Hi,
In can_vec_perm_const_p if we cannot directly permute a vector mode we
try to pun it with a byte mode. qimode_for_vec_perm checks gets the
mode size and uses that as number of elements for the new QImode vector.
This doesn't work for RVV mask vectors, though: First, their
precision might be smaller than a byte and second, there is no
way to easily pun them. The most common way would be a vector select
from {0, 0, ...} and {1, 1, ...} vectors. Therefore this patch checks
if the perm's innermode precision is a multiple of QImode's precision.
I hit this during testing of a riscv backend patch related to
gather/scatter (to be upstreamed soon). It manifests in around 10-20
FAILs in the test suite so I figured we don't need a separate test.
Bootstrapped and regtested on x86 and power10. Regtested on aarch64 and
riscv64.
Regards
Robin
gcc/ChangeLog:
* optabs-query.cc (qimode_for_vec_perm): Check if QImode's
precision divides the inner mode's precision.
---
gcc/optabs-query.cc | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/gcc/optabs-query.cc b/gcc/optabs-query.cc
index 5335d0d8401..484d7c2e205 100644
--- a/gcc/optabs-query.cc
+++ b/gcc/optabs-query.cc
@@ -358,7 +358,9 @@ can_conditionally_move_p (machine_mode mode)
opt_machine_mode
qimode_for_vec_perm (machine_mode mode)
{
- if (GET_MODE_INNER (mode) != QImode)
+ if (GET_MODE_INNER (mode) != QImode
+ && multiple_p (GET_MODE_PRECISION (GET_MODE_INNER (mode)),
+ GET_MODE_PRECISION (QImode)))
return related_vector_mode (mode, QImode, GET_MODE_SIZE (mode));
return opt_machine_mode ();
}
--
2.51.0
