From 1b9f9287a34a825e2811a625354a772433ca132b Mon Sep 17 00:00:00 2001
From: Pierre Chambart <pierre.chambart@ocamlpro.org>
Date: Thu, 18 Oct 2012 12:11:08 +0200
Subject: [PATCH] Generate sqrtsd instruction on amd64

---
 asmcomp/amd64/arch.ml      |    6 ++++++
 asmcomp/amd64/emit.mlp     |    4 ++++
 asmcomp/amd64/emit_nt.mlp  |    4 ++++
 asmcomp/amd64/selection.ml |   10 ++++++++++
 4 files changed, 24 insertions(+)

diff --git a/asmcomp/amd64/arch.ml b/asmcomp/amd64/arch.ml
index 6d66129..13a27a7 100644
--- a/asmcomp/amd64/arch.ml
+++ b/asmcomp/amd64/arch.ml
@@ -38,6 +38,8 @@ type specific_operation =
   | Ioffset_loc of int * addressing_mode (* Add a constant to a location *)
   | Ifloatarithmem of float_operation * addressing_mode
                                        (* Float arith operation with memory *)
+  | Isqrtf                             (* floating-point square root *)
+  | Ifloatsqrtf of addressing_mode     (* floating-point square root from memory *)
 and float_operation =
     Ifloatadd | Ifloatsub | Ifloatmul | Ifloatdiv
 
@@ -102,6 +104,10 @@ let print_specific_operation printreg op ppf arg =
       fprintf ppf "[%a] := \"%s\"" (print_addressing printreg addr) arg lbl
   | Ioffset_loc(n, addr) ->
       fprintf ppf "[%a] +:= %i" (print_addressing printreg addr) arg n
+  | Isqrtf ->
+      fprintf ppf "sqrtf %a" printreg arg.(0)
+  | Ifloatsqrtf addr ->
+     fprintf ppf "sqrtf float64[%a]" (print_addressing printreg addr) [|arg.(0)|]
   | Ifloatarithmem(op, addr) ->
       let op_name = function
       | Ifloatadd -> "+f"
diff --git a/asmcomp/amd64/emit.mlp b/asmcomp/amd64/emit.mlp
index 06f6aba..05d41b3 100644
--- a/asmcomp/amd64/emit.mlp
+++ b/asmcomp/amd64/emit.mlp
@@ -543,6 +543,10 @@ let emit_instr fallthrough i =
         `	addq	${emit_int n}, {emit_addressing addr i.arg 0}\n`
     | Lop(Ispecific(Ifloatarithmem(op, addr))) ->
         `	{emit_string(instr_for_floatarithmem op)}	{emit_addressing addr i.arg 1}, {emit_reg i.res.(0)}\n`
+    | Lop(Ispecific Isqrtf) ->
+        `	sqrtsd	{emit_reg i.arg.(0)}, {emit_reg i.res.(0)}\n`
+    | Lop(Ispecific(Ifloatsqrtf addr)) ->
+        `	sqrtsd	{emit_addressing addr i.arg 0}, {emit_reg i.res.(0)}\n`
     | Lreloadretaddr ->
         ()
     | Lreturn ->
diff --git a/asmcomp/amd64/emit_nt.mlp b/asmcomp/amd64/emit_nt.mlp
index bc1750c..9fc4c13 100644
--- a/asmcomp/amd64/emit_nt.mlp
+++ b/asmcomp/amd64/emit_nt.mlp
@@ -537,6 +537,10 @@ let emit_instr fallthrough i =
         `	add	QWORD PTR {emit_addressing addr i.arg 0}, {emit_int n}\n`
     | Lop(Ispecific(Ifloatarithmem(op, addr))) ->
         `	{emit_string(instr_for_floatarithmem op)}	{emit_reg i.res.(0)}, REAL8 PTR {emit_addressing addr i.arg 1}\n`
+    | Lop(Ispecific Isqrtf) ->
+        `	sqrtsd	{emit_reg i.res.(0)}, {emit_reg i.arg.(0)}\n`
+    | Lop(Ispecific(Ifloatsqrtf addr)) ->
+        `	sqrtsd	{emit_reg i.res.(0)}, REAL8 PTR {emit_addressing addr i.arg 0}\n`
     | Lreloadretaddr ->
         ()
     | Lreturn ->
diff --git a/asmcomp/amd64/selection.ml b/asmcomp/amd64/selection.ml
index cc07133..6a4e731 100644
--- a/asmcomp/amd64/selection.ml
+++ b/asmcomp/amd64/selection.ml
@@ -182,6 +182,16 @@ method! select_operation op args =
       self#select_floatarith true Imulf Ifloatmul args
   | Cdivf ->
       self#select_floatarith false Idivf Ifloatdiv args
+  | Cextcall("sqrt", _, false, _) ->
+     begin match args with
+       [Cop(Cload (Double|Double_u as chunk), [loc])] ->
+         let (addr, arg) = self#select_addressing chunk loc in
+         (Ispecific(Ifloatsqrtf addr), [arg])
+     | [arg] ->
+         (Ispecific Isqrtf, [arg])
+     | _ ->
+         assert false
+     end
   (* Recognize store instructions *)
   | Cstore Word ->
       begin match args with
-- 
1.7.10.4