disagreement in converting float32 result to float64 on 32-bit platform

[JeffBezanson]

This was seen on a 32-bit Xeon system:

julia> ndigf(n)=float64(log(float32(n)))

julia> ndigf(256)
5.5451774444736657

julia> float64(log(float32(256)))
5.5451774597167969

julia> log(256)
5.5451774444795623

The last answer is correct, the middle one is the float64 conversion of the correctly-rounded float32 answer, and the first one is something strange.

julia> num2hex(5.5451774444795623)
"40162e42fefa39ef"

julia> num2hex(5.5451774444736657)
"40162e42fefa2000"

Looks like the float64 answer with the last 13 bits zero'd.

Happens with some functions, e.g. log and sin, but not with sqrt.

[ViralBShah]

Might this have anything to do with the gcc optimization bug mentioned in the fdlibm readme?

[JeffBezanson]

Very likely. I tried the libm version and it doesn't have this issue.

[StefanKarpinski]

Jeff, can you check if this is fixed by compiling fdlibm with -O1 instead of -O2?

[JeffBezanson]

-O1 did not fix this, but it did fix the other one.

[JeffBezanson]

What am I smoking, it can't be the same bug, since this is calling logf, the float32 version, which already has the pointer cast bug fixed.

[JeffBezanson]

OK, this is just the classic extended precision issue. I can fix it by replacing

return y;

with

volatile float x;
x=y; return x;

in the fdlibm code.
Inside ndigf's llvm-generated code the result stays in an extended-precision register and so can be converted to float64 without first rounding to float32.
So what should we do? I could modify our code generator to force rounding of results returned by C functions.

[StefanKarpinski]

I think always rounding is the right thing to do. Consistency and determinism is more important than keeping extra bits around. This same issue has already caused both Java and PHP to have DOS bugs in the past few months, let's learn the lesson from them.

[ViralBShah]

I agree that let's go for consistency. That crlibm presentation discusses this issue.

Best to also check with Alan on this issue. He is probably not reading this thread.

-viral

On Jun 15, 2011, at 12:41 PM, StefanKarpinski wrote:

I think always rounding is the right thing to do. Consistency and determinism is more important than keeping extra bits around. This same issue has already caused both Java and PHP to have DOS bugs in the past few months, let's learn the lesson from them.

Reply to this email directly or view it on GitHub:
#41 (comment)

[StefanKarpinski]

He can't actually read this thread — this is private and AFAIK, Alan doesn't have a GitHub account.