Loop Rotate with -flooprotate

1 files changed, 50 insertions, 0 deletions

diff --git a/backend/Duplicateaux.ml b/backend/Duplicateaux.ml
index 2b13ab5d..fac0ba76 100644
--- a/backend/Duplicateaux.ml
+++ b/backend/Duplicateaux.ml
@@ -836,6 +836,56 @@ let unroll_inner_loops_body f code revmap =
     (!code', !revmap')
   end
 
+let extract_upto_icond f code head =
+  let rec extract h =
+    let inst = get_some @@ PTree.get h code in
+    match inst with
+    | Icond _ -> [h]
+    | _ -> ( match rtl_successors inst with
+        | [n] -> h :: (extract n)
+        | _ -> failwith "Found a node with more than one successor??"
+      )
+  in List.rev @@ extract head
+
+let rotate_inner_loop f code revmap iloop =
+  let header = extract_upto_icond f code iloop.head in
+  let limit = !Clflags.option_flooprotate in
+  if count_ignore_nops code header > limit then begin
+    debug "Loop Rotate: too many nodes to duplicate (%d > %d)" (List.length header) limit;
+    (code, revmap)
+  end else
+    let (code2, revmap2, dupheader, fwmap) = clone code revmap header in
+    let code' = ref code2 in
+    let head' = apply_map fwmap iloop.head in
+    begin
+      code' := change_pointers !code' iloop.head head' iloop.preds;
+      (!code', revmap2)
+    end
+
+let rotate_inner_loops f code revmap =
+  let is_loop_header = get_loop_headers code (f.fn_entrypoint) in
+  let inner_loops = get_inner_loops f code is_loop_header in
+  let code' = ref code in
+  let revmap' = ref revmap in
+  begin
+    print_inner_loops inner_loops;
+    List.iter (fun iloop ->
+      let (new_code, new_revmap) = rotate_inner_loop f !code' !revmap' iloop in
+      code' := new_code; revmap' := new_revmap
+    ) inner_loops;
+    (!code', !revmap')
+  end
+
+let loop_rotate f =
+  let entrypoint = f.fn_entrypoint in
+  let code = f.fn_code in
+  let revmap = make_identity_ptree code in
+  let (code, revmap) =
+    if !Clflags.option_flooprotate > 0 then
+      rotate_inner_loops f code revmap
+    else (code, revmap) in
+  ((code, entrypoint), revmap)
+
 let static_predict f =
   let entrypoint = f.fn_entrypoint in
   let code = f.fn_code in