Googletest export

Improve DoubleNearPredFormat output on bad epsilons

DoubleNearPredFormat will happily accept epsilon values (abs_error) that
are so small that they are meaningless. This turns EXPECT_NEAR into a
complicated and non-obvious version of EXPECT_EQ.

This change modifies DoubleNearPredFormat) so that when there is a
failure it calculates the smallest meaningful epsilon value, given the
input values, and then prints a message which explains what happened.

If a true equality test is wanted either pass a literal 0.0 as abs_error
or use EXPECT_EQ. If a check for being almost equal is wanted consider
using EXPECT_DOUBLE_EQ which, contrary to its name, verifies that the
two numbers are *almost* equal (within four ULPs).

With this change the flaky test mentioned in crbug.com/786046 gives this
output:

The difference between 4.2934311416234112e+18 and 4.2934311416234107e+18 is 512, where
4.2934311416234112e+18 evaluates to 4.2934311416234112e+18,
4.2934311416234107e+18 evaluates to 4.2934311416234107e+18.
The abs_error parameter 1.0 evaluates to 1 which is smaller than the minimum distance between doubles for numbers of this magnitude which is 512, thus making this EXPECT_NEAR check equivalent to EXPECT_EQUAL. Consider using EXPECT_DOUBLE_EQ instead.

Tested:
I confirmed that this change detects the bad epsilon value that caused
crbug.com/786046 in Chromium and added a test for the desired output.
PiperOrigin-RevId: 332946880
This commit is contained in:
Abseil Team 2020-09-21 18:22:01 -04:00 committed by vslashg
parent 7aca84427f
commit b5687db554
2 changed files with 31 additions and 0 deletions

View File

@ -1516,6 +1516,30 @@ AssertionResult DoubleNearPredFormat(const char* expr1,
const double diff = fabs(val1 - val2);
if (diff <= abs_error) return AssertionSuccess();
// Find the value which is closest to zero.
const double min_abs = std::min(fabs(val1), fabs(val2));
// Find the distance to the next double from that value.
const double epsilon =
nextafter(min_abs, std::numeric_limits<double>::infinity()) - min_abs;
// Detect the case where abs_error is so small that EXPECT_NEAR is
// effectively the same as EXPECT_EQUAL, and give an informative error
// message so that the situation can be more easily understood without
// requiring exotic floating-point knowledge.
// Don't do an epsilon check if abs_error is zero because that implies
// that an equality check was actually intended.
if (!isnan(val1) && !isnan(val2) && abs_error > 0 && abs_error < epsilon) {
return AssertionFailure()
<< "The difference between " << expr1 << " and " << expr2 << " is "
<< diff << ", where\n"
<< expr1 << " evaluates to " << val1 << ",\n"
<< expr2 << " evaluates to " << val2 << ".\nThe abs_error parameter "
<< abs_error_expr << " evaluates to " << abs_error
<< " which is smaller than the minimum distance between doubles for "
"numbers of this magnitude which is "
<< epsilon
<< ", thus making this EXPECT_NEAR check equivalent to "
"EXPECT_EQUAL. Consider using EXPECT_DOUBLE_EQ instead.";
}
return AssertionFailure()
<< "The difference between " << expr1 << " and " << expr2
<< " is " << diff << ", which exceeds " << abs_error_expr << ", where\n"

View File

@ -3084,6 +3084,13 @@ TEST_F(DoubleTest, EXPECT_NEAR) {
EXPECT_NONFATAL_FAILURE(EXPECT_NEAR(1.0, 1.5, 0.25), // NOLINT
"The difference between 1.0 and 1.5 is 0.5, "
"which exceeds 0.25");
// At this magnitude adjacent doubles are 512.0 apart, so this triggers a
// slightly different failure reporting path.
EXPECT_NONFATAL_FAILURE(
EXPECT_NEAR(4.2934311416234112e+18, 4.2934311416234107e+18, 1.0),
"The abs_error parameter 1.0 evaluates to 1 which is smaller than the "
"minimum distance between doubles for numbers of this magnitude which is "
"512");
}
// Tests ASSERT_NEAR.