This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
RIF: [PATCH] Improved driver for vectorizer testsuite
- From: "Bonzini Paolo" <paolo dot bonzini at lu dot unisi dot ch>
- To: "Janis Johnson" <janis187 at us dot ibm dot com>
- Cc: "GCC Patches" <gcc-patches at gcc dot gnu dot org>
- Date: Sat, 9 Apr 2005 10:11:27 +0200
- Subject: RIF: [PATCH] Improved driver for vectorizer testsuite
> Then again, perhaps there should be a new
> effective-target keyword that is true when we are compiling using
> target-specific options, and some tests can be skipped or run only
> when that is true.
That could be simd_hw: including it would *not* compile without target-specific options. I see several chicken-and-egg problems though:
1) Some generic vectorization tests would fail with hardware vectorization! For example, a tests checks that multiplication or 16-bit addition are *not* vectorized with generic vectors, because it's worthless. Tests such as this must not be compiled with hardware vectorization options.
2) I would add this effective-target keyword to all vectorization tests, because currently no test can pass if UNITS_PER_SIMD_WORD is zero. But many of them would work with generic vectorization, so I would have to undo a lot of changes when the generic vectorization patch goes in. Otherwise, I could modify vect.exp only after generic vectorization is approved, but this conflicts with the need to test generic vectorization.
3) On a separate side, vect_* detection may need improving when many sets of target options are passed. For example -msse does not support vect_double; still, dg-require-effective-target will let the test run.
Also, I'm new to DejaGNU so I prefer to do this in small steps. So, would it be ok to check-in the patch as is (as a "good start", and especially something that works independently of generic vectorization) and improve on it in a follow-up patch?
Paolo