interpret method gets compiled twice by ION

[mbebenita]

We should figure out why this is, and make sure it is never compiled more than once. The method is very large and probably takes a while to compile.

[marco-c]

(We're marking with the "intex" label the things that block us from landing the intex branch, with "intex-followup" things that we need to do but that don't block us from landing it)

[mbebenita]

This is a fascinating problem. SpiderMonkey, after executing the J2ME interpreter loop decides to compile it with Baseline and then ION. Unfortunately, the type information recorded by SpiderMonkey when it first compiles the interpreter loop is not complete since it is likely that parts of the interpreter loop were never executed (for example the code dealing with OSR or bailouts). When those parts of the code are executed, we end up bailing out of ION and back to baseline, then back into ION for another compilation, and so on. If we are really unlucky, SpiderMonkey might give up on ION all together and we never execute in optimized code.

The only way to stay in ION from the very beginning is to probably warm up the interpreter loop evenly, before it get compiled by ION. Another thing that might help is it factor out infrequently used parts of the interpreter loop into functions.

I pulled a histogram of the startup sequence, showing the top 80% of executed bytecodes.

With JIT Enabled:

 ALOAD_0        13.09% 23033
 GETFIELD       6.58%  11573
 INVOKEVIRTUAL  5.26%  9259 
 ILOAD          3.85%  6780 
 PUTFIELD       3.67%  6466 
 DUP            3.65%  6424 
 ALOAD_1        3.56%  6268 
 INVOKESPECIAL  3.15%  5536 
 BIPUSH         2.77%  4869 
 RETURN         2.58%  4546 
 ICONST_0       2.58%  4537 
 ILOAD_2        2.37%  4174 
 INVOKESTATIC   2.36%  4151 
 ILOAD_1        2.00%  3526 
 ALOAD          1.94%  3411 
 ICONST_1       1.88%  3315 
 GETSTATIC      1.59%  2801 
 ILOAD_3        1.57%  2769 
 GOTO           1.53%  2696 
 ISTORE         1.53%  2692 
 NEW            1.44%  2536 
 IFEQ           1.40%  2467 
 IRETURN        1.32%  2317 
 IADD           1.26%  2223 
 ARETURN        1.18%  2076 
 LDC            1.14%  2011 
 ALOAD_2        1.05%  1856 
 IFNE           0.79%  1397 
 IF_ICMPGE      0.79%  1391 
 ALOAD_3        0.79%  1384 
 PUTSTATIC      0.79%  1383 

With JIT Disabled:

 ALOAD_0        11.88% 2576326
 GETFIELD       9.65%  2093871
 ILOAD          9.06%  1965620
 ILOAD_2        5.78%  1254358
 GOTO           3.93%  852107 
 ILOAD_1        3.40%  736872 
 IF_ICMPGE      3.32%  720257 
 BIPUSH         2.84%  615380 
 IINC           2.80%  607237 
 ISTORE         2.42%  524609 
 ICONST_1       2.31%  501056 
 ICONST_0       2.18%  473358 
 IADD           2.17%  471250 
 INVOKEVIRTUAL  2.01%  435600 
 IRETURN        2.00%  433069 
 IASTORE        1.94%  421679 
 IF_ICMPEQ      1.94%  420708 
 DUP            1.56%  338008 
 ICONST_4       1.53%  331056 
 CALOAD         1.49%  322349 
 ISTORE_2       1.31%  284207 
 IALOAD         1.28%  278660 
 ILOAD_3        1.28%  278233 
 ALOAD_1        1.21%  262003 
 PUTFIELD       1.16%  251748 

[mbebenita]

1843 is an effort to reduce the size of the interpreter loop.