Javascript:SpiderMonkey:OdinMonkey: Difference between revisions

Revision as of 12:32, 30 November 2013

Cache compiled asm.js modules: caching landed on FF28 (Nightly). Follow-up improvements: cache from workers (bug 941830), cache more than one module per origin (bug 944821).
Math.fround is in ES6 and SpiderMonkey and optimized in Ion. Add to asm.js/Odin (bug 904918))
Standardize then implement SIMD extensions to JS (bug 894105) and then incorporate into asm.js (TODO: bug)
Land experimental SharedArrayBuffer (pref'd off, bug 933001) and use this to add experimental pthreads support to Emscripten and evaluate on threaded codebases in preparation for a standards proposal.

Add stack-walking support for asm.js code:
- Per-function information in the builtin FF profiler
- Make asm.js calls show up in Error.stack.
Remove parsing limitations associated with huge JS files:
- Fix lazy-parsing/decompression memory-spike and performance hit for huge js files (bug 938385).
- Remove num-locals (bug 916612) and num-lines (bug 916564) parser limits.
Mitigate 32-bit OOM problems: (bug 865959)
Standardize then implement an ArrayBuffer.resize (bug 927182, to allow growable heap) and ArrayBuffer.discard (bug 855669, to allow madvise(DONTNEED))
Add && and || to asm.js to avoid sub-optimal control flow (bug 919052).

Experiment with using LLVM as the compiler backend for asm.js code.
Extend asm.js to better support dynamically linking asm.js modules (a mutable, callable array argument to asm.js module, like the PLT)
Compress asm.js cache entries
Add enough try/catch to asm.js to allow compilation of C++ exceptions using the zero-cost implementation strategy
Start general JS parsing before the entire JS script has completed downloading so that asm.js compilation can overlap
Standardize and implement value objects (bug 749786) then add int64/uint64 to asm.js
asm.js parser : bug 854061
Support GC allocated Java-like objects to allow asm.js generation from JVM/CLR bytecode. Use Typed Objects to describe class structure. Use real vtables.
fall back to baseline jit for very large functions (maybe not necessary with Emscripten outlining): bug 875174
FunctionFuture : bug 854627

native vs. asm.js differential profiling : bug 863924
Revisit primary Ion algorithms (phi placement, gvn, regalloc) to avoid the polynomial compile-time blowup we sometimes see on huge functions
improve heap access performance : bug 865516, bug 865523, bug 869606, bug 870743, bug 897425
improve x86/ARM asm.js-to-asm.js ABI : bug 861785
further improve FFI calls : bug 882399, bug 886411
Optimize calls into asm.js code (to avoid the general CallAsmJS C++ trampoline.
Use a signal handler to:
- remove stack overflow checks
- remove idiv branching
- remove double-to-int conversion branching (see also bug 818750)
Use useAny when lowering new AsmJS MIR nodes
Optimize CFG to minimize branching : bug 844779
Don't spill non-volatile registers at calls out to C++
Align loop headers on natural boundaries : bug 867001
Further optimize multiplication by a constant to use lea tricks for small constants (e.g. x*3 == x<<1+x)

Credit to John Howard

@@ Line 21: / Line 21: @@
 * Compress asm.js cache entries
 * Add enough try/catch to asm.js to allow compilation of C++ exceptions using the zero-cost implementation strategy
-* Start general JS parsing before the entire JS script has completed downloadeding so that asm.js compilation can overlap
+* Start general JS parsing before the entire JS script has completed downloading so that asm.js compilation can overlap
 * Standardize and implement value objects ({{bug|749786}}) then add int64/uint64 to asm.js
 * asm.js parser : {{bug|854061}}