https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104015
--- Comment #10 from rsandifo at gcc dot gnu.org <rsandifo at gcc dot gnu.org> --- (In reply to Kewen Lin from comment #9) > I totally agree this test case can be fragile when facing different > vectorisation strategies, but I'm not sure if leaving the exact number > checkings still has a bit value since this case seems the only case to catch > the redundant re-try (at least on Power?), once we fix it not to check any > times, we might miss some sensitive coverage on some useless re-tries, > though the compiling time influence is very very tiny (seems not a big deal > :)). Checking the number of tries might be useful, but if so, I think it should be done by a test that was written for that specific purpose. The tst can then be updated if the original way of checking the number of tries no longer worked for some reason. The test would likely be target-specific (since, for example, the expectations are different for Power 9 and Power 10). But this test was added to test how the SLP permute code handled groups of 3 elements. IMO that's the only thing that it should test. Even the existing target selectors are so complicated as to be almost incomprehensible. :-)
