https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63269
Bug ID: 63269
Summary: libgo/math test failures in TestLog2
Product: gcc
Version: 5.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: go
Assignee: ian at airs dot com
Reporter: vogt at linux dot vnet.ibm.com
CC: cmang at google dot com
There are several flaws in the math library test code. On s390x, the test
TestLog2 fails for three reasons:
1) This part of the test is too strict for i == 0:
for i := -1074; i <= 1023; i++ {
f := Ldexp(1, i)
l := Log2(f)
if l != float64(i) {
t.Errorf("Log2(2**%d) = %g, want %d", i, l, i)
}
}
It should really use the veryclose function() as there is no guarantee that the
result of this calculation from log10.go is exact:
return Log(frac)*(1/Ln2) + float64(exp)
With frac == 0.5 and exp == 1, the result may be just close to 0, not exactly
zero:
* If the compiler generated two instructions, one to mulitply and one to add,
the result is (with a small error delta):
Product: Log(0.5)*(1/Ln2) = -1 + delta -> rounded to -1
Sum: -1 + 1 = 0 -> rounded to 0
Result: 0
* If the compiler generates a single multiply and add instruction, only the
final result is rouced, i.e.
Product: Log(0.5)*(1/Ln2) = -1 + delta (not rounded)
Sum: -1 + delta + 1 = delta -> rounded to delta'
Result: delta' != 0
So, the fix is to use the "veryclose" error tolerance.
2) There's a bug in the tolerance() function:
if a != 0 {
e = e * a
if e < 0 {
e = -e
}
}
This snippet uses the defective value a but should use the expected value b
instead:
if b != 0 {
e = e * b
if e < 0 {
e = -e
}
}
Otherwise, bad things happen for an expected value 0 when the defective value
is very close, i.e. they never match.
3) The alike() function does allow an error tolerance. This kicks in on s390x
in the case that the expected value is 0 and the defective value is not
identical.
switch {
case IsNaN(a) && IsNaN(b):
return true
case a == b:
return Signbit(a) == Signbit(b)
}
A possible fix would be to detect this one situation and return true, and rely
on more specific tests to chack that the error is small ebough.
switch {
case IsNaN(a) && IsNaN(b):
return true
case a == 0 && !IsNaN(b) && !IsInf(b, 0):
// allow deviations when the expected value is zero
return true
case a == b:
return Signbit(a) == Signbit(b)
}
