Don't attempt to schedule instrcution beyond an Icond if it is the last

instruction of the superblock.

1 files changed, 29 insertions, 26 deletions

diff --git a/scheduling/MyRTLpathScheduleraux.ml b/scheduling/MyRTLpathScheduleraux.ml
index 40b1e5eb..1beab9f2 100644
--- a/scheduling/MyRTLpathScheduleraux.ml
+++ b/scheduling/MyRTLpathScheduleraux.ml
@@ -434,11 +434,11 @@ let update_liveins liveins live_renames =
 
 let replace_iconds_by_ocmps sb code ~next_free_reg =
   let module P = Camlcoq.P in
-  (* Don't modify the arguments *)
-  (* XXX: not necessary since instructions is never modified, only code *)
-  let instructions = Array.copy sb.instructions in
+  (* We won't replace an Icond if it is the last instruction of the superblock, then it
+   * cannot be moved. *)
+  let last_pc = sb.instructions.(Array.length sb.instructions - 1) in
   let (code, _previous_icond, next_free_reg) =
-    ArrayLabels.fold_left instructions
+    ArrayLabels.fold_left sb.instructions
     ~init:(code, None, next_free_reg)
     ~f:(fun (code, previous_icond_proxy_reg, next_free_reg) pc ->
       let inst = get_some @@ PTree.get pc code in
@@ -451,25 +451,29 @@ let replace_iconds_by_ocmps sb code ~next_free_reg =
           let code' = PTree.set pc istore' code in
           (code', previous_icond_proxy_reg, next_free_reg) )
       | Icond(cond, args, ifso, ifnot, prediction) ->
-        (match prediction with
-        | None ->
-          (* Case only happens at the very end of the path; no transformation necessary *)
-          assert(sb.instructions.(Array.length sb.instructions - 1) = pc);
+        (* Don't transform the last Icond of a superblock *)
+        if pc = last_pc then
           (code, None, next_free_reg)
-        | Some true -> (
-          let ocmp = match previous_icond_proxy_reg with
-          | None -> Iop((Op.Ocmp cond), args, next_free_reg, ifso)
-          | Some r -> Iop((Op.Ocmp cond), r::args, next_free_reg, ifso)
-          in
-          let code' = PTree.set pc ocmp code in
-          (code', Some next_free_reg, P.succ next_free_reg) )
-        | Some false ->  (
-          let ocmp = match previous_icond_proxy_reg with
-          | None -> Iop((Op.Ocmp cond), args, next_free_reg, ifnot)
-          | Some r -> Iop((Op.Ocmp cond), r::args, next_free_reg, ifnot)
-          in
-          let code' = PTree.set pc ocmp code in
-          (code', Some next_free_reg, P.succ next_free_reg) ))
+        else
+          (match prediction with
+          | None ->
+            (* Case only happens at the very end of the path; no transformation necessary *)
+            assert(sb.instructions.(Array.length sb.instructions - 1) = pc);
+            (code, None, next_free_reg)
+          | Some true -> (
+            let ocmp = match previous_icond_proxy_reg with
+            | None -> Iop((Op.Ocmp cond), args, next_free_reg, ifso)
+            | Some r -> Iop((Op.Ocmp cond), r::args, next_free_reg, ifso)
+            in
+            let code' = PTree.set pc ocmp code in
+            (code', Some next_free_reg, P.succ next_free_reg) )
+          | Some false ->  (
+            let ocmp = match previous_icond_proxy_reg with
+            | None -> Iop((Op.Ocmp cond), args, next_free_reg, ifnot)
+            | Some r -> Iop((Op.Ocmp cond), r::args, next_free_reg, ifnot)
+            in
+            let code' = PTree.set pc ocmp code in
+            (code', Some next_free_reg, P.succ next_free_reg) ))
       | Iload(trap, chunk, addr, args, dest, succ) -> (
         if !Machine.config.Machine.has_non_trapping_loads then
           (code, previous_icond_proxy_reg, next_free_reg)
@@ -483,8 +487,7 @@ let replace_iconds_by_ocmps sb code ~next_free_reg =
       | _ -> (code, previous_icond_proxy_reg, next_free_reg)
     )
   in
-  let sb' = {sb with instructions = instructions} in
-  (sb', code, next_free_reg)
+  (code, next_free_reg)
 
 let is_store = function
   | Istore _ -> true
@@ -494,9 +497,9 @@ let is_store = function
  * Iconds into Ocmps for the purpose of dependency calculations. *)
 let ideal_schedule' sb code ~next_free_reg =
   let old_debug_flag = !debug_flag in
-  let (fake_sb, fake_code, _next_free_reg) = replace_iconds_by_ocmps sb code ~next_free_reg in
+  let (fake_code, _next_free_reg) = replace_iconds_by_ocmps sb code ~next_free_reg in
   (* Does PTree.empty work or do I need to map the entries to Regset.empty *)
-  let fake_sb = { fake_sb with liveins = PTree.empty } in
+  let fake_sb = { sb with liveins = PTree.empty } in
   (* Copied from RTLpathScheduleraux.schedule_superblock *)
   let nr_instr_sb = Array.length sb.instructions in
   assert (nr_instr_sb = Array.length fake_sb.instructions);