aboutsummaryrefslogtreecommitdiffstats
path: root/kvx/PostpassSchedulingOracle.ml
diff options
context:
space:
mode:
Diffstat (limited to 'kvx/PostpassSchedulingOracle.ml')
-rw-r--r--kvx/PostpassSchedulingOracle.ml33
1 files changed, 20 insertions, 13 deletions
diff --git a/kvx/PostpassSchedulingOracle.ml b/kvx/PostpassSchedulingOracle.ml
index 3f4520a6..3eb0b95f 100644
--- a/kvx/PostpassSchedulingOracle.ml
+++ b/kvx/PostpassSchedulingOracle.ml
@@ -34,6 +34,7 @@ type real_instruction =
| Addw | Andw | Compw | Mulw | Orw | Sbfw | Sbfxw | Sraw | Srlw | Sllw | Srsw | Rorw | Xorw
| Addd | Andd | Compd | Muld | Ord | Sbfd | Sbfxd | Srad | Srld | Slld | Srsd | Xord
| Nandw | Norw | Nxorw | Nandd | Nord | Nxord | Andnw | Ornw | Andnd | Ornd
+ | Abdw | Abdd
| Maddw | Maddd | Msbfw | Msbfd | Cmoved
| Make | Nop | Extfz | Extfs | Insf
| Addxw | Addxd
@@ -47,7 +48,8 @@ type real_instruction =
| Faddd | Faddw | Fsbfd | Fsbfw | Fmuld | Fmulw
| Fmind | Fminw | Fmaxd | Fmaxw | Finvw
| Ffmaw | Ffmad | Ffmsw | Ffmsd
- | Fnarrowdw | Fwidenlwd | Floatwz | Floatuwz | Floatdz | Floatudz | Fixedwz | Fixeduwz | Fixeddz | Fixedudz
+ | Fnarrowdw | Fwidenlwd | Floatwz | Floatuwz | Floatdz | Floatudz
+ | Fixedw | Fixeduw | Fixedd | Fixedud
| Fcompw | Fcompd
type ab_inst_rec = {
@@ -86,12 +88,12 @@ let arith_rr_real = function
| Pfloatuwrnsz -> Floatuwz
| Pfloatudrnsz -> Floatudz
| Pfloatdrnsz -> Floatdz
- | Pfixedwrzz -> Fixedwz
- | Pfixeduwrzz -> Fixeduwz
- | Pfixeddrzz -> Fixeddz
- | Pfixedudrzz -> Fixedudz
- | Pfixeddrzz_i32 -> Fixeddz
- | Pfixedudrzz_i32 -> Fixedudz
+ | Pfixedwrzz -> Fixedw
+ | Pfixeduwrzz -> Fixeduw
+ | Pfixeddrzz -> Fixedd
+ | Pfixedudrzz -> Fixedud
+ | Pfixeddrzz_i32 -> Fixedd
+ | Pfixedudrzz_i32 -> Fixedud
let arith_rrr_real = function
| Pcompw it -> Compw
@@ -142,6 +144,8 @@ let arith_rrr_real = function
| Pfminw -> Fminw
| Pfmaxd -> Fmaxd
| Pfmaxw -> Fmaxw
+ | Pabdw -> Abdw
+ | Pabdl -> Abdd
let arith_rri32_real = function
| Pcompiw it -> Compw
@@ -167,6 +171,7 @@ let arith_rri32_real = function
| Psrlil -> Srld
| Psrail -> Srad
| Psrxil -> Srsd
+ | Pabdiw -> Abdw
let arith_rri64_real = function
| Pcompil it -> Compd
@@ -183,6 +188,7 @@ let arith_rri64_real = function
| Pnxoril -> Nxord
| Pandnil -> Andnd
| Pornil -> Ornd
+ | Pabdil -> Abdd
let arith_arr_real = function
@@ -602,11 +608,11 @@ let rec_to_usage r =
(match encoding with None | Some U6 | Some S10 -> alu_lite
| Some U27L5 | Some U27L10 -> alu_lite_x
| Some E27U27L10 -> alu_lite_y)
- | Addxw ->
+ | Addxw | Abdw ->
(match encoding with None | Some U6 | Some S10 -> alu_lite
| Some U27L5 | Some U27L10 -> alu_lite_x
| _ -> raise InvalidEncoding)
- | Addxd ->
+ | Addxd | Abdd ->
(match encoding with None | Some U6 | Some S10 -> alu_lite
| Some U27L5 | Some U27L10 -> alu_lite_x
| Some E27U27L10 -> alu_lite_y)
@@ -643,7 +649,7 @@ let rec_to_usage r =
(* TODO: check *)
| Srsw | Srsd | Rorw -> (match encoding with None | Some U6 -> alu_lite | _ -> raise InvalidEncoding)
| Extfz | Extfs | Insf -> (match encoding with None -> alu_lite | _ -> raise InvalidEncoding)
- | Fixeduwz | Fixedwz | Floatwz | Floatuwz | Fixeddz | Fixedudz | Floatdz | Floatudz -> mau
+ | Fixeduw | Fixedw | Floatwz | Floatuwz | Fixedd | Fixedud | Floatdz | Floatudz -> mau
| Lbs | Lbz | Lhs | Lhz | Lws | Ld | Lq | Lo ->
(match encoding with None | Some U6 | Some S10 -> lsu_auxw
| Some U27L5 | Some U27L10 -> lsu_auxw_x
@@ -656,8 +662,8 @@ let rec_to_usage r =
| Get -> bcu_tiny_tiny_mau_xnop
| Fnegd | Fnegw | Fabsd | Fabsw | Fwidenlwd
| Fmind | Fmaxd | Fminw | Fmaxw -> alu_lite
- | Fnarrowdw -> alu_full
- | Faddd | Faddw | Fsbfd | Fsbfw | Fmuld | Fmulw | Finvw
+ | Finvw | Fnarrowdw -> alu_full
+ | Faddd | Faddw | Fsbfd | Fsbfw | Fmuld | Fmulw
| Ffmad | Ffmaw | Ffmsd | Ffmsw -> mau
@@ -680,8 +686,9 @@ let real_inst_to_latency = function
| Addd | Andd | Compd | Ord | Sbfd | Sbfxd | Srad | Srsd | Srld | Slld | Xord | Make
| Extfs | Extfz | Insf | Fcompw | Fcompd | Cmoved | Addxw | Addxd
| Fmind | Fmaxd | Fminw | Fmaxw
+ | Abdw | Abdd
-> 1
- | Floatwz | Floatuwz | Fixeduwz | Fixedwz | Floatdz | Floatudz | Fixeddz | Fixedudz -> 4
+ | Floatwz | Floatuwz | Fixeduw | Fixedw | Floatdz | Floatudz | Fixedd | Fixedud -> 4
| Mulw | Muld | Maddw | Maddd | Msbfw | Msbfd -> 2 (* FIXME - WORST CASE. If it's S10 then it's only 1 *)
| Lbs | Lbz | Lhs | Lhz | Lws | Ld | Lq | Lo -> 3
| Sb | Sh | Sw | Sd | Sq | So -> 1 (* See kvx-Optimization.pdf page 19 *)