Hi Adam,
The changes look good. With respect to the other coefficients, I agree
with you that it's probably not worth the trouble to make them into
constants. It was just that one specific literal value in each file
that jumped out at me. Thanks also for the overflow explanation. I
would agree that given the extent to which you are exercising that code
there would be some evidence if an overflow was causing issues.
--Jamil
On 12/7/2018 8:53 AM, Adam Petcher wrote:
Updated webrev: http://cr.openjdk.java.net/~apetcher/8208648/webrev.01/
Thanks for looking at this. See below.
On 12/6/2018 8:03 PM, Jamil Nimeh wrote:
Hi Adam, comments/questions below (mostly simple stuff, nothing major):
* IntegerPolynomial.java
o The comment block for multByInt should reflect your changes
you made, namely the removal of "r" from the signature.
Fixed.
o 88: In the case of multiplying two very large long values
would we ever see those cause an overflow. If so, is it OK
to have them overflow before the reduce operation?
The potential for overflow is the hardest part of all this, and any
overflow will cause serious problems. You can do some calculations to
ensure that overflow doesn't happen. For the multiplication, these
calculations are pretty simple and can be done manually. For an n-bit
representation, each limb has a maximum magnitude of 2^(n + 1) going
into the multiplication. If there are k limbs, then the maximum
magnitude of any limb after the multiplication is 2^(2n + 2 + log k).
There is no overflow during multiplication as long as this maximum is
strictly less than 2^63. For example, the P-521 field uses 19 limbs
with 28 bits each. So the maximum limb magnitude after multiplication
is 2^(2*28 + 2 + log 19) =~ 2^(62.25).
Doing these sorts of calculations to ensure that there is no overflow
during the carry/reduce sequence is much more tedious. I have a model
of this arithmetic that does all this calculation for me, and that is
what I used to ensure there is no overflow. Also, the unit test
(TestIntegerModuloP.java) does a large number of random operations,
and I suspect that it would be very unlikely for an accidental
overflow to get past that test.
o
o 420-425: Looks like this was a copy of the comment block from
conditionalSwap(). Maybe needs to be tailored to what the
conditionalAssign method does?
Fixed.
o IntegerPolynomialP256.java, IntegerPolynomialP384.java,
IntegerPolynomialP521.java
+ In the carryReduce0 and carryReduce methods, you have
many uses of an integer literal (33554432 for P256,
134217728 for P384/521). Should these be made as private
static final long values?
o P256OrderField.java, P384OrderField.java, P521OrderField.java
+ Similar question about using a static final long vs.
repeated int literal values in the carryReduce* methods.
o With respect to the last two main bullet items: Would
conversion to a static final value be difficult due to the
fact that they are generated from Fieldgen.jsh?
I changed the literal that was used during the carry operations to a
class constant called CARRY_ADD, and made it more clear what it is (it
is 1 << n for some n). The only literals remaining are the
coefficients that are used in reduction. Making these constants is a
bit more complicated, because there are several cases to consider, and
there are optimizations for each of these cases to make the code more
clear. For example, if the coefficient is 1, then don't bother
multiplying by it. I can make these class constants, too, but I'm not
sure it is worth the trouble. Let me know what you think.
Changing the generator to produce better generated code is no problem,
so please let me know if you find more opportunities to make this code
more clear.
