[Fortran] Disabling error tests and tests not yet working with HLFIR #47
Conversation
Patch required before enabling HLFIR by default: llvm/llvm-project#72090 Three batches of tests must be disabled. 1. test using non constant length character elemental fuctions 2. test using pointer assignments inside FORALL 3. "error" test that previously "wrongly" pass (did not throw errors) For the first two batches, lowering without HFLIR handled a few cases , but did not fully handle the features correctly. With HLFIR, it was decided to make all cases TODO until the feature are handled correctly. They are not very common, hence I think the "feature regression" is acceptable for now compared to what HLFIR brings. For, the last batch, the test suite is not set to run error checking tests (verifying a runtime error that is expected from the runtime given an invalid Fortran input), so the test suits considered a "pass" tests that where run with no error while invalid. HLFIR with -O1 uses the runtime in cases where the previous lowering did things inline, the usage of the runtime allow catching these invalid program as expected, but since the test suit is not set up yet to deal with negative runtime tests, they must be disabled.
|
I can see a lot of new (compared to the upstream HLFIR allow/deny lists) test failures running the testsuite with optimization enabled. It looks like something to do with Other than that, I can see that your lists differ a bit from the upstream alow/deny list. If the allow/deny list is just out of date with bugfixes then that isn't a problem, its probably not worth fixing now, as I presume we can remove these after HLFIR is default. |
Thanks, yes I agree this seems like something that should be fixed before enabling HLFIR by default. I will hold on.
Good point. I forgot about this list, it is a bit outdated (some HLFIR TODOs were implemented since making the HLFIR allow/deny list and the related tests are not disabled in this patch). I agree the HLFIR list should be updated or simply removed in a later patch. |
There was a problem hiding this comment.
My comment earlier was a false alarm. I'm sorry for the noise. I had built my flang-new with a broken clang, causing my flang-new to segfault in one of the HLFIR passes. Rebuilding flang-new with a newer clang resolved the issue.
For future reference, the issue was reproducible with flang-new -flang-experimental-hlfir -O3 Fortran/gfortran/torture/execute/intrinsic_transpose.f90 -o /dev/null.
Patch required before enabling HLFIR by default:
llvm/llvm-project#72090
Three batches of tests must be disabled.
For the first two batches, lowering without HFLIR handled a few cases , but did not fully handle the features correctly. With HLFIR, it was decided to make all cases TODO until the feature are handled correctly. They are not very common, hence I think the "feature regression" is acceptable for now compared to what HLFIR brings.
For, the last batch, the test suite is not set to run error checking tests (verifying a runtime error that is expected from the runtime given an invalid Fortran input), so the test suits considered a "pass" tests that where run with no error while invalid.
HLFIR with -O1 uses the runtime in cases where the previous lowering did things inline, the usage of the runtime allow catching these invalid program as expected, but since the test suit is not set up yet to deal with negative runtime tests, they must be disabled.