From 637133336187452f8d726e305770cd07262db8e3 Mon Sep 17 00:00:00 2001
From: florian <florian@freepascal.org>
Date: Mon, 3 Oct 2005 12:31:31 +0000
Subject: [PATCH] + sqr uses sse on x86 if possible * the result type of sqr is
 equal to the argument in case of usual floats

git-svn-id: trunk@1265 -
---
 compiler/ninl.pas        | 17 +++++++++++++++--
 compiler/x86/nx86inl.pas | 14 ++++++++++++--
 2 files changed, 27 insertions(+), 4 deletions(-)

diff --git a/compiler/ninl.pas b/compiler/ninl.pas
index 24a6989bf5..dee1edf137 100644
--- a/compiler/ninl.pas
+++ b/compiler/ninl.pas
@@ -1206,6 +1206,20 @@ implementation
           end;
 
 
+      procedure setfloatresulttype;
+        begin
+          if (left.resulttype.def.deftype=floatdef) and
+            (tfloatdef(left.resulttype.def).typ in [s32real,s64real,s80real,s128real]) then
+            resulttype:=left.resulttype
+          else
+            begin
+              inserttypeconv(left,pbestrealtype^);
+              resulttype:=pbestrealtype^;
+            end;
+        end;
+
+
+
       var
          vl,vl2    : TConstExprInt;
          vr        : bestreal;
@@ -1935,8 +1949,7 @@ implementation
                   else
                    begin
                      set_varstate(left,vs_used,[vsf_must_be_valid]);
-                     inserttypeconv(left,pbestrealtype^);
-                     resulttype:=pbestrealtype^;
+                     setfloatresulttype;
                    end;
                 end;
 
diff --git a/compiler/x86/nx86inl.pas b/compiler/x86/nx86inl.pas
index 9d065f3d06..d587e7f6bd 100644
--- a/compiler/x86/nx86inl.pas
+++ b/compiler/x86/nx86inl.pas
@@ -246,8 +246,18 @@ implementation
      procedure tx86inlinenode.second_sqr_real;
 
        begin
-         load_fpu_location;
-         emit_reg_reg(A_FMUL,S_NO,NR_ST0,NR_ST0);
+         if use_sse(resulttype.def) then
+           begin
+             secondpass(left);
+             location_force_mmregscalar(exprasmlist,left.location,false);
+             location:=left.location;
+             cg.a_opmm_loc_reg(exprasmlist,OP_MUL,left.location.size,left.location,left.location.register,mms_movescalar);
+           end
+         else
+           begin
+             load_fpu_location;
+             emit_reg_reg(A_FMUL,S_NO,NR_ST0,NR_ST0);
+           end;
        end;
 
      procedure tx86inlinenode.second_sqrt_real;