Improve compilation speed

[peq]

Here is an overview of the compilation times for EBR for the different phases when compiled with -runcompiletimefunctions -stacktraces -inline -localOptimizations -measure.

parse files: 5592ms
Typecheck program: 6296ms
Translate program to Im: 1347ms
 to intermediate lang: 1347ms
Run compiletime functions: 1180ms
Transform program to Jass: 46558ms
 translate classes: 409ms
 add stack traces: 12706ms
 inlining: 1025ms
  inline functions: 987ms
 eliminate tuples: 1197ms
 remove func refs: 14ms
 remove cyclic functions: 1228ms
 flatten: 190ms
 local optimizations: 29150ms
  Constant and Copy Propagated: 1295ms
  Useless function calls removed: 592ms
  Globals Inlined: 417ms
  Branches merged: 7114ms
  Simple Rewrites: 2668ms
  Temp variables merged: 2981ms
  Local variables merged: 12214ms
 null setting: 281ms
 translate to jass: 256ms
Print Jass: 104ms
Run PJass: 708ms

It looks like "add stack traces" , " Local variables merged", "Temp variables merged", "Branches merged" would be good targets to make faster.

[Frotty]

I think it would be nice if we could setup a test suite of some wurst resources (stdlib, fdef. ebr. itt) and then let jenkins compile it with master and the opt-branch so we can directly see and validate changes made.

[peq]

Current state:

parse files: 5324ms
Typecheck program: 7020ms
Translate program to Im: 1585ms
 to intermediate lang: 1585ms
Run compiletime functions: 1232ms
Transform program to Jass: 25817ms
 Eliminate generics: 152ms
 translate classes: 314ms
 add stack traces: 504ms
 inlining: 1307ms
 eliminate tuples: 506ms
 remove func refs: 18ms
 remove cyclic functions: 318ms
 flatten: 68ms
 local optimizations: 22105ms
  Constant and Copy Propagated: 1221ms
  Useless function calls removed: 738ms
  Globals Inlined: 352ms
  Branches merged: 303ms
  Simple Rewrites: 1441ms
  Temp variables merged: 2162ms
  Local variables merged: 14466ms
 null setting: 226ms
 translate to jass: 286ms
Print Jass: 147ms
Run PJass: 179ms

Probably also helped that some debug checks are disabled by default now.

[Frotty]

should display sum after the value so we can see where the times differ. Typecheck seems to take 1000ms longer almost e.g. I will run this also and post results.

Snowball Fight old:

#############################
Run times:
parse files: 1767ms
Typecheck program: 2654ms
Translate program to Im: 804ms
 to intermediate lang: 804ms
Run compiletime functions: 1076ms
Transform program to Jass: 4971ms
 Eliminate generics: 82ms
 translate classes: 223ms
 add stack traces: 267ms
 inlining: 361ms
 eliminate tuples: 135ms
 remove func refs: 8ms
 remove cyclic functions: 80ms
 flatten: 18ms
 local optimizations: 3665ms
  Constant and Copy Propagated: 232ms
  Useless function calls removed: 134ms
  Globals Inlined: 101ms
  Branches merged: 44ms
  Simple Rewrites: 284ms
  Temp variables merged: 388ms
  Local variables merged: 2142ms
 null setting: 46ms
 translate to jass: 76ms
Print Jass: 59ms
Run PJass: 258ms

new:

#############################
Run times:
parse files: 1776ms
Typecheck program: 2588ms
Translate program to Im: 865ms
 to intermediate lang: 865ms
Run compiletime functions: 971ms
Transform program to Jass: 5076ms
 Eliminate generics: 87ms
 translate classes: 254ms
 add stack traces: 244ms
 inlining: 360ms
 eliminate tuples: 130ms
 remove func refs: 9ms
 remove cyclic functions: 75ms
 flatten: 19ms
 local optimizations: 3760ms
  Constant and Copy Propagated: 243ms
  Useless function calls removed: 135ms
  Globals Inlined: 100ms
  Branches merged: 49ms
  Simple Rewrites: 312ms
  Temp variables merged: 403ms
  Local variables merged: 2129ms
 null setting: 48ms
 translate to jass: 82ms
Print Jass: 76ms
Run PJass: 263ms

[Frotty]

I think the JVM would need to be warmed up and have enough heap to not cause any garbage collection during build to produce useful results. Right now without opimizations I get heavy fluctuations of up to 1000ms in most tasks, e.g. Transform program to Jass and Run compiletime functions (900 vs 1600).

[Frotty]

Several performance improvements have been done, builds are now much faster. If you still have specific performance bottlenecks open a separate ticket with measurements and reproduction map/project.