1 This document describes system properties that are used for internal 2 debugging and instrumentation purposes, along with the system loggers, 3 which are used for the same thing. 4 5 This document is intended as a developer resource, and it is not 6 needed as Nashorn documentation for normal usage. Flags and system 7 properties described herein are subject to change without notice. 8 9 ===================================== 10 1. System properties used internally 11 ===================================== 12 13 This documentation of the system property flags assume that the 14 default value of the flag is false, unless otherwise specified. 15 16 SYSTEM PROPERTY: -Dnashorn.unstable.relink.threshold=x 17 18 This property controls how many call site misses are allowed before a 19 callsite is relinked with "apply" semantics to never change again. 20 In the case of megamorphic callsites, this is necessary, or the 21 program would spend all its time swapping out callsite targets. Dynalink 22 has a default value (currently 8 relinks) for this property if it 23 is not explicitly set. 24 25 26 SYSTEM PROPERTY: -Dnashorn.compiler.splitter.threshold=x 27 28 This will change the node weight that requires a subgraph of the IR to 29 be split into several classes in order not to run out of bytecode space. 30 The default value is 0x8000 (32768). 31 32 33 SYSTEM PROPERTY: -Dnashorn.compiler.intarithmetic 34 35 Arithmetic operations in Nashorn (except bitwise ones) typically 36 coerce the operands to doubles (as per the JavaScript spec). To switch 37 this off and remain in integer mode, for example for "var x = a&b; var 38 y = c&d; var z = x*y;", use this flag. This will force the 39 multiplication of variables that are ints to be done with the IMUL 40 bytecode and the result "z" to become an int. 41 42 WARNING: Note that is is experimental only to ensure that type support 43 exists for all primitive types. The generated code is unsound. This 44 will be the case until we do optimizations based on it. There is a CR 45 in Nashorn to do better range analysis, and ensure that this is only 46 done where the operation can't overflow into a wider type. Currently 47 no overflow checking is done, so at the moment, until range analysis 48 has been completed, this option is turned off. 49 50 We've experimented by using int arithmetic for everything and putting 51 overflow checks afterwards, which would recompute the operation with 52 the correct precision, but have yet to find a configuration where this 53 is faster than just using doubles directly, even if the int operation 54 does not overflow. Getting access to a JVM intrinsic that does branch 55 on overflow would probably alleviate this. 56 57 There is also a problem with this optimistic approach if the symbol 58 happens to reside in a local variable slot in the bytecode, as those 59 are strongly typed. Then we would need to split large sections of 60 control flow, so this is probably not the right way to go, while range 61 analysis is. There is a large difference between integer bytecode 62 without overflow checks and double bytecode. The former is 63 significantly faster. 64 65 66 SYSTEM PROPERTY: -Dnashorn.codegen.debug, -Dnashorn.codegen.debug.trace=<x> 67 68 See the description of the codegen logger below. 69 70 71 SYSTEM_PROPERTY: -Dnashorn.fields.debug 72 73 See the description on the fields logger below. 74 75 76 SYSTEM PROPERTY: -Dnashorn.fields.dual 77 78 When this property is true, Nashorn will attempt to use primitive 79 fields for AccessorProperties (currently just AccessorProperties, not 80 spill properties). Memory footprint for script objects will increase, 81 as we need to maintain both a primitive field (a long) as well as an 82 Object field for the property value. Ints are represented as the 32 83 low bits of the long fields. Doubles are represented as the 84 doubleToLongBits of their value. This way a single field can be used 85 for all primitive types. Packing and unpacking doubles to their bit 86 representation is intrinsified by the JVM and extremely fast. 87 88 While dual fields in theory runs significantly faster than Object 89 fields due to reduction of boxing and memory allocation overhead, 90 there is still work to be done to make this a general purpose 91 solution. Research is ongoing. 92 93 In the future, this might complement or be replaced by experimental 94 feature sun.misc.TaggedArray, which has been discussed on the mlvm 95 mailing list. TaggedArrays are basically a way to share data space 96 between primitives and references, and have the GC understand this. 97 98 As long as only primitive values are written to the fields and enough 99 type information exists to make sure that any reads don't have to be 100 uselessly boxed and unboxed, this is significantly faster than the 101 standard "Objects only" approach that currently is the default. See 102 test/examples/dual-fields-micro.js for an example that runs twice as 103 fast with dual fields as without them. Here, the compiler, can 104 determine that we are dealing with numbers only throughout the entire 105 property life span of the properties involved. 106 107 If a "real" object (not a boxed primitive) is written to a field that 108 has a primitive representation, its callsite is relinked and an Object 109 field is used forevermore for that particular field in that 110 PropertyMap and its children, even if primitives are later assigned to 111 it. 112 113 As the amount of compile time type information is very small in a 114 dynamic language like JavaScript, it is frequently the case that 115 something has to be treated as an object, because we don't know any 116 better. In reality though, it is often a boxed primitive is stored to 117 an AccessorProperty. The fastest way to handle this soundly is to use 118 a callsite typecheck and avoid blowing the field up to an Object. We 119 never revert object fields to primitives. Ping-pong:ing back and forth 120 between primitive representation and Object representation would cause 121 fatal performance overhead, so this is not an option. 122 123 For a general application the dual fields approach is still slower 124 than objects only fields in some places, about the same in most cases, 125 and significantly faster in very few. This is due the program using 126 primitives, but we still can't prove it. For example "local_var a = 127 call(); field = a;" may very well write a double to the field, but the 128 compiler dare not guess a double type if field is a local variable, 129 due to bytecode variables being strongly typed and later non 130 interchangeable. To get around this, the entire method would have to 131 be replaced and a continuation retained to restart from. We believe 132 that the next steps we should go through are instead: 133 134 1) Implement method specialization based on callsite, as it's quite 135 frequently the case that numbers are passed around, but currently our 136 function nodes just have object types visible to the compiler. For 137 example "var b = 17; func(a,b,17)" is an example where two parameters 138 can be specialized, but the main version of func might also be called 139 from another callsite with func(x,y,"string"). 140 141 2) This requires lazy jitting as the functions have to be specialized 142 per callsite. 143 144 Even though "function square(x) { return x*x }" might look like a 145 trivial function that can always only take doubles, this is not 146 true. Someone might have overridden the valueOf for x so that the 147 toNumber coercion has side effects. To fulfil JavaScript semantics, 148 the coercion has to run twice for both terms of the multiplication 149 even if they are the same object. This means that call site 150 specialization is necessary, not parameter specialization on the form 151 "function square(x) { var xd = (double)x; return xd*xd; }", as one 152 might first think. 153 154 Generating a method specialization for any variant of a function that 155 we can determine by types at compile time is a combinatorial explosion 156 of byte code (try it e.g. on all the variants of am3 in the Octane 157 benchmark crypto.js). Thus, this needs to be lazy 158 159 3) Possibly optimistic callsite writes, something on the form 160 161 x = y; //x is a field known to be a primitive. y is only an object as 162 far as we can tell 163 164 turns into 165 166 try { 167 x = (int)y; 168 } catch (X is not an integer field right now | ClassCastException e) { 169 x = y; 170 } 171 172 Mini POC shows that this is the key to a lot of dual field performance 173 in seemingly trivial micros where one unknown object, in reality 174 actually a primitive, foils it for us. Very common pattern. Once we 175 are "all primitives", dual fields runs a lot faster than Object fields 176 only. 177 178 We still have to deal with objects vs primitives for local bytecode 179 slots, possibly through code copying and versioning. 180 181 182 SYSTEM PROPERTY: -Dnashorn.compiler.symbol.trace=[<x>[,*]], 183 -Dnashorn.compiler.symbol.stacktrace=[<x>[,*]] 184 185 When this property is set, creation and manipulation of any symbol 186 named "x" will show information about when the compiler changes its 187 type assumption, bytecode local variable slot assignment and other 188 data. This is useful if, for example, a symbol shows up as an Object, 189 when you believe it should be a primitive. Usually there is an 190 explanation for this, for example that it exists in the global scope 191 and type analysis has to be more conservative. 192 193 Several symbols names to watch can be specified by comma separation. 194 195 If no variable name is specified (and no equals sign), all symbols 196 will be watched 197 198 By using "stacktrace" instead of or together with "trace", stack 199 traces will be displayed upon symbol changes according to the same 200 semantics. 201 202 203 SYSTEM PROPERTY: nashorn.lexer.xmlliterals 204 205 If this property it set, it means that the Lexer should attempt to 206 parse XML literals, which would otherwise generate syntax 207 errors. Warning: there are currently no unit tests for this 208 functionality. 209 210 XML literals, when this is enabled, end up as standard LiteralNodes in 211 the IR. 212 213 214 SYSTEM_PROPERTY: nashorn.debug 215 216 If this property is set to true, Nashorn runs in Debug mode. Debug 217 mode is slightly slower, as for example statistics counters are enabled 218 during the run. Debug mode makes available a NativeDebug instance 219 called "Debug" in the global space that can be used to print property 220 maps and layout for script objects, as well as a "dumpCounters" method 221 that will print the current values of the previously mentioned stats 222 counters. 223 224 These functions currently exists for Debug: 225 226 "map" - print(Debug.map(x)) will dump the PropertyMap for object x to 227 stdout (currently there also exist functions called "embedX", where X 228 is a value from 0 to 3, that will dump the contents of the embed pool 229 for the first spill properties in any script object and "spill", that 230 will dump the contents of the growing spill pool of spill properties 231 in any script object. This is of course subject to change without 232 notice, should we change the script object layout. 233 234 "methodHandle" - this method returns the method handle that is used 235 for invoking a particular script function. 236 237 "identical" - this method compares two script objects for reference 238 equality. It is a == Java comparison 239 240 "dumpCounters" - will dump the debug counters' current values to 241 stdout. 242 243 Currently we count number of ScriptObjects in the system, number of 244 Scope objects in the system, number of ScriptObject listeners added, 245 removed and dead (without references). 246 247 We also count number of ScriptFunctions, ScriptFunction invocations 248 and ScriptFunction allocations. 249 250 Furthermore we count PropertyMap statistics: how many property maps 251 exist, how many times were property maps cloned, how many times did 252 the property map history cache hit, prevent new allocations, how many 253 prototype invalidations were done, how many time the property map 254 proto cache hit. 255 256 Finally we count callsite misses on a per callsite bases, which occur 257 when a callsite has to be relinked, due to a previous assumption of 258 object layout being invalidated. 259 260 261 SYSTEM PROPERTY: nashorn.methodhandles.debug, 262 nashorn.methodhandles.debug=create 263 264 If this property is enabled, each MethodHandle related call that uses 265 the java.lang.invoke package gets its MethodHandle intercepted and an 266 instrumentation printout of arguments and return value appended to 267 it. This shows exactly which method handles are executed and from 268 where. (Also MethodTypes and SwitchPoints). This can be augmented with 269 more information, for example, instance count, by subclassing or 270 further extending the TraceMethodHandleFactory implementation in 271 MethodHandleFactory.java. 272 273 If the property is specialized with "=create" as its option, 274 instrumentation will be shown for method handles upon creation time 275 rather than at runtime usage. 276 277 278 SYSTEM PROPERTY: nashorn.methodhandles.debug.stacktrace 279 280 This does the same as nashorn.methodhandles.debug, but when enabled 281 also dumps the stack trace for every instrumented method handle 282 operation. Warning: This is enormously verbose, but provides a pretty 283 decent "grep:able" picture of where the calls are coming from. 284 285 See the description of the codegen logger below for a more verbose 286 description of this option 287 288 289 SYSTEM PROPERTY: nashorn.scriptfunction.specialization.disable 290 291 There are several "fast path" implementations of constructors and 292 functions in the NativeObject classes that, in their original form, 293 take a variable amount of arguments. Said functions are also declared 294 to take Object parameters in their original form, as this is what the 295 JavaScript specification mandates. 296 297 However, we often know quite a lot more at a callsite of one of these 298 functions. For example, Math.min is called with a fixed number (2) of 299 integer arguments. The overhead of boxing these ints to Objects and 300 folding them into an Object array for the generic varargs Math.min 301 function is an order of magnitude slower than calling a specialized 302 implementation of Math.min that takes two integers. Specialized 303 functions and constructors are identified by the tag 304 @SpecializedFunction and @SpecializedConstructor in the Nashorn 305 code. The linker will link in the most appropriate (narrowest types, 306 right number of types and least number of arguments) specialization if 307 specializations are available. 308 309 Every ScriptFunction may carry specializations that the linker can 310 choose from. This framework will likely be extended for user defined 311 functions. The compiler can often infer enough parameter type info 312 from callsites for in order to generate simpler versions with less 313 generic Object types. This feature depends on future lazy jitting, as 314 there tend to be many calls to user defined functions, some where the 315 callsite can be specialized, some where we mostly see object 316 parameters even at the callsite. 317 318 If this system property is set to true, the linker will not attempt to 319 use any specialized function or constructor for native objects, but 320 just call the generic one. 321 322 323 SYSTEM PROPERTY: nashorn.tcs.miss.samplePercent=<x> 324 325 When running with the trace callsite option (-tcs), Nashorn will count 326 and instrument any callsite misses that require relinking. As the 327 number of relinks is large and usually produces a lot of output, this 328 system property can be used to constrain the percentage of misses that 329 should be logged. Typically this is set to 1 or 5 (percent). 1% is the 330 default value. 331 332 333 SYSTEM_PROPERTY: nashorn.profilefile=<filename> 334 335 When running with the profile callsite options (-pcs), Nashorn will 336 dump profiling data for all callsites to stderr as a shutdown hook. To 337 instead redirect this to a file, specify the path to the file using 338 this system property. 339 340 341 SYSTEM_PROPERTY: nashorn.regexp.impl=[jdk|joni] 342 343 This property defines the regular expression engine to be used by 344 Nashorn. The default implementation is "jdk" which is based on the 345 JDK's java.util.regex package. Set this property to "joni" to install 346 an implementation based on Joni, the regular expression engine used by 347 the JRuby project. 348 349 350 =============== 351 2. The loggers. 352 =============== 353 354 It is very simple to create your own logger. Use the DebugLogger class 355 and give the subsystem name as a constructor argument. 356 357 The Nashorn loggers can be used to print per-module or per-subsystem 358 debug information with different levels of verbosity. The loggers for 359 a given subsystem are available are enabled by using 360 361 --log=<systemname>[:<level>] 362 363 on the command line. 364 365 Here <systemname> identifies the name of the subsystem to be logged 366 and the optional colon and level argument is a standard 367 java.util.logging.Level name (severe, warning, info, config, fine, 368 finer, finest). If the level is left out for a particular subsystem, 369 it defaults to "info". Any log message logged as the level or a level 370 that is more important will be output to stderr by the logger. 371 372 Several loggers can be enabled by a single command line option, by 373 putting a comma after each subsystem/level tuple (or each subsystem if 374 level is unspecified). The --log option can also be given multiple 375 times on the same command line, with the same effect. 376 377 For example: --log=codegen,fields:finest is equivalent to 378 --log=codegen:info --log=fields:finest 379 380 The subsystems that currently support logging are: 381 382 383 * compiler 384 385 The compiler is in charge of turning source code and function nodes 386 into byte code, and installs the classes into a class loader 387 controlled from the Context. Log messages are, for example, about 388 things like new compile units being allocated. The compiler has global 389 settings that all the tiers of codegen (e.g. Lower and CodeGenerator) 390 use.s 391 392 393 * codegen 394 395 The code generator is the emitter stage of the code pipeline, and 396 turns the lowest tier of a FunctionNode into bytecode. Codegen logging 397 shows byte codes as they are being emitted, line number information 398 and jumps. It also shows the contents of the bytecode stack prior to 399 each instruction being emitted. This is a good debugging aid. For 400 example: 401 402 [codegen] #41 line:2 (f)_afc824e 403 [codegen] #42 load symbol x slot=2 404 [codegen] #43 {1:O} load int 0 405 [codegen] #44 {2:I O} dynamic_runtime_call GT:ZOI_I args=2 returnType=boolean 406 [codegen] #45 signature (Ljava/lang/Object;I)Z 407 [codegen] #46 {1:Z} ifeq ternary_false_5402fe28 408 [codegen] #47 load symbol x slot=2 409 [codegen] #48 {1:O} goto ternary_exit_107c1f2f 410 [codegen] #49 ternary_false_5402fe28 411 [codegen] #50 load symbol x slot=2 412 [codegen] #51 {1:O} convert object -> double 413 [codegen] #52 {1:D} neg 414 [codegen] #53 {1:D} convert double -> object 415 [codegen] #54 {1:O} ternary_exit_107c1f2f 416 [codegen] #55 {1:O} return object 417 418 shows a ternary node being generated for the sequence "return x > 0 ? 419 x : -x" 420 421 The first number on the log line is a unique monotonically increasing 422 emission id per bytecode. There is no guarantee this is the same id 423 between runs. depending on non deterministic code 424 execution/compilation, but for small applications it usually is. If 425 the system variable -Dnashorn.codegen.debug.trace=<x> is set, where x 426 is a bytecode emission id, a stack trace will be shown as the 427 particular bytecode is about to be emitted. This can be a quick way to 428 determine where it comes from without attaching the debugger. "Who 429 generated that neg?" 430 431 The --log=codegen option is equivalent to setting the system variable 432 "nashorn.codegen.debug" to true. 433 434 435 * lower 436 437 This is the first lowering pass. 438 439 Lower is a code generation pass that turns high level IR nodes into 440 lower level one, for example substituting comparisons to RuntimeNodes 441 and inlining finally blocks. 442 443 Lower is also responsible for determining control flow information 444 like end points. 445 446 447 * attr 448 449 The lowering annotates a FunctionNode with symbols for each identifier 450 and transforms high level constructs into lower level ones, that the 451 CodeGenerator consumes. 452 453 Lower logging typically outputs things like post pass actions, 454 insertions of casts because symbol types have been changed and type 455 specialization information. Currently very little info is generated by 456 this logger. This will probably change. 457 458 459 * finalize 460 461 This --log=finalize log option outputs information for type finalization, 462 the third tier of the compiler. This means things like placement of 463 specialized scope nodes or explicit conversions. 464 465 466 * fields 467 468 The --log=fields option (at info level) is equivalent to setting the 469 system variable "nashorn.fields.debug" to true. At the info level it 470 will only show info about type assumptions that were invalidated. If 471 the level is set to finest, it will also trace every AccessorProperty 472 getter and setter in the program, show arguments, return values 473 etc. It will also show the internal representation of respective field 474 (Object in the normal case, unless running with the dual field 475 representation) --- EOF ---