From bdb169a04f3c42716bf98bdab661a509e287065d Mon Sep 17 00:00:00 2001
From: Xavier Leroy <xavier.leroy@inria.fr>
Date: Nov 01 2013 18:10:18 +0000
Subject: Follow-up to commit r14254: optimization of immediate division and modulus for ARM64.  Completely untested.



git-svn-id: http://caml.inria.fr/svn/ocaml/trunk@14257 f963ae5c-01c2-4b8c-9fe0-0dff7051ff02

---

diff --git a/asmcomp/arm64/emit.mlp b/asmcomp/arm64/emit.mlp
index fc9649c..ac8216c 100644
--- a/asmcomp/arm64/emit.mlp
+++ b/asmcomp/arm64/emit.mlp
@@ -454,17 +454,41 @@ let emit_instr i =
     | Lop(Iintop Imod) ->
         `	sdiv	{emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, {emit_reg i.arg.(1)}\n`;
         `	msub	{emit_reg i.res.(0)}, {emit_reg reg_tmp1}, {emit_reg i.arg.(1)}, {emit_reg i.arg.(0)}\n`
-    | Lop(Iintop_imm(Idiv, n)) -> (* n is a power of 2 *)
+    | Lop(Iintop_imm(Idiv, n)) ->
         let l = Misc.log2 n in
-        `	asr	{emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, #63\n`;
-        `	add	{emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, {emit_reg reg_tmp1}, lsr {emit_int (64-l)}\n`;
-        `	asr	{emit_reg i.res.(0)}, {emit_reg reg_tmp1}, {emit_int l}\n`
-    | Lop(Iintop_imm(Imod, n)) -> (* n is a power of 2 *)
+        if n = 1 lsl l then begin
+          `	asr	{emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, #63\n`;
+          `	add	{emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, {emit_reg reg_tmp1}, lsr {emit_int (64-l)}\n`;
+          `	asr	{emit_reg i.res.(0)}, {emit_reg reg_tmp1}, {emit_int l}\n`
+        end else begin
+          let (m, p) = Selectgen.divimm_parameters (Nativeint.of_int n) in
+          emit_intconst reg_tmp1 m;
+          `	smulh	{emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, {emit_reg reg_tmp1}\n`;
+          if m < 0n then
+            `	add	{emit_reg reg_tmp1}, {emit_reg reg_tmp1}, {emit_reg i.arg.(0)}\n`;
+          if p > 0 then
+            `	asr	{emit_reg reg_tmp1}, #{emit_int p}\n`;
+          `	add	{emit_reg i.res.(0)}, {emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, lsr 63\n`
+        end
+    | Lop(Iintop_imm(Imod, n)) ->
         let l = Misc.log2 n in
-        `	asr	{emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, #63\n`;
-        `	add	{emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, {emit_reg reg_tmp1}, lsr {emit_int (64-l)}\n`;
-        `	asr	{emit_reg reg_tmp1}, {emit_reg reg_tmp1}, {emit_int l}\n`;
-        `	sub	{emit_reg i.res.(0)}, {emit_reg i.arg.(0)}, {emit_reg reg_tmp1}, lsl {emit_int l}\n`
+        if n = 1 lsl l then begin
+          `	asr	{emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, #63\n`;
+          `	add	{emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, {emit_reg reg_tmp1}, lsr {emit_int (64-l)}\n`;
+          `	asr	{emit_reg reg_tmp1}, {emit_reg reg_tmp1}, {emit_int l}\n`;
+          `	sub	{emit_reg i.res.(0)}, {emit_reg i.arg.(0)}, {emit_reg reg_tmp1}, lsl {emit_int l}\n`
+        end else begin
+          let (m, p) = Selectgen.divimm_parameters (Nativeint.of_int n) in
+          emit_intconst reg_tmp1 m;
+          `	smulh	{emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, {emit_reg reg_tmp1}\n`;
+          if m < 0n then
+            `	add	{emit_reg reg_tmp1}, {emit_reg reg_tmp1}, {emit_reg i.arg.(0)}\n`;
+          if p > 0 then
+            `	asr	{emit_reg reg_tmp1}, #{emit_int p}\n`;
+          `	add	{emit_reg reg_tmp1}, {emit_reg reg_tmp1}, {emit_reg i.arg.(0)}, lsr 63\n`;
+          emit_intconst reg_tmp2 (Nativeint.of_int n);
+          `	msub	{emit_reg i.res.(0)}, {emit_reg reg_tmp1}, {emit_reg reg_tmp2}, {emit_reg i.arg.(0)}\n`
+        end
     | Lop(Iintop op) ->
         let instr = name_for_int_operation op in
         `	{emit_string instr}     {emit_reg i.res.(0)}, {emit_reg i.arg.(0)}, {emit_reg i.arg.(1)}\n`
diff --git a/asmcomp/arm64/selection.ml b/asmcomp/arm64/selection.ml
index c74b282..6af158f 100644
--- a/asmcomp/arm64/selection.ml
+++ b/asmcomp/arm64/selection.ml
@@ -194,17 +194,17 @@ method! select_operation op args =
           (Iintop Imul, args)
       end
   (* Division and modulus *)
-  (* Recognize (x / cst) and (x % cst) only if cst is a power of 2. *)
+  (* Recognize (x / cst) and (x % cst) only if cst is > 0. *)
   | Cdivi ->
       begin match args with
-      | [arg; Cconst_int n] when n = 1 lsl Misc.log2 n ->
+      | [arg; Cconst_int n] when n > 0 ->
           ((if n = 1 then Imove else Iintop_imm(Idiv, n)), [arg])
       | _ ->
           (Iintop Idiv, args)
       end
   | Cmodi ->
       begin match args with
-      | [arg; Cconst_int n] when n = 1 lsl Misc.log2 n ->
+      | [arg; Cconst_int n] when n > 0 ->
           ((if n = 1 then Iconst_int 0n else Iintop_imm(Imod, n)), [arg])
       | _ ->
           (Iintop Imod, args)