[Ada] Long_Float'Read and Long_Float'Write use wrong subprogram

[Arnaud Charlet <charlet at adacore dot com>]

Several routines are provided by System.Stream_Attributes for floating
point types (W_SF, W_F, W_LF, W_LLF). The appropriate routine to call
for an attribute reference is determined using two criteria.
If a non-standard size or a stream size is specified, then we should
certainly use the size. But if we have two types the same (notably
Short_Float_Size = Float_Size which is close to universally true, and
Long_Long_Float_Size = Long_Float_Size, true on most targets except the
x86), then we would really rather use the root type, so that if people
want to fiddle with System.Stream_Attributes to get inter-target portable
streams, they get the size they expect. Consider in particular the case
of a stream written on an x86, with 96-bit Long_Long_Float being read
into a non-x86 target with 64 bit Long_Long_Float. A special version of
System.Stream_Attributes can deal with this provided the proper type is
always used.

This change fixes an inccorect test in the circuitry that implements
this selection scheme.

When compiled with -gnatG, the two attribute references in the below
subprogram must be expanded to calls to W_LF and I_LF, respectively.

$ cat p.adb
with Ada.Streams; use Ada.Streams;
procedure P
   (S : access Root_Stream_Type'Class;
    FI : Long_Float;
    FO : out Long_Float) is
begin
   Long_Float'Write (S, FI);
   Long_Float'Read  (S, FO);
end P;

$ gcc -c -gnatG p.adb
[...]
   $system__stream_attributes__w_lf (s, fi);
   fo := long_float!($system__stream_attributes__i_lf (s));
[...]

Tested on i686-pc-linux-gnu, committed on trunk

2008-08-20  Thomas Quinot  <quinot@adacore.com>

	* exp_strm.adb (Build_Elementary_Input_Call,
	Build_Elementary_Write_Call): Fix incorrect condition in circuitry that
	selects the stream attribute routines for long float types.

Attachment: difs
Description: Text document