This implements costing of vector construction and decomposition
to a custom (possibly vector) element type to be used for vectorizer
costing of the corresponding operations when dealing with
VMAT_STRIDED_SLP loads and stores.
Bootstrapped and tested on x86_64-unknown-linux-gnu.
I hope this is reasonable to follow. OK?
Thanks,
Richard.
* config/i386/i386.cc (ix86_vector_cd_cost): New function,
enhanced and split out from ...
(ix86_default_vector_cost): ... here.
---
gcc/config/i386/i386.cc | 49 ++++++++++++++++++++++-------------------
1 file changed, 26 insertions(+), 23 deletions(-)
diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
index 3f22f029f5e..7b85fc8c472 100644
--- a/gcc/config/i386/i386.cc
+++ b/gcc/config/i386/i386.cc
@@ -25589,6 +25589,31 @@ asm_preferred_eh_data_format (int code, int global)
return DW_EH_PE_absptr;
}
+/* Cost of constructing or destructing a vector in VECMODE from/to elements
+ of ELMODE. */
+static int
+ix86_vector_cd_cost (machine_mode vecmode, machine_mode elmode)
+{
+ if (GET_MODE_BITSIZE (vecmode) < 128)
+ return ((GET_MODE_BITSIZE (vecmode) / GET_MODE_BITSIZE (elmode) - 1)
+ * ix86_cost->sse_op);
+
+ int n = GET_MODE_BITSIZE (vecmode) / 128;
+ int cost = 0;
+ /* Element inserts/extracts into/from N SSE vectors, the possible
+ GPR <-> XMM moves have to be accounted for elsewhere. */
+ if (GET_MODE_BITSIZE (elmode) < 128)
+ cost += n * (128 / GET_MODE_BITSIZE (elmode) - 1) * ix86_cost->sse_op;
+ if (GET_MODE_BITSIZE (vecmode) >= 256
+ && GET_MODE_BITSIZE (elmode) < 256)
+ /* N/2 vinserti128/vextracti128 for SSE <-> AVX256. */
+ cost += n * ix86_vec_cost (V32QImode, ix86_cost->sse_op) / 2;
+ if (GET_MODE_BITSIZE (vecmode) == 512)
+ /* One vinserti64x4/vextracti64*4 for AVX256 <-> AVX512. */
+ cost += ix86_vec_cost (vecmode, ix86_cost->sse_op);
+ return cost;
+}
+
/* Worker for ix86_builtin_vectorization_cost and the fallback calls
from ix86_vector_costs::add_stmt_cost. */
static int
@@ -25679,29 +25704,7 @@ ix86_default_vector_cost (enum vect_cost_for_stmt
type_of_cost,
return ix86_vec_cost (mode, ix86_cost->sse_op);
case vec_construct:
- {
- int n = GET_MODE_NUNITS (mode);
- /* N - 1 element inserts into an SSE vector, the possible
- GPR -> XMM move is accounted for in add_stmt_cost. */
- if (GET_MODE_BITSIZE (mode) <= 128)
- return (n - 1) * ix86_cost->sse_op;
- /* One vinserti128 for combining two SSE vectors for AVX256. */
- else if (GET_MODE_BITSIZE (mode) == 256)
- return ((n - 2) * ix86_cost->sse_op
- + ix86_vec_cost (mode, ix86_cost->sse_op));
- /* One vinserti64x4 and two vinserti128 for combining SSE
- and AVX256 vectors to AVX512. */
- else if (GET_MODE_BITSIZE (mode) == 512)
- {
- machine_mode half_mode
- = mode_for_vector (GET_MODE_INNER (mode),
- GET_MODE_NUNITS (mode) / 2).require ();
- return ((n - 4) * ix86_cost->sse_op
- + 2 * ix86_vec_cost (half_mode, ix86_cost->sse_op)
- + ix86_vec_cost (mode, ix86_cost->sse_op));
- }
- gcc_unreachable ();
- }
+ return ix86_vector_cd_cost (mode, GET_MODE_INNER (mode));
default:
gcc_unreachable ();
--
2.51.0
