Comment author: @gasche

This is inconsistent with the rest of the testsuite, and it should
probably be replace by a test which checks the actual output of the compiler.

What would be the point of such a change? My idea in adding the test is that we might want to change the implementation (eg. to a more optimized one) later, and that this testcase could catch regression bugs at this time. Trying to test the compiler output seems harder, more fragile (you're sensitive to changes in heuristics), and likely to catch less bugs.

Comment author: @lpw25

What would be the point of such a change?

I guess it is just quite strange to have a single unit test in amongst all the non-unit tests. Every other test tests the behaviour of the compiler that has been built, whereas this test compiles some new code and tests the behaviour of that.

Perhaps it would make more sense if this test was testing the behaviour of the function within the appropriate compiler-libs. I know it wouldn't make much difference, but at least then it would be testing an output of the build process rather than testing the source directly.

Comment author: @gasche

I applied the patch you propose (BYTECODE_ONLY). Given that the rest of the suggestion seems hard to implement without weakening the safety guarantees, I'm fine with "strange" here and mark the bug resolved (but do feel free to complain here if the odd test does create another actual problem someday).