[M3commit] CVS Update: cm3

[Jay Krell]

CVSROOT:	/usr/cvs
Changes by:	jkrell at birch.	10/06/13 04:48:54

Modified files:
	cm3/m3-sys/cminstall/src/config-no-install/: ALPHA_OSF 
	                                             Unix.common 

Log message:
	-mieee is needed to support denormal floating point, such
	as m3-libs/m3core/src/float/IEEE/LongReal.i3
	MinPos: T = 4.9406564584124654D-324;
	(* The minimum positive value in T. *)
	
	MinPosNormal: T = 2.2250738585072014D-308;
	(* The minimum positive "normal" value in T; differs from MinPos
	only for implementations with denormalized numbers. *)
	
	Otherwise MinPos is NaN and multiplication with it here:
	
	m3-libs/arithmetic/src/basictypes/float/FloatTrans.ig
	Tiny = R.MinPos * FLOAT(1000.0, T); (* nearly 0.0 *)
	
	issues a SIGFPE, when R = LongReal.
	
	I also wonder if maybe MinPos should be adjusted to be normal
	or ALPHA_OSF should do so (i.e. not use generic IEEE directory, e.g.:
	VAX/LongReal.i3:  MinPos: T = 2.93873587705571880D-39;
	VAX/LongReal.i3:  MinPosNormal: T = MinPos;
	)
	
	Probably best to be like all the other architectures though,
	even if it costs some perf. You know, some people like floating point
	to be really really the same across all architectures..witness
	Java, Java -strictfp, and the lack of any Java for VAX.
	
	see:
	http://gcc.gnu.org/onlinedocs/gcc-4.5.0/gcc/DEC-Alpha-Options.html#DEC-Alpha-Options
	
	A program like this demontrates it, I think it was, from memory:
	
	int main()
	{
	long a = 0x3E8; /* NaN */
	double b = *(double*)&a;
	double c = b * 1000;;
	return 0;
	}
	
	or:
	int main()
	{
	double a = 4.9406564584124654e-324;
	double b = a * 1000;
	return 0;
	}
	
	We should probably hardcode this in parse.c for Alpha instead
	of relying on the config files.