SlideShare a Scribd company logo

JavaScriptCore's DFG JIT (JSConf EU 2012)

•
•
Igalia
Igalia

This document summarizes Andy Wingo's talk on JavaScriptCore's optimizing compiler called the DFG JIT. It begins by providing context on Wingo and an overview of JavaScriptCore's compilation pipeline. It then analyzes the performance of DFG JIT on various V8 benchmarks through experimentation with variable warmup times. Key optimizations identified include inlining, integer operations, and floating point arithmetic. The document ends by discussing ways to extract more data from JavaScriptCore like debug options and comparing performance to V8.

1 of 30
Download to read offline
The DFG JIT, Inside & Out JavaScriptCore’s Optimizing Compiler JSConf.EU 2012 Andy Wingo
wingo@igalia.com Compiler hacker at Igalia Contract work on language implementations V8, JavaScriptCore Schemer
Hubris “Now that JavaScriptCore is as fast as V8 on its own benchmark, it’s well past time to take a look inside JSC’s optimizing compiler, the DFG JIT.”
DFG Optimizing compiler for JSC LLInt -> Baseline JIT -> DFG JIT Makes hot code run fast But how good is it?
An empirical approach Getting good code What: V8 benchmarks When: Hacked V8 benchmarks How: Code dive
The V8 benchmarks The best performance possible from an optimizing compiler ❧ full second of warmup ❧ full second of runtime ❧ long run amortizes GC pauses
Baseline JIT vs DFG
Abusing the V8 benchmarks When does the DFG kick in? What does it do? Idea: V8 benchmarks with variable warmup ❧ after 0 ms of warmup ❧ after 5 ms of warmup ❧ after n ms of warmup Small fixed runtime (5 ms)
Caveats Very sensitive ❧ GC ❧ optimization pauses ❧ timer precision ... but then, so is real code Keep close eye on distribution of measurements
Richards Speedup: 3.7X Bit ops, properties, prototypes
TaskControlBlock.prototype.isHeldOrSuspended = function () { return (this.state & STATE_HELD) != 0 || (this.state == STATE_SUSPENDED); }; GetLocal 0x7f4d028abbf4: CheckStructure 0x7f4d028abbf8: 0x7f4d028abc02: 0x7f4d028abc05: GetByOffset 0x7f4d028abc0b: GetGlobalVar 0x7f4d028abc0f: 0x7f4d028abc19: BitAnd 0x7f4d028abc1c: 0x7f4d028abc1f: 0x7f4d028abc25: 0x7f4d028abc28: 0x7f4d028abc2e: mov -0x38(%r13), %rax mov $0x7f4d00109c80, %r11 cmp %r11, (%rax) jnz 0x7f4d028abd15 mov 0x38(%rax), %rax mov $0x7f4d479cdca8, %rdx mov (%rdx), %rdx cmp %r14, %rax jb 0x7f4d028abd2b cmp %r14, %rdx jb 0x7f4d028abd41 and %edx, %eax
CompareEq 0x7f4d028abc30: 0x7f4d028abc32: 0x7f4d028abc34: 0x7f4d028abc37: 0x7f4d028abc3a: xor %ecx, %ecx cmp %ecx, %eax setz %al movzx %al, %eax or $0x6, %eax LogicalNot 0x7f4d028abc3d: xor $0x1, %rax SetLocal 0x7f4d028abc41: mov %rax, 0x0(%r13) Branch 0x7f4d028abc45: test $0x1, %eax 0x7f4d028abc4b: jnz 0x7f4d028abc87 ... 0x7f4d028abc97: ret (End Of Main Path) ...
DeltaBlue Speedup: 4.4X Prototypes, inlining
Inlining At 20ms: Delaying optimization for Constraint.prototype.satisfy (in loop) because of insufficient profiling. Eventually succeeds after 4 more times and 20 more ms; see --maximumOptimizationDelay.
1000 cuts One function optimized about 20ms in: Planner.prototype.addConstraintsConsumingTo = function (v, coll) { var determining = v.determinedBy; var cc = v.constraints; for (var i = 0; i < cc.size(); i++) { var c = cc.at(i); if (c != determining && c.isSatisfied() coll.add(c); } } Many small marginal gains
Crypto Speedup: 4.1X Integers, arrays
function am3(i,x,w,j,c,n) { var this_array = this.array; var w_array = w.array; var xl = x&0x3fff, xh = x>>14; while(--n >= 0) { var l = this_array[i]&0x3fff; var h = this_array[i++]>>14; var m = xh*l+h*xl; l = xl*l+((m&0x3fff)<<14)+w_array[j]+c; c = (l>>28)+(m>>14)+xh*h; w_array[j++] = l&0xfffffff; } return c; }
var l = this_array[i]&0x3fff GetLocal: this_array 0x7f4d02909bf6: mov 0x0(%r13), %r10 GetLocal: i (int32; type check hoisted) 0x7f4d02909bfa: mov -0x40(%r13), %eax GetButterfly: this_array 0x7f4d02909bfe: mov 0x8(%r10), %rdx GetByVal: this_array[i] (array check hoisted) 0x7f4d02909c02: cmp -0x4(%rdx), %eax 0x7f4d02909c05: jae 0x7f4d02909ed2 0x7f4d02909c0b: mov 0x10(%rdx,%rax,8), %rcx 0x7f4d02909c10: test %rcx, %rcx 0x7f4d02909c13: jz 0x7f4d02909ee8 BitAnd: 0x7f4d02909c19: cmp %r14, %rcx 0x7f4d02909c1c: jb 0x7f4d02909efe 0x7f4d02909c22: mov %rcx, %rbx 0x7f4d02909c25: and $0x3fff, %ebx
RayTrace Speedup: 2.5X Floating point, objects with floating-point fields
normalize() normalize : function() { var m = this.magnitude(); return new Flog.RayTracer.Vector(this.x / m, this.y / m, this.z / m); }, DFG inlines as it compiles: inlines this.magnitude() ArithDiv: 0x7f4d0298164b: divsd %xmm1, %xmm0 SetLocal: 0x7f4d0298164f: movd %xmm0, %rdx 0x7f4d02981654: sub %r14, %rdx 0x7f4d02981657: mov %rdx, 0x20(%r13) No typed fields (yet)
EarleyBoyer Speedup: 2.0X Function calls, small short-lived allocations
EarleyBoyer “Performance is a distribution, not a value” Wide distribution indicates nonuniform performance Cause in this case: nonincremental mark GC
RegExp Speedup: 1.2X Regexp compiler test; DFG of no help
Splay Speedup: 1.4X GC test, huge variance
NavierStokes Speedup: 3.0X Floating point arrays, large floating-point functions
No automagic double arrays GetByVal: 0x7f4d02acec1f: 0x7f4d02acec23: 0x7f4d02acec29: 0x7f4d02acec2e: 0x7f4d02acec31: cmp -0x4(%rcx), %r9d jae 0x7f4d02acee0b mov 0x10(%rcx,%r9,8), %rbx test %rbx, %rbx jz 0x7f4d02acee21 GetLocal: 0x7f4d02acec37: Int32ToDouble: 0x7f4d02acec3b: 0x7f4d02acec3e: 0x7f4d02acec44: 0x7f4d02acec47: 0x7f4d02acec4d: 0x7f4d02acec50: 0x7f4d02acec53: 0x7f4d02acec58: 0x7f4d02acec5d: mov -0x50(%r13), %rdi cmp %r14, %rbx jae 0x7f4d02acec5d test %rbx, %r14 jz 0x7f4d02acee37 mov %rbx, %rsi add %r14, %rsi movd %rsi, %xmm0 jmp 0x7f4d02acec61 cvtsi2sd %ebx, %xmm0
Getting data out of JSC jsc --options jsc -d jsc --showDFGDisassembly=true -DJIT_ENABLE_VERBOSE=1, DJIT_ENABLE_VERBOSE_OSR=1 and timestamping hacks on dataLog
Comparative Literature V8 vs JSC: fight! Does JSC beat V8? Does JSC meet V8? Does V8 beat JSC?
Yes
Questions? ❧ igalia.com/compilers ❧ wingolog.org ❧ @andywingo ❧ wingolog.org/pub/jsconf-eu-2012slides.pdf

More Related Content

JavaScriptCore's DFG JIT (JSConf EU 2012)

  • 1. The DFG JIT, Inside & Out JavaScriptCore’s Optimizing Compiler JSConf.EU 2012 Andy Wingo
  • 2. wingo@igalia.com Compiler hacker at Igalia Contract work on language implementations V8, JavaScriptCore Schemer
  • 3. Hubris “Now that JavaScriptCore is as fast as V8 on its own benchmark, it’s well past time to take a look inside JSC’s optimizing compiler, the DFG JIT.”
  • 4. DFG Optimizing compiler for JSC LLInt -> Baseline JIT -> DFG JIT Makes hot code run fast But how good is it?
  • 5. An empirical approach Getting good code What: V8 benchmarks When: Hacked V8 benchmarks How: Code dive
  • 6. The V8 benchmarks The best performance possible from an optimizing compiler ❧ full second of warmup ❧ full second of runtime ❧ long run amortizes GC pauses
  • 8. Abusing the V8 benchmarks When does the DFG kick in? What does it do? Idea: V8 benchmarks with variable warmup ❧ after 0 ms of warmup ❧ after 5 ms of warmup ❧ after n ms of warmup Small fixed runtime (5 ms)
  • 9. Caveats Very sensitive ❧ GC ❧ optimization pauses ❧ timer precision ... but then, so is real code Keep close eye on distribution of measurements
  • 10. Richards Speedup: 3.7X Bit ops, properties, prototypes
  • 11. TaskControlBlock.prototype.isHeldOrSuspended = function () { return (this.state & STATE_HELD) != 0 || (this.state == STATE_SUSPENDED); }; GetLocal 0x7f4d028abbf4: CheckStructure 0x7f4d028abbf8: 0x7f4d028abc02: 0x7f4d028abc05: GetByOffset 0x7f4d028abc0b: GetGlobalVar 0x7f4d028abc0f: 0x7f4d028abc19: BitAnd 0x7f4d028abc1c: 0x7f4d028abc1f: 0x7f4d028abc25: 0x7f4d028abc28: 0x7f4d028abc2e: mov -0x38(%r13), %rax mov $0x7f4d00109c80, %r11 cmp %r11, (%rax) jnz 0x7f4d028abd15 mov 0x38(%rax), %rax mov $0x7f4d479cdca8, %rdx mov (%rdx), %rdx cmp %r14, %rax jb 0x7f4d028abd2b cmp %r14, %rdx jb 0x7f4d028abd41 and %edx, %eax
  • 12. CompareEq 0x7f4d028abc30: 0x7f4d028abc32: 0x7f4d028abc34: 0x7f4d028abc37: 0x7f4d028abc3a: xor %ecx, %ecx cmp %ecx, %eax setz %al movzx %al, %eax or $0x6, %eax LogicalNot 0x7f4d028abc3d: xor $0x1, %rax SetLocal 0x7f4d028abc41: mov %rax, 0x0(%r13) Branch 0x7f4d028abc45: test $0x1, %eax 0x7f4d028abc4b: jnz 0x7f4d028abc87 ... 0x7f4d028abc97: ret (End Of Main Path) ...
  • 14. Inlining At 20ms: Delaying optimization for Constraint.prototype.satisfy (in loop) because of insufficient profiling. Eventually succeeds after 4 more times and 20 more ms; see --maximumOptimizationDelay.
  • 15. 1000 cuts One function optimized about 20ms in: Planner.prototype.addConstraintsConsumingTo = function (v, coll) { var determining = v.determinedBy; var cc = v.constraints; for (var i = 0; i < cc.size(); i++) { var c = cc.at(i); if (c != determining && c.isSatisfied() coll.add(c); } } Many small marginal gains
  • 17. function am3(i,x,w,j,c,n) { var this_array = this.array; var w_array = w.array; var xl = x&0x3fff, xh = x>>14; while(--n >= 0) { var l = this_array[i]&0x3fff; var h = this_array[i++]>>14; var m = xh*l+h*xl; l = xl*l+((m&0x3fff)<<14)+w_array[j]+c; c = (l>>28)+(m>>14)+xh*h; w_array[j++] = l&0xfffffff; } return c; }
  • 18. var l = this_array[i]&0x3fff GetLocal: this_array 0x7f4d02909bf6: mov 0x0(%r13), %r10 GetLocal: i (int32; type check hoisted) 0x7f4d02909bfa: mov -0x40(%r13), %eax GetButterfly: this_array 0x7f4d02909bfe: mov 0x8(%r10), %rdx GetByVal: this_array[i] (array check hoisted) 0x7f4d02909c02: cmp -0x4(%rdx), %eax 0x7f4d02909c05: jae 0x7f4d02909ed2 0x7f4d02909c0b: mov 0x10(%rdx,%rax,8), %rcx 0x7f4d02909c10: test %rcx, %rcx 0x7f4d02909c13: jz 0x7f4d02909ee8 BitAnd: 0x7f4d02909c19: cmp %r14, %rcx 0x7f4d02909c1c: jb 0x7f4d02909efe 0x7f4d02909c22: mov %rcx, %rbx 0x7f4d02909c25: and $0x3fff, %ebx
  • 19. RayTrace Speedup: 2.5X Floating point, objects with floating-point fields
  • 20. normalize() normalize : function() { var m = this.magnitude(); return new Flog.RayTracer.Vector(this.x / m, this.y / m, this.z / m); }, DFG inlines as it compiles: inlines this.magnitude() ArithDiv: 0x7f4d0298164b: divsd %xmm1, %xmm0 SetLocal: 0x7f4d0298164f: movd %xmm0, %rdx 0x7f4d02981654: sub %r14, %rdx 0x7f4d02981657: mov %rdx, 0x20(%r13) No typed fields (yet)
  • 21. EarleyBoyer Speedup: 2.0X Function calls, small short-lived allocations
  • 22. EarleyBoyer “Performance is a distribution, not a value” Wide distribution indicates nonuniform performance Cause in this case: nonincremental mark GC
  • 23. RegExp Speedup: 1.2X Regexp compiler test; DFG of no help
  • 25. NavierStokes Speedup: 3.0X Floating point arrays, large floating-point functions
  • 26. No automagic double arrays GetByVal: 0x7f4d02acec1f: 0x7f4d02acec23: 0x7f4d02acec29: 0x7f4d02acec2e: 0x7f4d02acec31: cmp -0x4(%rcx), %r9d jae 0x7f4d02acee0b mov 0x10(%rcx,%r9,8), %rbx test %rbx, %rbx jz 0x7f4d02acee21 GetLocal: 0x7f4d02acec37: Int32ToDouble: 0x7f4d02acec3b: 0x7f4d02acec3e: 0x7f4d02acec44: 0x7f4d02acec47: 0x7f4d02acec4d: 0x7f4d02acec50: 0x7f4d02acec53: 0x7f4d02acec58: 0x7f4d02acec5d: mov -0x50(%r13), %rdi cmp %r14, %rbx jae 0x7f4d02acec5d test %rbx, %r14 jz 0x7f4d02acee37 mov %rbx, %rsi add %r14, %rsi movd %rsi, %xmm0 jmp 0x7f4d02acec61 cvtsi2sd %ebx, %xmm0
  • 27. Getting data out of JSC jsc --options jsc -d jsc --showDFGDisassembly=true -DJIT_ENABLE_VERBOSE=1, DJIT_ENABLE_VERBOSE_OSR=1 and timestamping hacks on dataLog
  • 28. Comparative Literature V8 vs JSC: fight! Does JSC beat V8? Does JSC meet V8? Does V8 beat JSC?
  • 29. Yes
  • 30. Questions? ❧ igalia.com/compilers ❧ wingolog.org ❧ @andywingo ❧ wingolog.org/pub/jsconf-eu-2012slides.pdf