Hi,
for a generic vec_construct from scalar elements we need
to load each scalar element and move it over to a vector register.
This patch uses register-move cost and scalar_to_vec and multiplies
it with the number of elements in the vector.
This helps vectorization of e.g. x264 SATD with the default
-mvector-strict-align.
Regtested on rv64gcv_zvl512b.
Regards
Robin
PR target/118019
gcc/ChangeLog:
* config/riscv/riscv.cc (riscv_builtin_vectorization_cost):
Increase vec_construct cost.
---
gcc/config/riscv/riscv.cc | 8 +++++++-
1 file changed, 7 insertions(+), 1 deletion(-)
diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc
index be2ebf9d9c0..aa8a4562d9a 100644
--- a/gcc/config/riscv/riscv.cc
+++ b/gcc/config/riscv/riscv.cc
@@ -12263,7 +12263,13 @@ riscv_builtin_vectorization_cost (enum
vect_cost_for_stmt type_of_cost,
return fp ? common_costs->fp_stmt_cost : common_costs->int_stmt_cost;
case vec_construct:
- return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vectype));
+ {
+ /* TODO: This is too pessimistic in case we can splat. */
+ int regmove_cost = fp ? costs->regmove->FR2VR
+ : costs->regmove->GR2VR;
+ return (regmove_cost + common_costs->scalar_to_vec_cost)
+ * estimated_poly_value (TYPE_VECTOR_SUBPARTS (vectype));
+ }
default:
gcc_unreachable ();
--
2.47.1
