On Sat, Oct 05, 2019 at 12:07:29PM -0700, Andres Freund wrote: > +#define EXPECT_EQ_U32(result_expr, expected_expr) \ > + do { \ > + uint32 result = (result_expr); \ > + uint32 expected = (expected_expr); \ > + if (result != expected) \ > + elog(ERROR, \ > + "%s yielded %u, expected %s in file \"%s\" line %u", \ > + #result_expr, result, #expected_expr, __FILE__, __LINE__); \ > + } while (0) > ... > > > I wonder if we should put these (and a few more, for other types), into > a more general place. I would like to have them for writing both tests > like regress.c:test_atomic_ops(), and for writing assertions that > actually display useful error messages. For the former it makes sense > to ERROR out, for the latter they ought to abort, as currently. > > Seems like putting ASSERT_{EQ,LT,...}_{U32,S32,...} (or Assert_Eq_..., > but that'd imo look weirder than the inconsistency) into c.h would make > sense, and EXPECT_ somewhere in common/pg_test.h or such?
Sounds reasonable. For broader use, I would include the expected value, not just expected_expr: elog(ERROR, \ "%s yielded %u, expected %s (%u) in file \"%s\" line %u", \ #result_expr, result, #expected_expr, expected, __FILE__, __LINE__); \ I didn't do that for the atomics tests, where expected_expr is always trivial. The codebase has plenty of Assert(x == y) where either of x or y could have the surprising value.