Account for multiple errors in throwExceptionIfNecessary.

OpenSSL's error queue is somewhat difficult to use. It may push multiple
errors, not just a single one, as kind of a stack trace of sorts. It's
also unclear who is responsible for clearing the error queue after a
failed operation. (Changing this in BoringSSL is difficult due to API
constraints. See https://crbug.com/boringssl/38.)

There's three bugs here.

1. libcore's NativeBN code was not accounting for multiple errors in the
   queue. ERR_get_error popped one error, but there needs to be an
   ERR_clear_error call to clear everything.

2. libcore's NativeBN code checks whether the error queue is empty to
   determine whether to fail. This isn't quite right as it makes it
   sensitive to unrelated code failing to clear the error queue. (I also
   suspect there are codepaths in BoringSSL where failures don't hit the
   error queue just because it's so easy to forget.)

3. A recent change in BoringSSL caused BN_mod_inverse to push two errors
   rather than one if the input did not have an inverse. This caused 1
   and 2 to combine and manifest as a test failure in a later test.

   Although two errors is perfectly legitimate per OpenSSL's API,
   pushing ERR_R_INTERNAL_ERROR behind BN_R_NO_INVERSE doesn't make much
   sense, so https://boringssl-review.googlesource.com/c/10483/ brings it
   back to just the one error.

This CL fixes the first of three issues and should be sufficient to fix
the immediate test failure. The second will be fixed as a follow-up.

Bug: 30917411
Test: run cts -c org.apache.harmony.tests.java.math.BigIntegerTest
Change-Id: Ie3b3c2097cc85473a250e40efdd61729d04b9ba8

0 files changed, 0 insertions, 0 deletions