Do not duplicate nodes that don't need to when unrolling the body

1 files changed, 18 insertions, 4 deletions

diff --git a/backend/Duplicateaux.ml b/backend/Duplicateaux.ml
index dc95d5b9..9c4304e3 100644
--- a/backend/Duplicateaux.ml
+++ b/backend/Duplicateaux.ml
@@ -895,6 +895,18 @@ let unroll_inner_loops_single f code revmap =
 
 let is_some o = match o with Some _ -> true | None -> false
 
+let rec go_through_predicted code start final =
+  if start == final then
+    Some [start]
+  else
+    match rtl_successors_pref @@ get_some @@ PTree.get start code with
+    | [n] -> (
+        match go_through_predicted code n final with
+        | Some ln -> Some (start :: ln)
+        | None -> None
+      )
+    | _ -> None
+
 (* Unrolls the body of the inner loop once - duplicating the exit condition as well 
  * 1) Clones body into body'
  * 2) Links the last instruction of body (sb_final) into the first of body' 
@@ -911,13 +923,15 @@ let unroll_inner_loop_body code revmap iloop =
     debug "The loop body does not form a superblock OR we have predicted that we do not loop";
     (code, revmap)
   end else
-    let (code2, revmap2, dupbody, fwmap) = clone code revmap body in
+    let sb_final = get_some @@ iloop.sb_final in
+    let sb_body = get_some @@ go_through_predicted code iloop.head sb_final in
+    let (code2, revmap2, dupbody, fwmap) = clone code revmap sb_body in
     let code' = ref code2 in
     let head' = apply_map fwmap (iloop.head) in
-    let finals' = apply_map_list fwmap (iloop.finals) in
+    let sb_final' = apply_map fwmap sb_final in
     begin
-      code' := change_pointers !code' iloop.head head' [get_some @@ iloop.sb_final];
-      code' := change_pointers !code' head' iloop.head finals'; 
+      code' := change_pointers !code' iloop.head head' [sb_final];
+      code' := change_pointers !code' head' iloop.head [sb_final']; 
       (!code', revmap2)
     end