Changing weights system at asmblock level instead of asm

1 files changed, 83 insertions, 297 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 =