Javascript:SpiderMonkey:OdinMonkey: Difference between revisions
Jump to navigation
Jump to search
No edit summary |
No edit summary |
||
| Line 3: | Line 3: | ||
* [https://docs.google.com/spreadsheets/d/1PFa3aDxY6mffT8uoflCaFitX9lKj_Y4_aZwtMApIRiI/edit#gid=0 SharedArrayBuffer, Atomics, Futex] | * [https://docs.google.com/spreadsheets/d/1PFa3aDxY6mffT8uoflCaFitX9lKj_Y4_aZwtMApIRiI/edit#gid=0 SharedArrayBuffer, Atomics, Futex] | ||
== | == Possible further load-time optimizations == | ||
* Move code generation off of parsing thread: {{bug|959263}} | * Move code generation off of parsing thread: {{bug|959263}} | ||
* PGAOT (Profile-Guided Ahead-of-Time Compilation): | * PGAOT (Profile-Guided Ahead-of-Time Compilation): | ||
** Create a baseline asm.js compiler: super-fast codegen but with full static type info | ** Create a baseline asm.js compiler: super-fast codegen but with full static type info | ||
** Allow individual functions in an asm.js module to choose full or baseline (via "use" pragma on individual functions) | ** Allow individual functions in an asm.js module to choose full or baseline (via "use" pragma on individual functions) | ||
** Allow Emscripten users to change default (full or baseline), hand-annotate, or do profiling to place pragmas. | ** Allow Emscripten users to change default (full or baseline), hand-annotate, or do profiling | ||
to place pragmas. | |||
* Parse and AOT compile while downloading: {{bug|1061886}} | * Parse and AOT compile while downloading: {{bug|1061886}} | ||
* Remove async-script requirement for asm.js caching: | * Remove async-script requirement for asm.js caching: | ||
| Line 30: | Line 27: | ||
== Possible asm.js extensions that don't require new JS features == | == Possible asm.js extensions that don't require new JS features == | ||
* Zero-cost exception handling: {{bug|1065089}} | |||
* Extend asm.js to allow proper tail calls for all return types and do the tail-call optimization | * Extend asm.js to allow proper tail calls for all return types and do the tail-call optimization | ||
* Effectively support 'goto' (even computed 'goto') by formalizing a pattern of while+switch that gets compiled as if it was goto | * Effectively support 'goto' (even computed 'goto') by formalizing a pattern of while+switch that gets compiled as if it was goto | ||
| Line 46: | Line 44: | ||
* Print warning on out-of-bounds heap access: {{bug|879891}} | * Print warning on out-of-bounds heap access: {{bug|879891}} | ||
* Add about:config option to check for and emit warning on unaligned heap access | * Add about:config option to check for and emit warning on unaligned heap access | ||
* Fix asm.js caching browser interop ({{bug|1047098}}, {{bug|1047105}} | |||
== Current asm.js spec bugs == | == Current asm.js spec bugs == | ||
Revision as of 16:52, 17 October 2014
Active
- Continue iteration on SIMD proposal and implementation/optimization in Ion/Odin.
- SharedArrayBuffer, Atomics, Futex
Possible further load-time optimizations
- Move code generation off of parsing thread: bug 959263
- PGAOT (Profile-Guided Ahead-of-Time Compilation):
- Create a baseline asm.js compiler: super-fast codegen but with full static type info
- Allow individual functions in an asm.js module to choose full or baseline (via "use" pragma on individual functions)
- Allow Emscripten users to change default (full or baseline), hand-annotate, or do profiling
to place pragmas.
- Parse and AOT compile while downloading: bug 1061886
- Remove async-script requirement for asm.js caching:
- Maybe remove it from QuotaManager: bug 961057
- More likely, just take (large) script parsing off the main thread: bug 1084009
- asm.js parser: Avoid building an intermediate parse tree; type check during recursive descent bug 854061
Possible further throughput optimizations
- Optimize sin/cos: bug 967709
- Use callee-saved registers: bug 985065
- Heap access: bug 986981, bug 870743, bug 897425, bug 915157, bug 983131, bug 865523, bug 984478, bug 1056027
- Optimize add-with-overflow: bug 1043365
- Optimize min/max: bug 1060635
- For asm.js->asm.js calls on x86, pass the return value via xmm register, not fp stack
- Remove branching from double-to-int conversion using signal handler (see also bug 818750)
- Optimize calls into (bug 962641), out of (bug 1000632), and between (bug 982036) asm.js modules.
Possible asm.js extensions that don't require new JS features
- Zero-cost exception handling: bug 1065089
- Extend asm.js to allow proper tail calls for all return types and do the tail-call optimization
- Effectively support 'goto' (even computed 'goto') by formalizing a pattern of while+switch that gets compiled as if it was goto
- Extend asm.js to better support dynamically linking asm.js modules (a mutable, callable array argument to asm.js module, like the PLT)
- Add DataView to avoid the alignment mask on asm.js heap accesses (this is mostly a problem of DataView being slow in all browsers)
Proposed future JS features that would be useful to asm.js
- ArrayBuffer.prototype.discard (bug 855669): to allow madvise(DONTNEED)
- Typed Objects: (Typed Objects)
- This feature comes with float32/uint64/int64 value types, which would be useful as scalar types in asm.js.
- Formalize Typed Object "class declarations" in asm.js which could make asm.js a target for JVM/CLR/TypeScript (and allow these languages to reuse the builtin JS GC)
- FunctionPromise: bug 854627, or perhaps just make some functions (asm.js, no-free-variables) structured-cloneable
- Add way to programmatically observe whether asm.js validated: bug 952847
Usability improvements
- Print warning on out-of-bounds heap access: bug 879891
- Add about:config option to check for and emit warning on unaligned heap access
- Fix asm.js caching browser interop (bug 1047098, bug 1047105
Current asm.js spec bugs
https://github.com/dherman/asm.js/issues?state=open
Logo
Credit to John Howard