Re: RIF: [PATCH] Improved driver for vectorizer testsuite

[Paolo Bonzini <bonzini at gnu dot org>]

>   compile tests for generic vectors; default dg-do action is 'run'
>   for each kind of hardware vectors supported by this target
>     compile tests for hardware vectors
>     if test system supports these vectors
>       default dg-do action is 'run
>     else
>       default dg-do action is 'compile'

I see two problems:

1) most hardware vector testcases will fail when generic vectors are 
used, if only because ints do not vectorize on 32-bit machines.

2) generic vector testcases do not require funky effective-target 
keywords as long as they reason on "generic" types such as "half a word" 
and "a quarter of a word".

In fact when applying your patch, almost all tests were failing with 
generic vectors, and with only a bunch of exceptions it is very 
difficult to modify them and make them pass for both generic and 
hardware vectorization.  It took me an hour to find five easy to modify 
testcases.

In other words, I'm not sure it is worth to keep the two testsuites as 
strictly unified as you proposed.

Also, regarding -msse vs. -msse2 vectorization, the problem in that case 
is that the various effective target keywords often hold only for one of 
them.  E.g., -msse does not have vect_double but -msse2 has.  This would 
either:

1) require really a lot of changes to the target-supports.exp code.

2) require changing the testcases so that they use #ifdef's (enabled 
with -D options that are passed together with -m options) instead of a 
proliferation of vect_* effective-target keywords.  Not that I would not 
like that, but it entails a lot of side work (and possibly duplicated on 
autovect-branch).

I still think that for now it is better to keep it simple and use my 
original patch.  It is possible to work incrementally from there, as you 
did indeed starting from my patch.

Paolo