cmd/gc: fix floating point rounding again

Passes the expanded test in CL 100660044,
which gives me some confidence that it
might be right.

(The old code failed by not considering all the
low bits.)

LGTM=r
R=golang-codereviews, r, bradfitz
CC=golang-codereviews, iant, khr
https://golang.org/cl/99410051

src/cmd/gc/mparith3.c

@@ -203,7 +203,7 @@ static double
 mpgetfltN(Mpflt *a, int prec, int bias)
 {
 	int s, i, e, minexp;
-	uvlong v, vm;
+	uvlong v;
 	double f;
 
 	if(a->val.ovf && nsavederrors+nerrors == 0)
@@ -226,25 +226,23 @@ mpgetfltN(Mpflt *a, int prec, int bias)
 			return 0;
 	}
 
-	// pick up the mantissa and a rounding bit in a uvlong
-	s = prec+1;
+	// pick up the mantissa, a rounding bit, and a tie-breaking bit in a uvlong
+	s = prec+2;
 	v = 0;
 	for(i=Mpnorm-1; s>=Mpscale; i--) {
 		v = (v<<Mpscale) | a->val.a[i];
 		s -= Mpscale;
 	}
-	vm = v;
-	if(s > 0)
-		vm = (vm<<s) | (a->val.a[i]>>(Mpscale-s));
-
-	// continue with 64 more bits
-	s += 64;
-	for(; s>=Mpscale; i--) {
-		v = (v<<Mpscale) | a->val.a[i];
-		s -= Mpscale;
-	}
-	if(s > 0)
+	if(s > 0) {
 		v = (v<<s) | (a->val.a[i]>>(Mpscale-s));
+		if((a->val.a[i]&((1<<(Mpscale-s))-1)) != 0)
+			v |= 1;
+		i--;
+	}
+	for(; i >= 0; i--) {
+		if(a->val.a[i] != 0)
+			v |= 1;
+	}
 
 	// gradual underflow
 	e = Mpnorm*Mpscale + a->exp - prec;
@@ -253,23 +251,23 @@ mpgetfltN(Mpflt *a, int prec, int bias)
 		s = minexp - e;
 		if(s > prec+1)
 			s = prec+1;
-		v |= vm & ((1ULL<<s) - 1);
-		vm >>= s;
+		if((v & ((1<<s)-1)) != 0)
+			v |= 1<<s;
+		v >>= s;
 		e = minexp;
 	}
+	
+	// round to even
+	v |= (v&4)>>2;
+	v += v&1;
+	v >>= 2;
 
-//print("vm=%.16llux v=%.16llux\n", vm, v);
-	// round toward even
-	if(v != 0 || (vm&2ULL) != 0)
-		vm = (vm>>1) + (vm&1ULL);
-	else
-		vm >>= 1;
-
-	f = (double)(vm);
+	f = (double)(v);
 	f = ldexp(f, e);
 
 	if(a->val.neg)
 		f = -f;
+
 	return f;
 }