+ 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 -
This commit is contained in:
florian 2005-10-03 12:31:31 +00:00
parent c15e0911e5
commit 6371333361
2 changed files with 27 additions and 4 deletions

View File

@ -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;

View File

@ -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;