I found two small problem on cvt.s.d emulation. 1. Converting a denormalized number does not raise Inexact exception. 2. Any denormalized double precision numbers are converted to zero regardless rounding mode. If rounding mode was "up", a positive denormalized double precision number should be converted to minimal (denormalized) single precision number. Here is a patch. --- linux-sgi-cvs/arch/mips/math-emu/sp_fdp.c Mon Oct 22 10:29:56 2001 +++ linux.new/arch/mips/math-emu/sp_fdp.c Thu Feb 14 18:56:06 2002 @@ -55,6 +55,10 @@ case IEEE754_CLASS_DNORM: /* cant possibly be sp representable */ SETCX(IEEE754_UNDERFLOW); + SETCX(IEEE754_INEXACT); + if ((ieee754_csr.rm == IEEE754_RU && !xs) || + (ieee754_csr.rm == IEEE754_RD && xs)) + return ieee754sp_xcpt(ieee754sp_mind(xs), "fdp", x); return ieee754sp_xcpt(ieee754sp_zero(xs), "fdp", x); case IEEE754_CLASS_NORM: break; --- Atsushi Nemoto
