Hi Richard,
>> However, I wonder if there shouldn't be some effective-target keyword for
>> this difference instead of the target-specific xfail (probably not worth
>> it) or if restricing the xfail like this is just a hack.
>
> What's the actual difference here? I expect we use versioning for alignment,
> so the effective target requirement on pr33804 looks bogus to me. Is it
> that we lack VnQImode addition support? (with n a multiple of 4)
>
> What's the difference in behavior of 32bit vs 64bit sparc?
comparing the dumps, I find:
* sparc:
pr33804.c:8:19: missed: op not supported by target.
using word mode not possible.
pr33804.c:9:18: missed: not vectorized: relevant stmt not supported: _6 = _3
+ _5;
pr33804.c:8:19: note: unsupported SLP instance starting from: *_2 = _6;
pr33804.c:8:19: missed: unsupported SLP instances
pr33804.c:8:19: missed: no array mode for V8QI[4]
pr33804.c:8:19: missed: no array mode for V8QI[4]
permutation op not supported by target.
pr33804.c:9:18: missed: unsupported grouped store
pr33804.c:6:6: note: vectorized 0 loops in function.
* sparcv9:
pr33804.c:8:19: note: ***** Choosing vector mode V8QI
pr33804.c:8:19: optimized: loop vectorized using 8 byte vectors and unroll
factor 2
pr33804.c:6:6: note: vectorized 1 loops in function.
Rainer
--
-----------------------------------------------------------------------------
Rainer Orth, Center for Biotechnology, Bielefeld University
