Loop body unrolling

1 files changed, 39 insertions, 1 deletions

diff --git a/backend/Duplicateaux.ml b/backend/Duplicateaux.ml
index ec7a4d02..84dc92ac 100644
--- a/backend/Duplicateaux.ml
+++ b/backend/Duplicateaux.ml
@@ -783,7 +783,6 @@ let unroll_inner_loop_single code revmap iloop =
       debug "IHEAD: %d\n" (P.to_int iloop.head);
       code' := change_pointers !code' (iloop.head) head' (iloop.preds);
       code' := change_pointers !code' head' (iloop.head) [final'];
-      (* code' := PTree.set last_n' (change_single_next first_n @@ ptree_get_some last_n' !code') !code' *)
       (!code', revmap2)
     end
 
@@ -801,6 +800,41 @@ let unroll_inner_loops_single f code revmap =
     (!code', !revmap')
   end
 
+(* 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 into the first of body' 
+ * 3) Links the last instruction of body' into the first of body
+ *)
+let unroll_inner_loop_body code revmap iloop =
+  let body = HashedSet.PSet.elements (iloop.body) in
+  if count_ignore_nops code body > 1000 then begin (* FIXME *)
+    debug "Too many nodes in the loop body (%d > %d)" (List.length body) 1000;
+    (code, revmap)
+  end else
+    let (code2, revmap2, dupbody, fwmap) = clone code revmap body in
+    let code' = ref code2 in
+    let head' = apply_map fwmap (iloop.head) in
+    let final' = apply_map fwmap (iloop.final) in
+    begin
+      code' := change_pointers !code' iloop.head head' [iloop.final];
+      code' := change_pointers !code' head' iloop.head [final']; 
+      (!code', revmap2)
+    end
+
+let unroll_inner_loops_body 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) = unroll_inner_loop_body !code' !revmap' iloop in
+      code' := new_code; revmap' := new_revmap
+    ) inner_loops;
+    (!code', !revmap')
+  end
+
 let duplicate_aux f =
   (* initializing *)
   let entrypoint = f.fn_entrypoint in
@@ -819,6 +853,10 @@ let duplicate_aux f =
       unroll_inner_loops_single f code revmap
     else (code, revmap) in
 
+  (* unroll body *)
+  let (code, revmap) =
+    unroll_inner_loops_body f code revmap in
+
   (* static prediction bis *)
   let code =
     if !Clflags.option_fpredict then