Hi,
This changes the docs to reflect the vectorizer's assumption that
scatter stores are ordered. Both, scatters and strided stores are
affected but for strided stores we can version the vectorized loop.
Right now RVV's strided stores are the only ones we emit and those
are unordered (while there are two variants for the gather/scatter) and
would require version for runtime strides. If we don't want to always
version (in case we support a future ordered strided store), we'll
need to introduce either two optabs like
mask_len_(un)ordered_strided_store or add a param like bias that can be
queried. I didn't want to decide for one yet so the patch just
documents the status quo.
Regards
Robin
gcc/ChangeLog:
* doc/md.texi: Document that scatters are left-to-right.
---
gcc/doc/md.texi | 16 ++++++++++------
1 file changed, 10 insertions(+), 6 deletions(-)
diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
index a98a572cc01..67c036d6bcb 100644
--- a/gcc/doc/md.texi
+++ b/gcc/doc/md.texi
@@ -7439,11 +7439,11 @@ Mask elements @var{i} with @var{i} > (operand 4 +
operand 5) are ignored.
@mdindex scatter_store@var{m}@var{n}
@item @samp{scatter_store@var{m}@var{n}}
-Store a vector of mode @var{m} into several distinct memory locations.
-Operand 0 is a scalar base address and operand 1 is a vector of mode
-@var{n} containing offsets from that base. Operand 4 is the vector of
-values that should be stored, which has the same number of elements as
-@var{n}. For each element index @var{i}:
+Store a vector of mode @var{m} into several, not necessarily distinct,
+memory locations. Operand 0 is a scalar base address and operand 1 is
+a vector of mode @var{n} containing offsets from that base. Operand 4
+is the vector of values that should be stored, which has the same number
+of elements as @var{n}. For each element index @var{i}:
@itemize @bullet
@item
@@ -7460,6 +7460,10 @@ store element @var{i} of operand 4 to that address.
The value of operand 2 does not matter if the offsets are already
address width.
+If two or more offset elements point to the same location, it is expected that
+the one with the highest element index is being written last. This implies a
+left-to-right ordering of the store.
+
@mdindex mask_scatter_store@var{m}@var{n}
@item @samp{mask_scatter_store@var{m}@var{n}}
Like @samp{scatter_store@var{m}@var{n}}, but takes an extra mask operand as
@@ -7476,7 +7480,7 @@ Mask elements @var{i} with @var{i} > (operand 6 + operand
7) are ignored.
@mdindex mask_len_strided_store@var{m}
@item @samp{mask_len_strided_store@var{m}}
-Store a vector of mode m into several distinct memory locations.
+Store a vector of mode m into several, not necessarily distinct, memory
locations.
Operand 0 is a scalar base address and operand 1 is scalar stride of Pmode.
Operand 2 is the vector of values that should be stored, which is of mode
@var{m}.
operand 3 is mask operand, operand 4 is length operand and operand 5 is bias
operand.
--
2.54.0
