rabbit hole Handle Signals: https://docs.oracle.com/javase/8/docs/technotes/guides/troubleshoot/signals001.html#CIHBBDED Speculative Optimization Speculative Optimization Faster Faster IF IF not null are encountered (Uncommon Trap) not null are encountered (Uncommon Trap) Implicit Null Check? Field and Array access is null checked
31. Method Inlining Method Inlining
32. Method Inlining Method Inlining
33. Method Inlining Method Inlining -XX:+PrintInlining
34. Method Inlining Method Inlining -XX:+PrintInlining
35. Method Inlining Method Inlining -XX:+PrintInlining
36. Method Inlining Method Inlining Inlining is the most important optimization Combine caller and callee into one unit Expands the scope for other optimizations Bytecode size < 35 bytes Bytecode size < 325 bytes for inlining hot methods
37. Method Inlining Method Inlining Inlining is the most important optimization Combine caller and callee into one unit Expands the scope for other optimizations Bytecode size < 35 bytes Bytecode size < 325 bytes for inlining hot methods What about virtual call inlining?
38. Method Inlining Method Inlining Inlining is the most important optimization Combine caller and callee into one unit Expands the scope for other optimizations Bytecode size < 35 bytes Bytecode size < 325 bytes for inlining hot methods What about virtual call inlining?
39. Method Inlining Method Inlining Inlining is the most important optimization Combine caller and callee into one unit Expands the scope for other optimizations Bytecode size < 35 bytes Bytecode size < 325 bytes for inlining hot methods What about virtual call inlining? Monomorphic
40. Method Inlining Method Inlining Inlining is the most important optimization Combine caller and callee into one unit Expands the scope for other optimizations Bytecode size < 35 bytes Bytecode size < 325 bytes for inlining hot methods What about virtual call inlining? Monomorphic Bimorphic
41. Method Inlining Method Inlining Inlining is the most important optimization Combine caller and callee into one unit Expands the scope for other optimizations Bytecode size < 35 bytes Bytecode size < 325 bytes for inlining hot methods What about virtual call inlining? Monomorphic Bimorphic Megamorphic
42. Method Inlining Method Inlining Inlining is the most important optimization Combine caller and callee into one unit Expands the scope for other optimizations Bytecode size < 35 bytes Bytecode size < 325 bytes for inlining hot methods What about virtual call inlining? Black Magic: https://shipilev.net/blog/2015/black-magic-method-dispatch/ Monomorphic Bimorphic Megamorphic
43. Method Inlining Method Inlining Inlining is the most important optimization Combine caller and callee into one unit Expands the scope for other optimizations Bytecode size < 35 bytes Bytecode size < 325 bytes for inlining hot methods What about virtual call inlining? Black Magic: https://shipilev.net/blog/2015/black-magic-method-dispatch/ Monomorphic Bimorphic Megamorphic Class Hierarchy Analysis
44. Method Inlining Method Inlining Inlining is the most important optimization Combine caller and callee into one unit Expands the scope for other optimizations Bytecode size < 35 bytes Bytecode size < 325 bytes for inlining hot methods What about virtual call inlining? Black Magic: https://shipilev.net/blog/2015/black-magic-method-dispatch/ Monomorphic Bimorphic Megamorphic Class Hierarchy Analysis Devirtualization
45. On-Stack Replacement On-Stack Replacement
46. On-Stack Replacement On-Stack Replacement
47. On-Stack Replacement On-Stack Replacement
48. On-Stack Replacement On-Stack Replacement -XX:+PrintCompilation -XX:+PrintInlining
49. On-Stack Replacement On-Stack Replacement -XX:+PrintCompilation -XX:+PrintInlining
50. On-Stack Replacement On-Stack Replacement -XX:+PrintCompilation -XX:+PrintInlining Compilation is OSR
51. Dead Code Elimination Dead Code Elimination
52. Dead Code Elimination Dead Code Elimination Dead-Code: http://hg.openjdk.java.net/code-tools/jmh/file/tip/jmh-samples/src/main/java/org/openjdk/jmh/samples/JMHSample_08_DeadCode.java
53. Dead Code Elimination Dead Code Elimination Dead-Code: http://hg.openjdk.java.net/code-tools/jmh/file/tip/jmh-samples/src/main/java/org/openjdk/jmh/samples/JMHSample_08_DeadCode.java
54. Dead Code Elimination Dead Code Elimination Dead-Code: http://hg.openjdk.java.net/code-tools/jmh/file/tip/jmh-samples/src/main/java/org/openjdk/jmh/samples/JMHSample_08_DeadCode.java
55. Dead Code Elimination Dead Code Elimination Dead-Code: http://hg.openjdk.java.net/code-tools/jmh/file/tip/jmh-samples/src/main/java/org/openjdk/jmh/samples/JMHSample_08_DeadCode.java
56. Dead Code Elimination Dead Code Elimination Dead-Code: http://hg.openjdk.java.net/code-tools/jmh/file/tip/jmh-samples/src/main/java/org/openjdk/jmh/samples/JMHSample_08_DeadCode.java
57. Dead Code Elimination Dead Code Elimination Dead-Code: http://hg.openjdk.java.net/code-tools/jmh/file/tip/jmh-samples/src/main/java/org/openjdk/jmh/samples/JMHSample_08_DeadCode.java Result is not used and the entire method is optimized away!
58. Escape Analysis Escape Analysis
59. Escape Analysis Escape Analysis No synchronization lock when calling getNoEscapeSum method Escape Analysis for Java: http://www.research.ibm.com/people/j/jdchoi/escape.ps
60. Escape Analysis Escape Analysis No synchronization lock when calling getNoEscapeSum method No allocate a Sum object at all, just keep track of the individual fields of the object Escape Analysis for Java: http://www.research.ibm.com/people/j/jdchoi/escape.ps
61. Deoptimization Deoptimization The compiler can back to previous versions of compiled code Effect in performance when code is deoptimized Deoptimize when previous optimizations are no longer valid There are two cases of deoptimization not entrant -> don't let new calls enter zombi -> on this way to deadness
62. Loop Unrolling Loop Unrolling
63. Loop Unrolling Loop Unrolling
64. Loop Unrolling Loop Unrolling Unrolled!
65. Intrinsics Intrinsics Intrinsics: https://github.com/dmlloyd/openjdk/blob/jdk8u/jdk8u/hotspot/src/share/vm/classfile/vmSymbols.hpp#L603 Special code build-into the VM for a particular method Often use special hardware capabilities Not writen in Java
66. JITWatch JITWatch Jitwatch: https://github.com/AdoptOpenJDK/jitwatch
67. Conclusion Conclusion Please use JMH, Don't lie to yourself Code will be Optimized and De-optimized JVM require Warmup JVM is Smart Keep It Simple, Stupid!, is perfect for JIT to make his job
68. Conclusion Conclusion Please use JMH, Don't lie to yourself Code will be Optimized and De-optimized JVM require Warmup JVM is Smart Keep It Simple, Stupid!, is perfect for JIT to make his job Learn Assembly
69. Resources Resources Java Performance: http://shop.oreilly.com/product/0636920028499.do Optimizing Java: http://shop.oreilly.com/product/0636920042983.do Aleksey Shipilëv (blog): https://shipilev.net/ Nitsan Wakart (blog): http://psy-lob-saw.blogspot.com.ar/ JVM Source Code: https://github.com/dmlloyd/openjdk Open JDK Wiki: https://wiki.openjdk.java.net/dashboard.action
70. Questions? Questions?