Commit 64489255 authored by Michael Kruse's avatar Michael Kruse
Browse files

[Polly] Add greedy fusion algorithm. 

When the option -polly-loopfusion-greedy is set, the ScheduleOptimizer
tries to aggressively fuse any band it can and does not violate any
dependences.

As part if the implementation, the functionalty for copying a band
into an new schedule was extracted out of the ScheduleTreeRewriter.
parent 20a0c482

polly/docs/ReleaseNotes.rst

+12 −0
Original line number Diff line number Diff line
@@ -11,3 +11,15 @@ In Polly 14 the following important changes have been incorporated. 
  branch.



- Change ...



 * The command line option -polly-opt-fusion has been removed. What the

   flag does was frequently misunderstood and is rarely useful. However,

   the functionality is still accessible using

```

    -polly-isl-arg=--no-schedule-serialize-sccs

```



 * The command line option -polly-loopfusion-greedy has been added.

   This will agressively try to fuse any loop regardless of

   profitability. The is what users might have expected what

   -polly-opt-fusion=max would do.

polly/include/polly/ScheduleTreeTransform.h

+8 −0
Original line number Diff line number Diff line
@@ -240,6 +240,14 @@ isl::schedule_node applyRegisterTiling(isl::schedule_node Node, 
                                       llvm::ArrayRef<int> TileSizes,

                                       int DefaultTileSize);



/// Apply greedy fusion. That is, fuse any loop that is possible to be fused

/// top-down.

///

/// @param Sched  Sched tree to fuse all the loops in.

/// @param Deps   Validity constraints that must be preserved.

isl::schedule applyGreedyFusion(isl::schedule Sched,

                                const isl::union_map &Deps);



} // namespace polly



#endif // POLLY_SCHEDULETREETRANSFORM_H

polly/include/polly/Support/GICHelper.h

+6 −0
Original line number Diff line number Diff line
@@ -231,6 +231,12 @@ ISL_DUMP_OBJECT(val) 
ISL_DUMP_OBJECT(val_list)

//@}



/// Emit the equivaltent of the isl_*_dump output into a raw_ostream.

/// @{

void dumpIslObj(const isl::schedule_node &Node, llvm::raw_ostream &OS);

void dumpIslObj(__isl_keep isl_schedule_node *node, llvm::raw_ostream &OS);

/// @}



inline llvm::raw_ostream &operator<<(llvm::raw_ostream &OS,

                                     __isl_keep isl_union_map *Map) {

  OS << polly::stringFromIslObj(Map, "null");

polly/lib/Support/GICHelper.cpp

+21 −0
Original line number Diff line number Diff line
@@ -238,4 +238,25 @@ void neverCalled() { 
  polly::dumpIslObj(isl::val());

  polly::dumpIslObj(isl::val_list());

}



void polly::dumpIslObj(__isl_keep isl_schedule_node *node, raw_ostream &OS) {

  if (!node)

    return;



  isl_ctx *ctx = isl_schedule_node_get_ctx(node);

  isl_printer *p = isl_printer_to_str(ctx);

  p = isl_printer_set_yaml_style(p, ISL_YAML_STYLE_BLOCK);

  p = isl_printer_print_schedule_node(p, node);



  char *char_str = isl_printer_get_str(p);

  OS << char_str;



  free(char_str);

  isl_printer_free(p);

}



void polly::dumpIslObj(const isl::schedule_node &Node, raw_ostream &OS) {

  dumpIslObj(Node.get(), OS);

}



#endif

polly/lib/Transform/ScheduleOptimizer.cpp

+12 −0
Original line number Diff line number Diff line
@@ -97,6 +97,11 @@ static cl::opt<std::string> 
                      cl::desc("Maximize the band depth (yes/no)"), cl::Hidden,

                      cl::init("yes"), cl::ZeroOrMore, cl::cat(PollyCategory));



static cl::opt<bool>

    GreedyFusion("polly-loopfusion-greedy",

                 cl::desc("Aggressively try to fuse everything"), cl::Hidden,

                 cl::ZeroOrMore, cl::cat(PollyCategory));



static cl::opt<std::string> OuterCoincidence(

    "polly-opt-outer-coincidence",

    cl::desc("Try to construct schedules where the outer member of each band "
@@ -835,6 +840,13 @@ static bool runIslScheduleOptimizer( 
  if (Schedule.is_null())

    return false;



  if (GreedyFusion) {

    isl::union_map Validity = D.getDependences(

        Dependences::TYPE_RAW | Dependences::TYPE_WAR | Dependences::TYPE_WAW);

    Schedule = applyGreedyFusion(Schedule, Validity);

    assert(!Schedule.is_null());

  }



  // Apply post-rescheduling optimizations (if enabled) and/or prevectorization.

  const OptimizerAdditionalInfoTy OAI = {

      TTI, const_cast<Dependences *>(&D),