🚀 go-pugleaf

Hello,

	This is to report a strange behavior (bug) of Complex::sqrt()
	on IRIS Crimson. Since Steve Rumsby told me at comp.sys.sgi that
	he has the same problem with his R3000/Indigo, I suspect this is not
	a particular problem of R4000/Crimson.

	Symptom:
		Complex::sqrt() returns incorrect value. Moreover, each time
		it is called with an identical argument, it returns a
		different value.  Because of this, the test program
		tComplex.cc produces wrong output.

		I played with the method and found from printf(3) output that
		the return value of hypot(3) is incorrect. However, I still
		believe	hypot(3) is correct since it works fine when I test the
		routine with a test program (under native cc and gcc2.2.2).

	Note:
		Funny behaviors I noticed are:
		(1) g++ -O  and g++ -g  yields different result.
		(2) rewriting the code to a set of simpler expressions
		    using temporary variables makes -g version works.

		From these observations, I suspect if there is any
		proc-call convention problem or instruction scheduling
		problem.


	I would appreciate your help and please let me know if I could help
	you in any way. I'll be happy to send asm source etc.


Takashi Totsuka      totsuka@av.crl.sony.co.jp
		     SONY Corporation