Atsushi Nemoto wrote:
Excuse me for too late comment.
On Mon, 12 Oct 2009 14:57:18 -0700, Chris Dearman <chris@xxxxxxxx> wrote:
From: Nigel Stephens <nigel@xxxxxxxx>
This patch ensures that the sign bit is always updated
for NaN operands.
...
@@ -76,15 +74,12 @@ ieee754dp ieee754dp_abs(ieee754dp x)
CLEARCX;
FLUSHXDP;
+ /* Clear sign ALWAYS, irrespective of NaN */
+ DPSIGN(x) = 0;
+
if (xc == IEEE754_CLASS_SNAN) {
- SETCX(IEEE754_INVALID_OPERATION);
- return ieee754dp_nanxcpt(ieee754dp_indef(), "neg");
+ return ieee754dp_nanxcpt(ieee754dp_indef(), "abs");
}
- if (ieee754dp_isnan(x)) /* but not infinity */
- return ieee754dp_nanxcpt(x, "abs", x);
-
- /* quick fix up */
- DPSIGN(x) = 0;
return x;
}
Is there any reason for removal of SETCX(IEEE754_INVALID_OPERATION)
line here?
The older version of this fix ("Fix absd emulation" by Raghu Gandham)
did not remove the line.
Without this line, a signaling NaN will not raise a signal. It seems
not expected behaviour.
ieee754dp/sp_nanxcpt also sets the invalid exception bit so I think this
is duplicated code. I think the same fix should have been applied to
ieee754sp_neg/ieee754dp_neg for consistency.
Chris
--
Chris Dearman Desk: +1 408 530 5092 Cell: +1 408 398 5531
MIPS Technologies Inc 955 East Arques Ave, Sunnyvale CA 94085
