aboutsummaryrefslogblamecommitdiffstats
path: root/lib/hipe/TODO
blob: f166472df6d8b76d23389eea64936bbee49276c6 (plain) (tree)
































































































































                                                                                 
Bugfix
======
 P->current (Fix observable behaviour?)
 New calling convention for guard bifs (Recognize at load time).
 Long branches:
   timer:tc(hipe,c,[megaco_text_parser,[{timeout,infinity}]]).
   {4801210531,
     {error,[{problem,too_long_branch},{address,3381732},{length,-828622}]}}

Performance
===========

 Better handling of multimove in regalloc.
 Faster closure creation. (Can static fields be preallocated?)
 Expand pseudo-ops before scheduler (SPARC)
 Stack maps for SPARC + Make frames in Sparc not in RTL.
 Coalesce spill locations.

Feature
=======

 Stack traces from stack maps.

Cleanup
=======

 Speedup renaming and other bottlenecks in the compiler.
 Only calls with fail label should end basic blocks.
 Remove fail-entry-points from RTL (sparc/x86). 
 Cleanup hipe_*_registers.erl and interface/rules with regalloc.
 HiPE in bootstrap.
 Cleanup and merge loaders. (Better handling of data.) 	
 Re-examine switching code.

Extensions
==========

 Design strategy for finding all processes holding a certain closure. 
 Design strategy for native code unloading.
 mbufs: In guards -> throw away, in bifs -> trigger special GC. (fix for native.)
 Unified heap + process optimization (+ PE).
 Incremental GC.


Old list compiled by Thomas Lindgren (needs cleaning up)
========================================================

<h1>Experimental implementations</h1>
<h2>RTL</h2>
<UL>
  <LI> Algebraiska f�renklingar av uttryck (ex. reducera integer multiply,
      ta bort addition med 0, etc)
  <LI> Partial redundancy elimination 
</UL>

<h1>Unimplemented optimizations</h1>

<H2>Erlang/Core source-level-optimizations</H2>
<UL>
  <LI> "Context compilation"
  <LI> CDR-kodning
  <LI> List reuse
  <LI> Compilation by transformation
</UL>

<H2>Icode-optimizations</H2>
<UL>
  <LI> Convertion to loops from recursive programs
  <LI> Dominatorer
      (<a href="./thomasl/icode/dominators.erl">l�ngsamma</a>, snabba)
</UL>

<H2>RTL-optimizations</H2>
<UL>
  <LI> Common subexpression elimination
  <LI> Ta bort redundanta tester globalt (ex. upprepade typtester)
  <LI> Ordna om hopp (ex. byt ordning p� nil/cons-tester)
  <LI> Goto eliminering (= expandera uncond. jumps m. k�nd m�ltavla)
  <LI> Save/restore-placering: datafl�desanalys, interaktion m. catch-frames
  <LI> Loop optimeringar
      <UL>
	<LI> Dominatorer (se dominatorer f�r icode)
	<LI> Unrolling
	<LI> Invariant expression removal
      </UL>
  <LI> Partial redundancy elimination by lazy code motion
  <LI> Partially dead code
</UL>

<H2>Sparc-optimizations</H2>
<UL>
  <LI> Global register allocation
      <UL>
	<LI> <a href="./thomasl/regalloc/regalloc.erl">
	    Pessimistisk f�rgning</a>
	<LI> Optimistisk f�rgning (kan sl�s p� i samma fil som pessimistisk
	    f�rgning ovan).
	<LI> B�ttre ber�kning av spillkostnader
	<LI> Renaming
	<LI> Callee-saves register
	<LI> Live-range splitting
      </UL>
  <LI> Instruktionsschedulering
      <UL>
	<LI> Branch delay slot scheduling
	<LI> Load delay slot scheduling
	<LI> Spekulativa loads med lduwa
	<LI> Kollapsa serier av tester med bpr
	<LI> Begr�nsad predicated execution med movcc
      </UL>
  <LI> Kodlayout: statiska f�ruts�gelser om riktning av hopp,
      layout, s�tta branch prediction bits i hopp, etc.
  <LI> Prefetching av kod med SparcV9:s bpn.
</UL>

<H2>Other optimizations</H2>

Profile driven optimizations.
<UL>
  <LI> Utplacering av r�knare i CFG:er (per block, per b�ge)
  <LI> Statiska metoder att uppskatta exekveringstid (inom och mellan proc.)
  <LI> Feedback till program, annotering av CFG:er med profileringsinfo.
  <LI> Kodlayout med profileringsinfo.
  <LI> Skapa superblock
  <LI> Skapa hyperblock
  <LI> Plocka fram heta block, skapa en 'het' sub-CFG som hoppar
      till den kalla huvud-CFG:n vid behov.
  <LI> Optimering av het CFG, kodf�rflyttning fr�n het till kall CFG.
  <LI> Spawn-time specialization
</UL>