[all]: Rework Neon load/store and load/store pair

Currently Neon load/store is split in two instructions: one for handling immediate/register offset and another for handling post-index variant. The similar applies to Neon load/store pair. Note, missing support for pre-index variant for Neon load/store. Let's refactor those instructions to take advantage of existing infrastructure for handling addressing modes. That removes need for special instruction for post-index variant and as a bonus naturally brings support for pre-index variant for Neon load/store. Signed-off-by: Vladimir Murzin <[email protected]>
herd · Mar 15, 2024 · 2dd1e78 · 2dd1e78
1 parent d9f2328
commit 2dd1e78
Show file tree

Hide file tree

Showing 7 changed files with 299 additions and 450 deletions.
diff --git a/gen/AArch64Compile_gen.ml b/gen/AArch64Compile_gen.ml
@@ -582,7 +582,7 @@ module Make(Cfg:Config) : XXXCompile_gen.S =
       let emit_load_reg temporal st init rA =
         let r1,st = next_vreg st in
         let r2,st = next_vreg st in
-        let ldp = [I_LDP_SIMD(temporal,A64.VSIMD32,to_scalar r1,to_scalar r2,rA,0)] in
+        let ldp = [I_LDP_SIMD(temporal,A64.VSIMD32,to_scalar r1,to_scalar r2,rA,(0,A64.Idx))] in
         let r3,st = next_vreg st in
         let add = [I_ADD_SIMD (r3,r1,r2)] in
         let rX,st = next_reg st in
@@ -603,7 +603,7 @@ module Make(Cfg:Config) : XXXCompile_gen.S =
     module LDUR = struct
       let emit_load_reg st init rA =
         let r,st = next_scalar_reg st in
-        let ldur = [I_LDUR_SIMD(A64.VSIMD32,r,rA,None)] in
+        let ldur = [I_LDUR_SIMD(A64.VSIMD32,r,rA,0)] in
         let rX,st = next_reg st in
         let fmov = [I_FMOV_TG(A64.V32,rX,A64.VSIMD32,r)] in
         rX,init,lift_code (ldur@fmov),st
@@ -622,7 +622,7 @@ module Make(Cfg:Config) : XXXCompile_gen.S =
     module LDAPUR = struct
       let emit_load_reg st init rA =
         let r,st = next_scalar_reg st in
-        let ldur = [I_LDAPUR_SIMD(A64.VSIMD32,r,rA,None)] in
+        let ldur = [I_LDAPUR_SIMD(A64.VSIMD32,r,rA,0)] in
         let rX,st = next_reg st in
         let fmov = [I_FMOV_TG(A64.V32,rX,A64.VSIMD32,r)] in
         rX,init,lift_code (ldur@fmov),st
@@ -834,7 +834,7 @@ module Make(Cfg:Config) : XXXCompile_gen.S =
         let r1,st = next_vreg st in
         let r2,st = next_vreg st in
         let movi = List.mapi (fun i r -> movi_reg r (v+i)) [r1;r2] in
-        let stp = [I_STP_SIMD(temporal,A64.VSIMD32,to_scalar r1,to_scalar r2,rA,0)] in
+        let stp = [I_STP_SIMD(temporal,A64.VSIMD32,to_scalar r1,to_scalar r2,rA,(0,A64.Idx))] in
         init,pseudo movi@pseudo stp,st
 
       let emit_store n st p init loc v =
@@ -854,15 +854,15 @@ module Make(Cfg:Config) : XXXCompile_gen.S =
         let r2,st = next_vreg st in
         let movi = List.mapi (fun i r -> movi_reg r (v+i)) [r1;r2] in
         let adds = List.map (fun v -> add_simd v rB) [r1;r2] in
-        let stp = [I_STP_SIMD(temporal,A64.VSIMD32,to_scalar r1,to_scalar r2,rA,0)] in
+        let stp = [I_STP_SIMD(temporal,A64.VSIMD32,to_scalar r1,to_scalar r2,rA,(0,A64.Idx))] in
         init,lift_code(dup@movi@adds@stp),st
     end
 
     module STUR = struct
       let emit_store_reg st init rA v =
         let r,st = next_vreg st in
         let movi = [movi_reg r v] in
-        let stur = [I_STUR_SIMD(A64.VSIMD32,to_scalar r,rA,None)] in
+        let stur = [I_STUR_SIMD(A64.VSIMD32,to_scalar r,rA,0)] in
         init,lift_code(movi@stur),st
 
       let emit_store st p init loc v =
@@ -881,15 +881,15 @@ module Make(Cfg:Config) : XXXCompile_gen.S =
         let r1,st = next_vreg st in
         let movi = [movi_reg r1 v] in
         let adds = [add_simd r1 rB]in
-        let stur = [I_STUR_SIMD(A64.VSIMD32,to_scalar r1,rA,None)] in
+        let stur = [I_STUR_SIMD(A64.VSIMD32,to_scalar r1,rA,0)] in
         init,lift_code(dup@movi@adds@stur),st
     end
 
     module STLUR = struct
       let emit_store_reg st init rA v =
         let r,st = next_vreg st in
         let movi = [movi_reg r v] in
-        let stlur = [I_STLUR_SIMD(A64.VSIMD32,to_scalar r,rA,None)] in
+        let stlur = [I_STLUR_SIMD(A64.VSIMD32,to_scalar r,rA,0)] in
         init,lift_code(movi@stlur),st
 
       let emit_store st p init loc v =
@@ -908,7 +908,7 @@ module Make(Cfg:Config) : XXXCompile_gen.S =
         let r1,st = next_vreg st in
         let movi = [movi_reg r1 v] in
         let adds = [add_simd r1 rB]in
-        let stlur = [I_STLUR_SIMD(A64.VSIMD32,to_scalar r1,rA,None)] in
+        let stlur = [I_STLUR_SIMD(A64.VSIMD32,to_scalar r1,rA,0)] in
         init,lift_code(dup@movi@adds@stlur),st
     end
 

diff --git a/herd/AArch64Arch_herd.ml b/herd/AArch64Arch_herd.ml
@@ -90,16 +90,16 @@ module Make (C:Arch_herd.Config)(V:Value.AArch64) =
     | I_EOR_SIMD _| I_ERET| I_FENCE _| I_GC _| I_IC _| I_LD1 _| I_LD1M _| I_LD1R _ | I_LDAP1 _
     | I_LD2 _| I_LD2M _| I_LD2R _| I_LD3 _| I_LD3M _| I_LD3R _| I_LD4 _| I_LD4M _
     | I_LD4R _| I_LDAR _| I_LDARBH _| I_LDCT _| I_LDG _| I_LDOP _| I_LDOPBH _
-    | I_LDP _| I_LDP_P_SIMD _| I_LDP_SIMD _| I_LDPSW _| I_LDR _
-    | I_LDRSW _ | I_LDR_P_SIMD _ | I_LDAPUR_SIMD _
-    | I_LDR_SIMD _| I_LDRBH _| I_LDRS _| I_LDUR _| I_LDUR_SIMD _| I_LDXP _| I_MOV _ | I_FMOV_TG _
+    | I_LDP _| I_LDP_SIMD _| I_LDPSW _| I_LDR _
+    | I_LDRSW _ | I_LDR_SIMD _ | I_LDAPUR_SIMD _
+    | I_LDRBH _| I_LDRS _| I_LDUR _| I_LDUR_SIMD _| I_LDXP _| I_MOV _ | I_FMOV_TG _
     | I_ADDV _| I_DUP _ | I_MOV_FG _| I_MOV_S _| I_MOV_TG _| I_MOV_V _| I_MOV_VE _| I_MOVI_S _
     | I_MOVI_V _| I_MOVK _| I_MOVZ _| I_MOVN _| I_MRS _| I_MSR _| I_OP3 _| I_RBIT _
     | I_RET _
     | I_SBFM _| I_SC _| I_SEAL _| I_ST1 _| I_STL1 _| I_ST1M _| I_ST2 _| I_ST2M _| I_ST3 _
     | I_ST3M _| I_ST4 _| I_ST4M _| I_STCT _| I_STG _| I_STLR _| I_STLRBH _| I_STOP _
-    | I_STOPBH _| I_STP _| I_STP_P_SIMD _| I_STP_SIMD _| I_STR _ | I_STLUR_SIMD _
-    | I_STR_P_SIMD _| I_STR_SIMD _| I_STRBH _| I_STUR_SIMD _| I_STXP _| I_STXR _
+    | I_STOPBH _| I_STP _| I_STP_SIMD _| I_STR _ | I_STLUR_SIMD _
+    | I_STR_SIMD _| I_STRBH _| I_STUR_SIMD _| I_STXP _| I_STXR _
     | I_STXRBH _| I_STZG _| I_STZ2G _
     | I_SWP _| I_SWPBH _| I_SXTW _| I_TLBI _| I_UBFM _
     | I_UDF _| I_UNSEAL _ | I_ADDSUBEXT _ | I_ABS _ | I_REV _ | I_EXTR _
@@ -229,10 +229,10 @@ module Make (C:Arch_herd.Config)(V:Value.AArch64) =
       | I_LDOP (_,v,_,_,_,_) | I_STOP (_,v,_,_,_) ->
           Some (tr_variant v)
       | I_STZG _|I_STZ2G _ -> Some MachSize.granule
-      | I_LDR_SIMD (v,_,_,_,_) | I_LDR_P_SIMD (v,_,_,_)
-      | I_LDP_SIMD (_,v,_,_,_,_) | I_LDP_P_SIMD (_,v,_,_,_,_)
-      | I_STR_SIMD (v,_,_,_,_) | I_STR_P_SIMD (v,_,_,_)
-      | I_STP_SIMD (_,v,_,_,_,_) | I_STP_P_SIMD (_,v,_,_,_,_)
+      | I_LDR_SIMD (v,_,_,_)
+      | I_LDP_SIMD (_,v,_,_,_,_)
+      | I_STR_SIMD (v,_,_,_)
+      | I_STP_SIMD (_,v,_,_,_,_)
       | I_LDUR_SIMD (v,_,_,_) | I_STUR_SIMD (v,_,_,_)
       | I_LDAPUR_SIMD (v,_,_,_) | I_STLUR_SIMD (v,_,_,_) ->
           Some (tr_simd_variant v)
@@ -344,10 +344,8 @@ module Make (C:Arch_herd.Config)(V:Value.AArch64) =
       | I_LD3R _|I_LD4 _|I_LD4M _|I_LD4R _
       | I_ST1 _|I_ST1M _|I_ST2 _|I_ST2M _
       | I_ST3 _|I_ST3M _|I_ST4 _|I_ST4M _
-      | I_LDP_P_SIMD _|I_STP_P_SIMD _
       | I_LDP_SIMD _|I_STP_SIMD _
-      | I_LDR_SIMD _|I_LDR_P_SIMD _
-      | I_STR_SIMD _|I_STR_P_SIMD _
+      | I_LDR_SIMD _|I_STR_SIMD _
       | I_LDUR_SIMD _|I_LDAPUR_SIMD _|I_STUR_SIMD _|I_STLUR_SIMD _
       | I_MOV_VE _
       | I_MOV_V _|I_MOV_TG _|I_MOV_FG _
@@ -376,10 +374,8 @@ module Make (C:Arch_herd.Config)(V:Value.AArch64) =
       | I_LD4 _|I_LD4M _|I_LD4R _|I_ST1 _|I_STL1 _
       | I_ST1M _|I_ST2 _|I_ST2M _|I_ST3 _
       | I_ST3M _|I_ST4 _|I_ST4M _
-      | I_LDP_P_SIMD _|I_STP_P_SIMD _
       | I_LDP_SIMD _|I_STP_SIMD _
-      | I_LDR_SIMD _|I_LDR_P_SIMD _
-      | I_STR_SIMD _|I_STR_P_SIMD _
+      | I_LDR_SIMD _|I_STR_SIMD _
       | I_LDUR_SIMD _|I_LDAPUR_SIMD _|I_STUR_SIMD _|I_STLUR_SIMD _
       | I_MOV_VE _
       | I_MOV_V _|I_MOV_TG _|I_MOV_FG _

diff --git a/herd/AArch64Sem.ml b/herd/AArch64Sem.ml
@@ -1890,9 +1890,8 @@ module Make
       let simd_ldr = do_simd_ldr  Annot.N
       let simd_ldar = do_simd_ldr  Annot.Q
 
-      let do_simd_str an sz rs rd kr s ii =
-        get_ea rs kr s ii >>|
-        read_reg_neon true rd ii >>= fun (addr, v) ->
+      let do_simd_str an sz addr rd ii =
+        read_reg_neon true rd ii >>= fun v ->
         if sz == MachSize.S128 then
           do_write_mem_2_ops sz an aexp Access.VIR addr v ii >>= B.next2T
         else
@@ -1901,9 +1900,8 @@ module Make
       let simd_str = do_simd_str Annot.N
       let simd_stlr = do_simd_str Annot.L
 
-      let simd_str_p sz rs rd k ii =
-        read_reg_ord rs ii >>|
-        read_reg_neon true rd ii >>= fun (addr, v) ->
+      let simd_str_p sz addr rd rs k ii =
+        read_reg_neon true rd ii >>= fun v ->
         if sz == MachSize.S128 then
           (* 128-bit Neon LDR/STR and friends are split into two 64-bit
            * single-copy atomic accesses. *)
@@ -2625,57 +2623,85 @@ module Make
             !!!!(read_reg_ord rA ii >>= fun addr ->
             (store_m addr rs ii >>|
             post_kr rA addr kr ii))
-
-        | I_LDR_SIMD(var,r1,rA,kr,s) ->
+        | I_LDR_SIMD(var,r1,rA,MemExt.Reg(v,kr,sext,s)) ->
+            let access_size = tr_simd_variant var in
+            get_ea_reg rA v kr sext s ii >>= fun addr ->
+            simd_ldr access_size addr r1 ii >>= B.next1T
+        | I_LDR_SIMD(var,r1,rA,MemExt.Imm (k,Idx)) ->
+            let access_size = tr_simd_variant var in
+            get_ea_idx rA k ii >>= fun addr ->
+            simd_ldr access_size addr r1 ii >>= B.next1T
+        | I_LDR_SIMD(var,r1,rA,MemExt.Imm (k,PreIdx)) ->
             let access_size = tr_simd_variant var in
-            get_ea rA kr s ii >>= fun addr ->
+            get_ea_preindexed rA k ii >>= fun addr ->
             simd_ldr access_size addr r1 ii >>= B.next1T
-        | I_LDR_P_SIMD(var,r1,rA,k) ->
+        | I_LDR_SIMD(var,r1,rA,MemExt.Imm (k,PostIdx)) ->
             let access_size = tr_simd_variant var in
             read_reg_ord rA ii >>= fun addr ->
             simd_ldr access_size addr r1 ii >>|
             post_kr rA addr (K k) ii >>= B.next2T
         | I_LDUR_SIMD(var,r1,rA,k) ->
-            let access_size = tr_simd_variant var and
-            k = K (match k with Some k -> k | None -> 0) in
-            (get_ea rA k S_NOEXT ii >>= fun addr ->
-            simd_ldr access_size addr r1 ii) >>= B.next1T
+            let access_size = tr_simd_variant var in
+            get_ea rA (K k) S_NOEXT ii >>= fun addr ->
+            simd_ldr access_size addr r1 ii >>= B.next1T
         | I_LDAPUR_SIMD(var,r1,rA,k) ->
-            let access_size = tr_simd_variant var and
-            k = K (match k with Some k -> k | None -> 0) in
-            (get_ea rA k S_NOEXT ii >>= fun addr ->
-            simd_ldar access_size addr r1 ii) >>= B.next1T
-        | I_STR_SIMD(var,r1,rA,kr,s) ->
             let access_size = tr_simd_variant var in
-            simd_str access_size rA r1 kr s ii
-        | I_STR_P_SIMD(var,r1,rA,k) ->
+            get_ea rA (K k) S_NOEXT ii >>= fun addr ->
+            simd_ldar access_size addr r1 ii >>= B.next1T
+        | I_STR_SIMD(var,r1,rA,MemExt.Reg (v,kr,sext,s)) ->
+            let access_size = tr_simd_variant var in
+            get_ea_reg rA v kr sext s ii >>= fun addr ->
+            simd_str access_size addr r1 ii
+        | I_STR_SIMD(var,r1,rA,MemExt.Imm (k,Idx)) ->
+            let access_size = tr_simd_variant var in
+            get_ea_idx rA k ii >>= fun addr ->
+            simd_str access_size addr r1 ii
+        | I_STR_SIMD(var,r1,rA,MemExt.Imm (k,PreIdx)) ->
             let access_size = tr_simd_variant var in
-            simd_str_p access_size rA r1 (K k) ii
+            get_ea_preindexed rA k ii >>= fun addr ->
+            simd_str access_size addr r1 ii
+        | I_STR_SIMD(var,r1,rA,MemExt.Imm (k,PostIdx)) ->
+            let access_size = tr_simd_variant var in
+            read_reg_ord rA ii >>= fun addr ->
+            simd_str_p access_size addr r1 rA (K k) ii
         | I_STUR_SIMD(var,r1,rA,k) ->
-            let access_size = tr_simd_variant var and
-            k = K (match k with Some k -> k | None -> 0) in
-            simd_str access_size rA r1 k S_NOEXT ii
+            let access_size = tr_simd_variant var in
+            get_ea_idx rA k ii >>= fun addr ->
+            simd_str access_size addr r1 ii
         | I_STLUR_SIMD(var,r1,rA,k) ->
-            let access_size = tr_simd_variant var and
-            k = K (match k with Some k -> k | None -> 0) in
-            simd_stlr access_size rA r1 k S_NOEXT ii
-        | I_LDP_SIMD(tnt,var,r1,r2,r3,k) ->
-            get_ea_idx r3 k ii >>= fun addr ->
-            simd_ldp tnt var addr r1 r2 ii
-        | I_LDP_P_SIMD(tnt,var,r1,r2,r3,k) ->
-            read_reg_ord r3 ii >>= fun addr ->
-            (simd_ldp tnt var addr r1 r2 ii >>|
-            post_kr r3 addr (K k) ii) >>=
-            fun (b,()) -> M.unitT b
-        | I_STP_SIMD(tnt,var,r1,r2,r3,k) ->
-            get_ea_idx r3 k ii >>= fun addr ->
-            simd_stp tnt var addr r1 r2 ii
-        | I_STP_P_SIMD(tnt,var,r1,r2,r3,k) ->
-            read_reg_ord r3 ii >>= fun addr ->
-            simd_stp tnt var addr r1 r2 ii >>|
-            post_kr r3 addr (K k) ii >>=
-            fun (b,()) -> M.unitT b
-
+            let access_size = tr_simd_variant var in
+            get_ea_idx rA k ii >>= fun addr ->
+            simd_stlr access_size addr r1 ii
+        | I_LDP_SIMD(tnt,var,r1,r2,r3,idx) ->
+          begin
+            match idx with
+            | k,Idx ->
+                get_ea_idx r3 k ii >>= fun addr ->
+                  simd_ldp tnt var addr r1 r2 ii
+            | k,PreIdx ->
+                get_ea_preindexed r3 k ii >>= fun addr ->
+                  simd_ldp tnt var addr r1 r2 ii
+            | k,PostIdx ->
+                read_reg_ord r3 ii >>= fun addr ->
+                  (simd_ldp tnt var addr r1 r2 ii >>|
+                  post_kr r3 addr (K k) ii) >>=
+                  fun (b,()) -> M.unitT b
+          end
+        | I_STP_SIMD(tnt,var,r1,r2,r3,idx) ->
+          begin
+            match idx with
+            | k,Idx ->
+                get_ea_idx r3 k ii >>= fun addr ->
+                  simd_stp tnt var addr r1 r2 ii
+            | k,PreIdx ->
+                get_ea_preindexed r3 k ii >>= fun addr ->
+                  simd_stp tnt var addr r1 r2 ii
+            | k,PostIdx ->
+                read_reg_ord r3 ii >>= fun addr ->
+                  (simd_stp tnt var addr r1 r2 ii >>|
+                    post_kr r3 addr (K k) ii) >>=
+                  fun (b,()) -> M.unitT b
+          end
         (* Morello instructions *)
         | I_ALIGND(rd,rn,k) ->
             check_morello inst ;

diff --git a/jingle/AArch64Arch_jingle.ml b/jingle/AArch64Arch_jingle.ml
@@ -659,10 +659,9 @@ include Arch.MakeArch(struct
     | I_ST2 _ | I_ST2M _
     | I_ST3 _ | I_ST3M _
     | I_ST4 _ | I_ST4M _
-    | I_LDP_SIMD _ | I_LDP_P_SIMD _
-    | I_STP_SIMD _ | I_STP_P_SIMD _
-    | I_LDR_SIMD _ | I_LDR_P_SIMD _
-    | I_STR_SIMD _ | I_STR_P_SIMD _
+    | I_LDP_SIMD _
+    | I_STP_SIMD _
+    | I_LDR_SIMD _ | I_STR_SIMD _
     | I_LDUR_SIMD _ | I_LDAPUR_SIMD _
     | I_STUR_SIMD _ | I_STLUR_SIMD _
     | I_ADDV _ | I_DUP _ | I_FMOV_TG _