Irreducible transform

[agle]

Add irreducible loop transform and pull across more tests

[agle]

The result for Fig 6.a differs to Basil; we only identify two loops whereas basil identifiers 3 (2 irreducible).

https://github.com/agle/bincaml/actions/runs/23731415109/job/69125853957?pr=108#step:8:28

Here we pick just b as the irreducible loop; whereas basil picks h1 and h2. seems suspicious. If I swap the iteration order of successors it picks the same headers as basil (but other things fail because I've hardcoded the header pointers list in a bunch of other tests), @katrinafyi does what its done in the test log look like a valid choice? The test passes otherwise, (all irreducible loops removed and become reducible) I'm just not fully convinced by rotating the loops in my mind that just b -> h3 is a valid loop forest.

[katrinafyi]

this one before transform? I'll have a think

  { block = ("%S", 0); dfs_pos = 1; loop = NonLoop }
  , { block = ("%h3", 1); dfs_pos = 2;
    loop =
    PrimaryHeader {primary_header = None; headers = {("%h3", 1)};
      nodes =
      {("%h3", 1), ("%x", 2), ("%y", 3), ("%b0", 4), ("%h2", 5), ("%h1", 6), ("%b", 7), ("%z", 8), ("%a", 9), ("%back", 10)}}
    }
  , { block = ("%x", 2); dfs_pos = 3;
    loop = LoopParticipant {primary_header = ("%h3", 1)} }
  , { block = ("%y", 3); dfs_pos = 4;
    loop = LoopParticipant {primary_header = ("%h3", 1)} }
  , { block = ("%b0", 4); dfs_pos = 5;
    loop = LoopParticipant {primary_header = ("%h3", 1)} }
  , { block = ("%b", 7); dfs_pos = 6;
    loop =
    PrimaryHeader {primary_header = (Some ("%h3", 1));
      headers = {("%h2", 5), ("%b", 7)};
      nodes = {("%h2", 5), ("%h1", 6), ("%b", 7), ("%z", 8), ("%a", 9)}}
    }
  , { block = ("%z", 8); dfs_pos = 7;
    loop = LoopParticipant {primary_header = ("%b", 7)} }
  , { block = ("%h1", 6); dfs_pos = 8;
    loop = LoopParticipant {primary_header = ("%b", 7)} }
  , { block = ("%a", 9); dfs_pos = 8;
    loop = LoopParticipant {primary_header = ("%b", 7)} }
  , { block = ("%h2", 5); dfs_pos = 9;
    loop = LoopParticipant {primary_header = ("%b", 7)} }
  , { block = ("%back", 10); dfs_pos = 9;
    loop = LoopParticipant {primary_header = ("%h3", 1)} }
  , { block = ("%exit", 11); dfs_pos = 10; loop = NonLoop }
  

[katrinafyi]

do you have the full sbasil loop infos?

[agle]

  case class BlockLoopInfo(
    val b: Block,
    val iloop_header: Option[Block],
    val dfsp_pos: Int,
    val headers: Set[Block],
    val nodes: Set[Block]
}

Before transform:

  + BlockLoopInfo(%S,None,1,Set(),Set()) 
  + BlockLoopInfo(%h3,None,2,Set(%h3),HashSet(%back, %h2, %y, %b0, %z, %b, %h1, %x, %h3, %a)) 
  + BlockLoopInfo(%x,Some(%h3),3,Set(),Set()) 
  + BlockLoopInfo(%h2,Some(%h3),4,Set(%h2, %b),HashSet(%h1, %h2, %z, %a, %b)) 
  + BlockLoopInfo(%y,Some(%h3),4,Set(),Set()) 
  + BlockLoopInfo(%h1,Some(%h2),5,Set(%h1, %b),Set(%h1, %z, %b)) 
  + BlockLoopInfo(%b0,Some(%h3),5,Set(),Set()) 
  + BlockLoopInfo(%b,Some(%h1),6,Set(),Set()) 
  + BlockLoopInfo(%z,Some(%h1),7,Set(),Set()) 
  + BlockLoopInfo(%a,Some(%h2),8,Set(),Set()) 
  + BlockLoopInfo(%back,Some(%h3),9,Set(),Set()) 
  + BlockLoopInfo(%exit,None,10,Set(),Set()) 

After transform


  + BlockLoopInfo(%S,None,1,Set(),Set()) 
  + BlockLoopInfo(%h3,None,2,Set(%h3),HashSet(%h1_loop_N, %back, %h2, %y, %h2_loop_N, %b0, %z, %b, %h1, %x, %h3, %a)) 
  + BlockLoopInfo(%x,Some(%h3),3,Set(),Set()) 
  + BlockLoopInfo(%y,Some(%h3),4,Set(),Set()) 
  + BlockLoopInfo(%b0,Some(%h3),5,Set(),Set()) 
  + BlockLoopInfo(%h2_loop_N,Some(%h3),6,Set(%h2_loop_N),HashSet(%h1_loop_N, %h1, %h2, %h2_loop_N, %z, %a, %b)) 
  + BlockLoopInfo(%h2,Some(%h2_loop_N),7,Set(),Set()) 
  + BlockLoopInfo(%h1_loop_N,Some(%h2_loop_N),8,Set(%h1_loop_N),Set(%h1_loop_N, %z, %b, %h1)) 
  + BlockLoopInfo(%h1,Some(%h1_loop_N),9,Set(),Set()) 
  + BlockLoopInfo(%b,Some(%h1_loop_N),10,Set(),Set()) 
  + BlockLoopInfo(%z,Some(%h1_loop_N),11,Set(),Set()) 
  + BlockLoopInfo(%a,Some(%h2_loop_N),12,Set(),Set()) 
  + BlockLoopInfo(%back,Some(%h3),13,Set(),Set()) 
  + BlockLoopInfo(%exit,None,14,Set(),Set()) 

[katrinafyi]

Yeah I'm happy with both the sbasil and bincaml forests. If you visit b first (like bincaml), then there is no sub-loop because all the inner loops go through b. If you visit h2 first, then there's a subloop without h2 via h1.

As long as the transform works, it should be fine.

[katrinafyi]

Suggested change

	doc = "Remove blocks unreachable from entry";
	doc = "Transform irreducible loops to reducible ones";