Changing weights system at asmblock level instead of asm

2 files changed, 154 insertions, 537 deletions

diff --git a/aarch64/OpWeightsAsm.ml b/aarch64/OpWeightsAsm.ml
index c1d0583a..ca120228 100644
--- a/aarch64/OpWeightsAsm.ml
+++ b/aarch64/OpWeightsAsm.ml
@@ -11,109 +11,17 @@
 (*                                                             *)
 (* *************************************************************)
 
+open Asmblock
+
 (*type called_function = (Registers.reg, AST.ident) Datatypes.sum*)
 
-type real_instruction =
-  | Add
-  | Adr
-  | Adrp
-  | And
-  | Asr
-  | B
-  | Bic
-  | Bl
-  | Blr
-  | Br
-  | Cbnz
-  | Cbz
-  | Cls
-  | Clz
-  | Cmn
-  | Cmp
-  | Csel
-  | Cset
-  | Eon
-  | Eor
-  | Fabs
-  | Fadd
-  | Fcmp
-  | Fcsel
-  | Fcvt
-  | Fcvtzs
-  | Fcvtzu
-  | Fdiv
-  | Fmadd
-  | Fmov
-  | Fmsub
-  | Fmul
-  | Fnmadd
-  | Fnmul
-  | Fnmsub
-  | Fneg
-  | Fsqrt
-  | Fsub
-  | Ldaxr
-  | Ldp
-  | Ldr
-  | Ldrb
-  | Ldrh
-  | Ldrsb
-  | Ldrsh
-  | Ldrsw
-  | Lr
-  | Lsl
-  | Lsr
-  | Madd
-  | Mov
-  | Movk
-  | Movn
-  | Movz
-  | Msub
-  | Nop
-  | Orn
-  | Orr
-  | Ret
-  | Rev
-  | Rev16
-  | Ror
-  | Sbfiz
-  | Sbfx
-  | Scvtf
-  | Sdiv
-  | Smulh
-  | Stlxr
-  | Stp
-  | Str
-  | Strb
-  | Strh
-  | Sub
-  | Sxtb
-  | Sxth
-  | Sxtw
-  | Tbnz
-  | Tbz
-  | Tst
-  | Ubfiz
-  | Ubfx
-  | Ucvtf
-  | Udiv
-  | Umulh
-  | Uxtb
-  | Uxth
-  | Uxtw
-  | Uxtx
-  (* Pseudo instr *)
-  | Btbl
-  | Allocframe
-  | Freeframe
-  | Builtin
-  | Cvtx2w
+type instruction = PBasic of Asmblock.basic | PControl of Asmblock.control
 
 type opweights = {
   pipelined_resource_bounds : int array;
   nr_non_pipelined_units : int;
-  latency_of_op : real_instruction -> int -> int;
-  resources_of_op : real_instruction -> int -> int array;
+  latency_of_op : instruction -> int -> int;
+  resources_of_op : instruction -> int -> int array;
       (*non_pipelined_resources_of_op : Op.operation -> int -> int array;*)
       (*latency_of_load : AST.trapping_mode -> AST.memory_chunk -> Op.addressing -> int -> int;*)
       (*resources_of_load : AST.trapping_mode -> AST.memory_chunk -> Op.addressing -> int -> int array;*)
@@ -130,197 +38,86 @@ module Cortex_A53 = struct
   (* instr ; ALU ; MAC; LSU *)
   let nr_non_pipelined_units = 1
 
-  let latency_of_op (i : real_instruction) (nargs : int) =
+  let latency_of_op (i : instruction) (nargs : int) =
     match i with
-    | Add -> 1
-    | Adr -> 1
-    | Adrp -> 1 (* XXX *)
-    | And -> 1
-    | Asr -> 2
-    | B -> 1
-    | Bic -> 1
-    | Bl -> 1
-    | Blr -> 1
-    | Br -> 1
-    | Cbnz -> 1
-    | Cbz -> 1
-    | Cls -> 1
-    | Clz -> 1
-    | Cmn -> 6
-    | Cmp -> 6
-    | Csel -> 2
-    | Cset -> 2
-    | Eon -> 1
-    | Eor -> 1
-    | Fabs -> 6
-    | Fadd -> 1
-    | Fcmp -> 6
-    | Fcsel -> 1
-    | Fcvt -> 6
-    | Fcvtzs -> 6
-    | Fcvtzu -> 6
-    | Fdiv -> 50
-    | Fmadd -> 1
-    | Fmov -> 1
-    | Fmsub -> 1
-    | Fmul -> 6
-    | Fnmadd -> 1
-    | Fnmul -> 1
-    | Fnmsub -> 1
-    | Fneg -> 6
-    | Fsqrt -> 1
-    | Fsub -> 6
-    | Ldaxr -> 3
-    | Ldp -> 3
-    | Ldr -> 3
-    | Ldrb -> 3
-    | Ldrh -> 3
-    | Ldrsb -> 3
-    | Ldrsh -> 3
-    | Ldrsw -> 3
-    | Lr -> 3
-    | Lsl -> 2
-    | Lsr -> 2
-    | Madd -> 4
-    | Mov -> 1 (* XXX *)
-    | Movk -> 1
-    | Movn -> 1
-    | Movz -> 1
-    | Msub -> 4
-    | Nop -> 1
-    | Orn -> 1
-    | Orr -> 1
-    | Ret -> 1
-    | Rev -> 1
-    | Rev16 -> 1
-    | Ror -> 2
-    | Sbfiz -> 2 (* XXX *)
-    | Sbfx -> 2
-    | Scvtf -> 6
-    | Sdiv -> 50
-    | Smulh -> 4
-    | Stlxr -> 2
-    | Stp -> 2
-    | Str -> 2
-    | Strb -> 2
-    | Strh -> 2
-    | Sub -> 1
-    | Sxtb -> 1
-    | Sxth -> 1
-    | Sxtw -> 1
-    | Tbnz -> 1
-    | Tbz -> 1
-    | Tst -> 6
-    | Ubfiz -> 2 (* XXX *)
-    | Ubfx -> 2
-    | Ucvtf -> 6
-    | Udiv -> 50
-    | Umulh -> 4
-    | Uxtb -> 1
-    | Uxth -> 1
-    | Uxtw -> 1
-    | Uxtx -> 1
-    | Btbl -> 1
-    | Allocframe -> 51
-    | Freeframe -> 51
-    | Builtin -> 51
-    | Cvtx2w -> 51
-
-  let resources_of_op (i : real_instruction) (nargs : int) =
+    | PBasic (PArith (PArithP (_, _))) -> 1
+    | PBasic (PArith (PArithPP (i', _, _))) -> (
+        match i' with
+        | Psbfiz (_, _, _) | Psbfx (_, _, _) | Pubfiz (_, _, _) | Pubfx (_, _, _)
+          ->
+            2
+        | Pfcvtds | Pfcvtsd
+        | Pfcvtzs (_, _)
+        | Pfcvtzu (_, _)
+        | Pfabs _ | Pfneg _
+        | Pscvtf (_, _)
+        | Pucvtf (_, _) ->
+            6
+        | _ -> 1)
+    | PBasic (PArith (PArithPPP (i', _, _, _))) -> (
+        match i' with
+        | Pasrv _ | Plslv _ | Plsrv _ | Prorv _ | Paddext _ | Psubext _ -> 2
+        | Psmulh | Pumulh -> 4
+        | Pfdiv _ | Psdiv _ | Pudiv _ -> 50
+        | _ -> 6)
+    | PBasic (PArith (PArithRR0R (_, _, _, _))) -> 2
+    | PBasic (PArith (PArithRR0 (_, _, _))) -> 1
+    | PBasic (PArith (PArithARRRR0 (_, _, _, _, _))) -> 4
+    | PBasic (PArith (PArithComparisonPP (_, _, _))) -> 6
+    | PBasic (PArith (PArithComparisonR0R (_, _, _))) -> 1
+    | PBasic (PArith (PArithComparisonP (i', _))) -> (
+        match i' with Pfcmp0 _ -> 6 | _ -> 1)
+    | PBasic (PArith (Pcset (_, _))) | PBasic (PArith (Pcsel (_, _, _, _))) -> 6
+    | PBasic (PArith _) -> 1
+    | PBasic (PLoad (_, _, _)) -> 3
+    | PBasic (PStore (_, _, _)) -> 3
+    | PBasic (Pallocframe (_, _)) -> 3
+    | PBasic (Pfreeframe (_, _)) -> 1
+    | PBasic (Ploadsymbol (_, _)) -> 1
+    | PBasic (Pcvtsw2x (_, _)) -> 2
+    | PBasic (Pcvtuw2x (_, _)) -> 2
+    | PBasic (Pcvtx2w _) -> 1
+    | PControl _ -> 6
+
+  let resources_of_op (i : instruction) (nargs : int) =
     match i with
-    | Add -> [| 1; 1; 0; 0 |]
-    | Adr -> [| 1; 1; 0; 0 |]
-    | Adrp -> [| 1; 1; 0; 0 |]
-    | And -> [| 1; 1; 0; 0 |]
-    | Asr -> [| 1; 1; 0; 0 |]
-    | B -> [| 1; 1; 0; 0 |]
-    | Bic -> [| 1; 1; 0; 0 |]
-    | Bl -> [| 1; 1; 0; 0 |]
-    | Blr -> [| 1; 1; 0; 0 |]
-    | Br -> [| 1; 1; 0; 0 |]
-    | Cbnz -> [| 1; 1; 0; 0 |]
-    | Cbz -> [| 1; 1; 0; 0 |]
-    | Cls -> [| 1; 1; 0; 0 |]
-    | Clz -> [| 1; 1; 0; 0 |]
-    | Cmn -> [| 1; 1; 1; 0 |]
-    | Cmp -> [| 1; 1; 1; 0 |]
-    | Csel -> [| 1; 1; 1; 0 |]
-    | Cset -> [| 1; 1; 1; 0 |]
-    | Eon -> [| 1; 1; 0; 0 |]
-    | Eor -> [| 1; 1; 0; 0 |]
-    | Fabs -> [| 1; 1; 1; 0 |]
-    | Fadd -> [| 1; 1; 1; 0 |]
-    | Fcmp -> [| 1; 1; 1; 0 |]
-    | Fcsel -> [| 1; 1; 0; 0 |]
-    | Fcvt -> [| 1; 1; 1; 0 |]
-    | Fcvtzs -> [| 1; 1; 1; 0 |]
-    | Fcvtzu -> [| 1; 1; 1; 0 |]
-    | Fdiv -> [| 1; 1; 1; 0 |]
-    | Fmadd -> [| 1; 1; 0; 0 |]
-    | Fmov -> [| 1; 1; 0; 0 |]
-    | Fmsub -> [| 1; 1; 0; 0 |]
-    | Fmul -> [| 1; 1; 1; 0 |]
-    | Fnmadd -> [| 1; 1; 0; 0 |]
-    | Fnmul -> [| 1; 1; 0; 0 |]
-    | Fnmsub -> [| 1; 1; 0; 0 |]
-    | Fneg -> [| 1; 1; 1; 0 |]
-    | Fsqrt -> [| 1; 1; 0; 0 |]
-    | Fsub -> [| 1; 1; 1; 0 |]
-    | Ldaxr -> [| 1; 0; 0; 1 |]
-    | Ldp -> [| 1; 0; 0; 1 |]
-    | Ldr -> [| 1; 0; 0; 1 |]
-    | Ldrb -> [| 1; 0; 0; 1 |]
-    | Ldrh -> [| 1; 0; 0; 1 |]
-    | Ldrsb -> [| 1; 0; 0; 1 |]
-    | Ldrsh -> [| 1; 0; 0; 1 |]
-    | Ldrsw -> [| 1; 0; 0; 1 |]
-    | Lr -> [| 1; 0; 0; 1 |]
-    | Lsl -> [| 1; 1; 0; 0 |]
-    | Lsr -> [| 1; 1; 0; 0 |]
-    | Madd -> [| 1; 1; 1; 0 |]
-    | Mov -> [| 1; 1; 0; 0 |]
-    | Movk -> [| 1; 1; 0; 0 |]
-    | Movn -> [| 1; 1; 0; 0 |]
-    | Movz -> [| 1; 1; 0; 0 |]
-    | Msub -> [| 1; 1; 1; 0 |]
-    | Nop -> [| 1; 1; 0; 0 |]
-    | Orn -> [| 1; 1; 0; 0 |]
-    | Orr -> [| 1; 1; 0; 0 |]
-    | Ret -> [| 1; 1; 0; 0 |]
-    | Rev -> [| 1; 1; 0; 0 |]
-    | Rev16 -> [| 1; 1; 0; 0 |]
-    | Ror -> [| 1; 1; 0; 0 |]
-    | Sbfiz -> [| 1; 1; 0; 0 |]
-    | Sbfx -> [| 1; 1; 0; 0 |]
-    | Scvtf -> [| 1; 1; 1; 0 |]
-    | Sdiv -> [| 1; 0; 0; 0 |]
-    | Smulh -> [| 1; 1; 0; 0 |]
-    | Stlxr -> [| 1; 0; 0; 1 |]
-    | Stp -> [| 1; 0; 0; 1 |]
-    | Str -> [| 1; 0; 0; 1 |]
-    | Strb -> [| 1; 0; 0; 1 |]
-    | Strh -> [| 1; 0; 0; 1 |]
-    | Sub -> [| 1; 1; 0; 0 |]
-    | Sxtb -> [| 1; 1; 0; 0 |]
-    | Sxth -> [| 1; 1; 0; 0 |]
-    | Sxtw -> [| 1; 1; 0; 0 |]
-    | Tbnz -> [| 1; 1; 0; 0 |]
-    | Tbz -> [| 1; 1; 0; 0 |]
-    | Tst -> [| 1; 1; 1; 0 |]
-    | Ubfiz -> [| 1; 1; 0; 0 |]
-    | Ubfx -> [| 1; 1; 0; 0 |]
-    | Ucvtf -> [| 1; 1; 1; 0 |]
-    | Udiv -> [| 1; 0; 0; 0 |]
-    | Umulh -> [| 1; 1; 0; 0 |]
-    | Uxtb -> [| 1; 1; 0; 0 |]
-    | Uxth -> [| 1; 1; 0; 0 |]
-    | Uxtw -> [| 1; 1; 0; 0 |]
-    | Uxtx -> [| 1; 1; 0; 0 |]
-    | Btbl -> [| 1; 1; 0; 0 |]
-    | Allocframe -> [| 1; 1; 1; 1 |]
-    | Freeframe -> [| 1; 1; 1; 1 |]
-    | Builtin -> [| 1; 1; 1; 1 |]
-    | Cvtx2w -> [| 1; 1; 1; 1 |]
+    | PBasic (PArith (PArithP (i', _))) -> (
+        match i' with
+        | Pfmovimmd _ | Pfmovimms _ -> [| 1; 1; 0; 1 |]
+        | _ -> [| 1; 1; 0; 0 |])
+    | PBasic (PArith (PArithPP (i', _, _))) -> (
+        match i' with
+        | Pfcvtds | Pfcvtsd
+        | Pfcvtzs (_, _)
+        | Pfcvtzu (_, _)
+        | Pfabs _ | Pfneg _
+        | Pscvtf (_, _)
+        | Pucvtf (_, _) ->
+            [| 1 ; 1; 1; 0 |]
+        | _ -> [| 1; 1; 0; 0 |])
+    | PBasic (PArith (PArithPPP (i', _, _, _))) -> (
+        match i' with
+        | Pasrv _ | Plslv _ | Plsrv _ | Prorv _ | Paddext _ | Psubext _ -> [| 1; 1; 0; 0 |]
+        | Pfdiv _ | Psdiv _ | Pudiv _ -> [| 1; 0; 0; 0 |]
+        | _ -> [| 1; 1; 1; 0 |])
+    | PBasic (PArith (PArithRR0R (_, _, _, _))) -> [| 1; 1; 0; 0 |]
+    | PBasic (PArith (PArithRR0 (_, _, _))) -> [| 1; 1; 0; 0 |]
+    | PBasic (PArith (PArithARRRR0 (_, _, _, _, _))) -> [| 1; 1; 1; 0 |]
+    | PBasic (PArith (PArithComparisonPP (_, _, _))) -> [| 1; 1; 1; 0 |]
+    | PBasic (PArith (PArithComparisonR0R (_, _, _))) -> [| 1; 1; 0; 0 |]
+    | PBasic (PArith (PArithComparisonP (i', _))) -> (
+      match i' with Pfcmp0 _ -> [| 1; 1; 1; 0 |] | _ -> [| 1; 1; 0; 0 |])
+    | PBasic (PArith (Pcset (_, _))) | PBasic (PArith (Pcsel (_, _, _, _))) -> [| 1; 1; 1; 0 |]
+    | PBasic (PArith _) -> [| 1; 1; 0; 0 |]
+    | PBasic (PLoad (_, _, _)) -> [| 1; 0; 0; 1 |]
+    | PBasic (PStore (_, _, _)) -> [| 1; 0; 0; 1 |]
+    | PBasic (Pallocframe (_, _)) -> [| 1; 0; 0; 1 |]
+    | PBasic (Pfreeframe (_, _)) -> [| 1; 1; 0; 0 |]
+    | PBasic (Ploadsymbol (_, _)) -> [| 1; 1; 0; 0 |]
+    | PBasic (Pcvtsw2x (_, _)) -> [| 1; 1; 0; 0 |]
+    | PBasic (Pcvtuw2x (_, _)) -> [| 1; 1; 0; 0 |]
+    | PBasic (Pcvtx2w _) -> [| 1; 1; 0; 0 |]
+    | PControl _ -> resource_bounds
 
   (*let non_pipelined_resources_of_op (op : operation) (nargs : int) =
       match op with
@@ -342,17 +139,6 @@ module Cortex_A53 = struct
           [| 1; 1; 1; 0 |]
        | _ -> [| 1; 1; 0; 0 |] )*)
 
-  let latency_of_load trap chunk _ _ = 3
-
-  let latency_of_call _ _ = 6
-
-  let resources_of_load trap chunk addressing nargs = [| 1; 0; 0; 1 |]
-
-  let resources_of_store chunk addressing nargs = [| 1; 0; 0; 1 |]
-
-  let resources_of_call _ _ = resource_bounds
-
-  let resources_of_builtin _ = resource_bounds
 end
 
 let get_opweights () : opweights =
diff --git a/aarch64/PostpassSchedulingOracle.ml b/aarch64/PostpassSchedulingOracle.ml
index 9263babc..fbcdb55d 100644
--- a/aarch64/PostpassSchedulingOracle.ml
+++ b/aarch64/PostpassSchedulingOracle.ml
@@ -23,12 +23,10 @@ let debug = false
  * Extracting infos from Asm instructions
  *)
 
-type instruction = PBasic of basic | PControl of control
-
 type location = Reg of Asm.preg | Mem | IREG0_XZR
 
 type ab_inst_rec = {
-  inst : real_instruction;
+  inst : instruction;
   write_locs : location list;
   read_locs : location list;
   is_control : bool;
@@ -38,173 +36,6 @@ type ab_inst_rec = {
 
 exception OpaqueInstruction
 
-let arith_p_real = function
-  | Padrp (_, _) -> Adrp
-  | Pmovz (_, _, _) -> Movz
-  | Pmovn (_, _, _) -> Movn
-  | Pfmovimms _ ->
-      Fmov
-      (* XXX We could also use load, but Fmov may be more convenient for tuning *)
-  | Pfmovimmd _ -> Fmov
-
-let arith_pp_real = function
-  | Pmov -> Mov
-  | Pmovk (_, _, _) -> Movk
-  | Paddadr (_, _) -> Add
-  | Psbfiz (_, _, _) -> Sbfiz
-  | Psbfx (_, _, _) -> Sbfx
-  | Pubfiz (_, _, _) -> Ubfiz
-  | Pubfx (_, _, _) -> Ubfx
-  | Pfmov -> Fmov
-  | Pfcvtds -> Fcvt
-  | Pfcvtsd -> Fcvt
-  | Pfabs _ -> Fabs
-  | Pfneg _ -> Fneg
-  | Pscvtf (_, _) -> Scvtf
-  | Pucvtf (_, _) -> Ucvtf
-  | Pfcvtzs (_, _) -> Fcvtzs
-  | Pfcvtzu (_, _) -> Fcvtzu
-  | Paddimm (_, _) -> Add
-  | Psubimm (_, _) -> Sub
-
-let arith_ppp_real = function
-  | Pasrv _ -> Asr
-  | Plslv _ -> Lsl
-  | Plsrv _ -> Lsr
-  | Prorv _ -> Ror
-  | Psmulh -> Smulh
-  | Pumulh -> Umulh
-  | Psdiv _ -> Sdiv
-  | Pudiv _ -> Udiv
-  | Paddext _ -> Add
-  | Psubext _ -> Sub
-  | Pfadd _ -> Fadd
-  | Pfdiv _ -> Fdiv
-  | Pfmul _ -> Fmul
-  | Pfsub _ -> Fsub
-
-let arith_rr0r_real = function
-  | Padd (_, _) -> Add
-  | Psub (_, _) -> Sub
-  | Pand (_, _) -> And
-  | Pbic (_, _) -> Bic
-  | Peon (_, _) -> Eon
-  | Peor (_, _) -> Eor
-  | Porr (_, _) -> Orr
-  | Porn (_, _) -> Orn
-
-let arith_rr0_real = function
-  | Pandimm (_, _) -> And
-  | Peorimm (_, _) -> Eor
-  | Porrimm (_, _) -> Orr
-
-let arith_arrrr0_real = function Pmadd _ -> Madd | Pmsub _ -> Msub
-
-let arith_comparison_p_real = function
-  | Pfcmp0 _ -> Fcmp
-  | Pcmpimm (_, _) -> Cmp
-  | Pcmnimm (_, _) -> Cmn
-  | Ptstimm (_, _) -> Tst
-
-let arith_comparison_pp_real = function
-  | Pcmpext _ -> Cmp
-  | Pcmnext _ -> Cmn
-  | Pfcmp _ -> Fcmp
-
-let arith_comparison_r0r_real = function
-  | Pcmp (_, _) -> Cmp
-  | Pcmn (_, _) -> Cmn
-  | Ptst (_, _) -> Tst
-
-let arith_comparison_rr0r_real = function
-  | Padd (_, _) -> Add
-  | Psub (_, _) -> Sub
-  | Pand (_, _) -> And
-  | Pbic (_, _) -> Bic
-  | Peon (_, _) -> Eon
-  | Peor (_, _) -> Eor
-  | Porr (_, _) -> Orr
-  | Porn (_, _) -> Orn
-
-let arith_comparison_rr0_real = function
-  | Pandimm (_, _) -> And
-  | Peorimm (_, _) -> Eor
-  | Porrimm (_, _) -> Orr
-
-let arith_comparison_arrrr0_real = function Pmadd _ -> Madd | Pmsub _ -> Msub
-
-let store_rs_a_real = function
-  | Pstrw -> Str
-  | Pstrw_a -> Str
-  | Pstrx -> Str
-  | Pstrx_a -> Str
-  | Pstrb -> Strb
-  | Pstrh -> Strh
-  | Pstrs -> Str
-  | Pstrd -> Str
-  | Pstrd_a -> Str
-
-let load_rd_a_real = function
-  | Pldrw -> Ldr
-  | Pldrw_a -> Ldr
-  | Pldrx -> Ldr
-  | Pldrx_a -> Ldr
-  | Pldrb _ -> Ldrb
-  | Pldrsb _ -> Ldrsb
-  | Pldrh _ -> Ldrh
-  | Pldrsh _ -> Ldrsh
-  | Pldrzw -> Ldr
-  | Pldrsw -> Ldrsw
-  | Pldrs -> Ldr
-  | Pldrd -> Ldr
-  | Pldrd_a -> Ldr
-
-let loadsymbol_real = Ldr
-
-let cvtsw2x_real = Sxtw
-
-let cvtuw2x_real = Uxtw
-
-let cset_real = Cset
-
-let csel_real = Csel
-
-let fmovi_real = Fmov
-
-let fnmul_real = Fnmul
-
-let b_real = B
-
-let bc_real = B
-
-let bl_real = Bl
-
-let bs_real = B
-
-let blr_real = Blr
-
-let br_real = Br
-
-let ret_real = Ret
-
-let cbnz_real = Cbnz
-
-let cbz_real = Cbz
-
-let tbnz_real = Tbnz
-
-let tbz_real = Tbz
-
-let btbl_real = Btbl
-
-let allocframe_real = Allocframe
-
-let freeframe_real = Freeframe
-
-let builtin_real = Builtin
-
-let cvtx2w_real = Cvtx2w
-
 let is_XZR = function IREG0_XZR -> true | _ -> false
 
 let reg_of_pc = Reg Asm.PC
@@ -232,17 +63,17 @@ let get_arith_p_wlocs = function
   | Pfmovimmd _ -> [ reg_of_ireg Asm.X16 ]
   | _ -> []
 
-let arith_p_rec i rd =
+let arith_p_rec i i' rd =
   {
-    inst = arith_p_real i;
-    write_locs = [ rd ] @ get_arith_p_wlocs i;
+    inst = i;
+    write_locs = [ rd ] @ get_arith_p_wlocs i';
     read_locs = [];
     is_control = false;
   }
 
 let arith_pp_rec i rd rs =
   {
-    inst = arith_pp_real i;
+    inst = i;
     write_locs = [ rd ];
     read_locs = [ rs ];
     is_control = false;
@@ -250,7 +81,7 @@ let arith_pp_rec i rd rs =
 
 let arith_ppp_rec i rd r1 r2 =
   {
-    inst = arith_ppp_real i;
+    inst = i;
     write_locs = [ rd ];
     read_locs = [ r1; r2 ];
     is_control = false;
@@ -259,7 +90,7 @@ let arith_ppp_rec i rd r1 r2 =
 let arith_rr0r_rec i rd r1 r2 =
   let rlocs = if is_XZR r1 then [ r2 ] else [ r1; r2 ] in
   {
-    inst = arith_rr0r_real i;
+    inst = i;
     write_locs = [ rd ];
     read_locs = rlocs;
     is_control = false;
@@ -268,7 +99,7 @@ let arith_rr0r_rec i rd r1 r2 =
 let arith_rr0_rec i rd r1 =
   let rlocs = if is_XZR r1 then [] else [ r1 ] in
   {
-    inst = arith_rr0_real i;
+    inst = i;
     write_locs = [ rd ];
     read_locs = rlocs;
     is_control = false;
@@ -277,7 +108,7 @@ let arith_rr0_rec i rd r1 =
 let arith_arrrr0_rec i rd r1 r2 r3 =
   let rlocs = if is_XZR r3 then [ r1; r2 ] else [ r1; r2; r3 ] in
   {
-    inst = arith_arrrr0_real i;
+    inst = i;
     write_locs = [ rd ];
     read_locs = rlocs;
     is_control = false;
@@ -285,7 +116,7 @@ let arith_arrrr0_rec i rd r1 r2 r3 =
 
 let arith_comparison_pp_rec i r1 r2 =
   {
-    inst = arith_comparison_pp_real i;
+    inst = i;
     write_locs = flags_wlocs;
     read_locs = [ r1; r2 ];
     is_control = false;
@@ -294,7 +125,7 @@ let arith_comparison_pp_rec i r1 r2 =
 let arith_comparison_r0r_rec i r1 r2 =
   let rlocs = if is_XZR r1 then [ r2 ] else [ r1; r2 ] in
   {
-    inst = arith_comparison_r0r_real i;
+    inst = i;
     write_locs = flags_wlocs;
     read_locs = rlocs;
     is_control = false;
@@ -302,7 +133,7 @@ let arith_comparison_r0r_rec i r1 r2 =
 
 let arith_comparison_p_rec i r1 =
   {
-    inst = arith_comparison_p_real i;
+    inst = i;
     write_locs = flags_wlocs;
     read_locs = [ r1 ];
     is_control = false;
@@ -320,7 +151,7 @@ let get_eval_addressing_rlocs a =
 
 let load_rec ld rd a =
   {
-    inst = load_rd_a_real ld;
+    inst = ld;
     write_locs = [ rd ];
     read_locs = [ Mem ] @ get_eval_addressing_rlocs a;
     is_control = false;
@@ -328,39 +159,39 @@ let load_rec ld rd a =
 
 let store_rec st r a =
   {
-    inst = store_rs_a_real st;
+    inst = st;
     write_locs = [ Mem ];
     read_locs = [ r; Mem ] @ get_eval_addressing_rlocs a;
     is_control = false;
   }
 
-let loadsymbol_rec rd id =
+let loadsymbol_rec i rd id =
   {
-    inst = loadsymbol_real;
+    inst = i;
     write_locs = [ rd ];
     read_locs = [ Mem ];
     is_control = false;
   }
 
-let cvtsw2x_rec rd r1 =
+let cvtsw2x_rec i rd r1 =
   {
-    inst = cvtsw2x_real;
+    inst = i;
     write_locs = [ rd ];
     read_locs = [ r1 ];
     is_control = false;
   }
 
-let cvtuw2x_rec rd r1 =
+let cvtuw2x_rec i rd r1 =
   {
-    inst = cvtuw2x_real;
+    inst = i;
     write_locs = [ rd ];
     read_locs = [ r1 ];
     is_control = false;
   }
 
-let cvtx2w_rec rd =
+let cvtx2w_rec i rd =
   {
-    inst = cvtx2w_real;
+    inst = i;
     write_locs = [ rd ];
     read_locs = [ rd ];
     is_control = false;
@@ -381,50 +212,50 @@ let get_testcond_rlocs c =
   | Asm.TCgt -> [ reg_of_cr Asm.CN; reg_of_cr Asm.CZ; reg_of_cr Asm.CV ]
   | Asm.TCle -> [ reg_of_cr Asm.CN; reg_of_cr Asm.CZ; reg_of_cr Asm.CV ]
 
-let cset_rec rd c =
+let cset_rec i rd c =
   {
-    inst = cset_real;
+    inst = i;
     write_locs = [ rd ];
     read_locs = get_testcond_rlocs c;
     is_control = false;
   }
 
-let csel_rec rd r1 r2 c =
+let csel_rec i rd r1 r2 c =
   {
-    inst = csel_real;
+    inst = i;
     write_locs = [ rd ];
     read_locs = [ r1; r2 ] @ get_testcond_rlocs c;
     is_control = false;
   }
 
-let fmovi_rec fsz rd r1 =
+let fmovi_rec i fsz rd r1 =
   let rlocs = if is_XZR r1 then [] else [ r1 ] in
   {
-    inst = fmovi_real;
+    inst = i;
     write_locs = [ rd ];
     read_locs = rlocs;
     is_control = false;
   }
 
-let fnmul_rec fsz rd r1 r2 =
+let fnmul_rec i fsz rd r1 r2 =
   {
-    inst = fnmul_real;
+    inst = i;
     write_locs = [ rd ];
     read_locs = [ r1; r2 ];
     is_control = false;
   }
 
-let allocframe_rec sz linkofs =
+let allocframe_rec i sz linkofs =
   {
-    inst = allocframe_real;
+    inst = i;
     write_locs = [ Mem; regXSP; reg_of_ireg Asm.X16; reg_of_ireg Asm.X29 ];
     read_locs = [ regXSP; Mem ];
     is_control = false;
   }
 
-let freeframe_rec sz linkofs =
+let freeframe_rec i sz linkofs =
   {
-    inst = freeframe_real;
+    inst = i;
     write_locs = [ Mem; regXSP; reg_of_ireg Asm.X16 ];
     read_locs = [ regXSP; Mem ];
     is_control = false;
@@ -432,44 +263,44 @@ let freeframe_rec sz linkofs =
 
 let arith_rec i =
   match i with
-  | PArithP (i', rd) -> arith_p_rec i' (reg_of_dreg rd)
-  | PArithPP (i', rd, rs) -> arith_pp_rec i' (reg_of_dreg rd) (reg_of_dreg rs)
+  | PArithP (i', rd) -> arith_p_rec (PBasic (PArith i)) i' (reg_of_dreg rd)
+  | PArithPP (i', rd, rs) -> arith_pp_rec (PBasic (PArith i)) (reg_of_dreg rd) (reg_of_dreg rs)
   | PArithPPP (i', rd, r1, r2) ->
-      arith_ppp_rec i' (reg_of_dreg rd) (reg_of_dreg r1) (reg_of_dreg r2)
+      arith_ppp_rec (PBasic (PArith i)) (reg_of_dreg rd) (reg_of_dreg r1) (reg_of_dreg r2)
   | PArithRR0R (i', rd, r1, r2) ->
-      arith_rr0r_rec i' (reg_of_ireg rd) (reg_of_ireg0 r1) (reg_of_ireg r2)
+      arith_rr0r_rec (PBasic (PArith i)) (reg_of_ireg rd) (reg_of_ireg0 r1) (reg_of_ireg r2)
   | PArithRR0 (i', rd, r1) ->
-      arith_rr0_rec i' (reg_of_ireg rd) (reg_of_ireg0 r1)
+      arith_rr0_rec (PBasic (PArith i)) (reg_of_ireg rd) (reg_of_ireg0 r1)
   | PArithARRRR0 (i', rd, r1, r2, r3) ->
-      arith_arrrr0_rec i' (reg_of_ireg rd) (reg_of_ireg r1) (reg_of_ireg r2)
+      arith_arrrr0_rec (PBasic (PArith i)) (reg_of_ireg rd) (reg_of_ireg r1) (reg_of_ireg r2)
         (reg_of_ireg0 r3)
   | PArithComparisonPP (i', r1, r2) ->
-      arith_comparison_pp_rec i' (reg_of_dreg r1) (reg_of_dreg r2)
+      arith_comparison_pp_rec (PBasic (PArith i)) (reg_of_dreg r1) (reg_of_dreg r2)
   | PArithComparisonR0R (i', r1, r2) ->
-      arith_comparison_r0r_rec i' (reg_of_ireg0 r1) (reg_of_ireg r2)
-  | PArithComparisonP (i', r1) -> arith_comparison_p_rec i' (reg_of_dreg r1)
-  | Pcset (rd, c) -> cset_rec (reg_of_ireg rd) c
-  | Pfmovi (fsz, rd, r1) -> fmovi_rec fsz (reg_of_freg rd) (reg_of_ireg0 r1)
+      arith_comparison_r0r_rec (PBasic (PArith i)) (reg_of_ireg0 r1) (reg_of_ireg r2)
+  | PArithComparisonP (i', r1) -> arith_comparison_p_rec (PBasic (PArith i)) (reg_of_dreg r1)
+  | Pcset (rd, c) -> cset_rec (PBasic (PArith i)) (reg_of_ireg rd) c
+  | Pfmovi (fsz, rd, r1) -> fmovi_rec (PBasic (PArith i)) fsz (reg_of_freg rd) (reg_of_ireg0 r1)
   | Pcsel (rd, r1, r2, c) ->
-      csel_rec (reg_of_dreg rd) (reg_of_dreg r1) (reg_of_dreg r2) c
+      csel_rec (PBasic (PArith i)) (reg_of_dreg rd) (reg_of_dreg r1) (reg_of_dreg r2) c
   | Pfnmul (fsz, rd, r1, r2) ->
-      fnmul_rec fsz (reg_of_freg rd) (reg_of_freg r1) (reg_of_freg r2)
+      fnmul_rec (PBasic (PArith i)) fsz (reg_of_freg rd) (reg_of_freg r1) (reg_of_freg r2)
 
 let basic_rec i =
   match i with
   | PArith i' -> arith_rec i'
-  | PLoad (ld, rd, a) -> load_rec ld (reg_of_dreg rd) a
-  | PStore (st, r, a) -> store_rec st (reg_of_dreg r) a
-  | Pallocframe (sz, linkofs) -> allocframe_rec sz linkofs
-  | Pfreeframe (sz, linkofs) -> freeframe_rec sz linkofs
-  | Ploadsymbol (rd, id) -> loadsymbol_rec (reg_of_ireg rd) id
-  | Pcvtsw2x (rd, r1) -> cvtsw2x_rec (reg_of_ireg rd) (reg_of_ireg r1)
-  | Pcvtuw2x (rd, r1) -> cvtuw2x_rec (reg_of_ireg rd) (reg_of_ireg r1)
-  | Pcvtx2w rd -> cvtx2w_rec (reg_of_ireg rd)
-
-let builtin_rec ef args res =
+  | PLoad (ld, rd, a) -> load_rec (PBasic i) (reg_of_dreg rd) a
+  | PStore (st, r, a) -> store_rec (PBasic i) (reg_of_dreg r) a
+  | Pallocframe (sz, linkofs) -> allocframe_rec (PBasic i) sz linkofs
+  | Pfreeframe (sz, linkofs) -> freeframe_rec (PBasic i) sz linkofs
+  | Ploadsymbol (rd, id) -> loadsymbol_rec (PBasic i) (reg_of_ireg rd) id
+  | Pcvtsw2x (rd, r1) -> cvtsw2x_rec (PBasic i) (reg_of_ireg rd) (reg_of_ireg r1)
+  | Pcvtuw2x (rd, r1) -> cvtuw2x_rec (PBasic i) (reg_of_ireg rd) (reg_of_ireg r1)
+  | Pcvtx2w rd -> cvtx2w_rec (PBasic i) (reg_of_ireg rd)
+
+let builtin_rec i ef args res =
   {
-    inst = builtin_real;
+    inst = i;
     write_locs = [ Mem ];
     read_locs = [ Mem ];
     is_control = true;
@@ -479,84 +310,84 @@ let ctl_flow_rec i =
   match i with
   | Pb lbl ->
       {
-        inst = b_real;
+        inst = (PControl (PCtlFlow i));
         write_locs = [ reg_of_pc ];
         read_locs = [ reg_of_pc ];
         is_control = true;
       }
   | Pbc (c, lbl) ->
       {
-        inst = bc_real;
+        inst = (PControl (PCtlFlow i));
         write_locs = [ reg_of_pc ];
         read_locs = [ reg_of_pc ];
         is_control = true;
       }
   | Pbl (id, sg) ->
       {
-        inst = bl_real;
+        inst = (PControl (PCtlFlow i));
         write_locs = [ reg_of_ireg Asm.X30; reg_of_pc ];
         read_locs = [ reg_of_pc ];
         is_control = true;
       }
   | Pbs (id, sg) ->
       {
-        inst = bs_real;
+        inst = (PControl (PCtlFlow i));
         write_locs = [ reg_of_pc ];
         read_locs = [];
         is_control = true;
       }
   | Pblr (r, sg) ->
       {
-        inst = blr_real;
+        inst = (PControl (PCtlFlow i));
         write_locs = [ reg_of_ireg Asm.X30; reg_of_pc ];
         read_locs = [ reg_of_ireg r ];
         is_control = true;
       }
   | Pbr (r, sg) ->
       {
-        inst = br_real;
+        inst = (PControl (PCtlFlow i));
         write_locs = [ reg_of_pc ];
         read_locs = [ reg_of_ireg r ];
         is_control = true;
       }
   | Pret r ->
       {
-        inst = ret_real;
+        inst = (PControl (PCtlFlow i));
         write_locs = [ reg_of_pc ];
         read_locs = [ reg_of_ireg r ];
         is_control = true;
       }
   | Pcbnz (sz, r, lbl) ->
       {
-        inst = cbnz_real;
+        inst = (PControl (PCtlFlow i));
         write_locs = [ reg_of_pc ];
         read_locs = [ reg_of_ireg r; reg_of_pc ];
         is_control = true;
       }
   | Pcbz (sz, r, lbl) ->
       {
-        inst = cbz_real;
+        inst = (PControl (PCtlFlow i));
         write_locs = [ reg_of_pc ];
         read_locs = [ reg_of_ireg r; reg_of_pc ];
         is_control = true;
       }
   | Ptbnz (sz, r, n, lbl) ->
       {
-        inst = tbnz_real;
+        inst = (PControl (PCtlFlow i));
         write_locs = [ reg_of_pc ];
         read_locs = [ reg_of_ireg r; reg_of_pc ];
         is_control = true;
       }
   | Ptbz (sz, r, n, lbl) ->
       {
-        inst = tbz_real;
+        inst = (PControl (PCtlFlow i));
         write_locs = [ reg_of_pc ];
         read_locs = [ reg_of_ireg r; reg_of_pc ];
         is_control = true;
       }
   | Pbtbl (r1, tbl) ->
       {
-        inst = btbl_real;
+        inst = (PControl (PCtlFlow i));
         write_locs = [ reg_of_ireg Asm.X16; reg_of_ireg Asm.X17; reg_of_pc ];
         read_locs = [ reg_of_ireg r1; reg_of_pc ];
         is_control = true;
@@ -564,7 +395,7 @@ let ctl_flow_rec i =
 
 let control_rec i =
   match i with
-  | Pbuiltin (ef, args, res) -> builtin_rec ef args res
+  | Pbuiltin (ef, args, res) -> builtin_rec (PControl i) ef args res
   | PCtlFlow i' -> ctl_flow_rec i'
 
 let rec basic_recs body =