On 04/11/2025 13:57, Christopher Bazley wrote:
On 28/10/2025 13:29, Richard Biener wrote:
Isn't SLP_TREE_CAN_USE_PARTIAL_VECTORS_P redundant given
SLP_TREE_CAN_USE_MASK_P || SLP_TREE_CAN_USE_LEN_P should be exactly
this?
SLP_TREE_CAN_USE_PARTIAL_VECTORS_P might be sth for an SLP instance
(or a subgraph with multiple entries (instances)) if we want to have
consistent len vs. mask use? (but I see no particular reason to force
consistency)
SLP_TREE_CAN_USE_PARTIAL_VECTORS_P is initialised to true and may
subsequently be set to false via vect_cannot_use_partial_vectors.
vect_analyze_stmt uses the value of SLP_TREE_CAN_USE_PARTIAL_VECTORS_P
to decided whether to return a failure result in cases where
tail-predication is required. If SLP_TREE_CAN_USE_MASK_P ||
SLP_TREE_CAN_USE_LEN_P were used for that purpose instead, it would
follow that neither SLP_TREE_CAN_USE_MASK_P nor SLP_TREE_CAN_USE_LEN_P
could be set to true in cases where vect_cannot_use_partial_vectors
might subsequently be called (which seems impossible because
vect_load_lanes_supported can be called with different values of
'count', and we cannot predict those values) , or else that
vect_cannot_use_partial_vectors would have to set both
SLP_TREE_CAN_USE_MASK_P and SLP_TREE_CAN_USE_LEN_P to false.
Setting both flags to false in vect_cannot_use_partial_vectors would
be trivial, but SLP_TREE_CAN_USE_PARTIAL_VECTORS_P has another
purpose: it gives the return value of vect_can_use_partial_vectors_p.
How did you envisage the return value of
vect_can_use_partial_vectors_p being decided for BB SLP? If it always
returns true then I think that the vectoriser might carry on trying to
use partial vectors when it should have already given up; if it
returns SLP_TREE_CAN_USE_MASK_P || SLP_TREE_CAN_USE_LEN_P then that
would prevent check_load_store_for_partial_vectors from being called
for the first time.
Together, the three flags allow the following states to be represented:
1. Might be able to operate on partial vectors, but don't yet know
whether we would use len or mask.
2. Might be able to operate on partial vectors with len.
3. Might be able to operate on partial vectors with mask.
4. (Invalid) Might be able to operate on partial vectors with both len
and mask.
5. Cannot operate on partial vectors.
6. (Strictly redundant) Cannot operate on partial vectors although we
previously thought we might be able to use len.
7. (Strictly redundant) Cannot operate on partial vectors although we
previously thought we might be able to use mask.
8. (Invalid) Cannot operate on partial vectors although we previously
thought we might be able to use both len and mask.
It would be impossible to encode the four states that neither invalid
nor redundant in only two bits. In any case, my goal was to keep the
new logic for BB SLP as close as possible to the existing logic for
loop vectorisation.
Sorry, it would clearly be possible to encode four values in 2 bits (00,
01, 10, 11) by encoding the "Might be able to operate on partial
vectors" state as both bits clear and "Cannot operate on partial
vectors" state as "Might be able to operate on partial vectors with
both len and mask" (both bits set). A reversal of the current encoding
would probably make more sense: set both 'len' and 'mask' bits at the
start, clear the 'mask' bit when 'len' is chosen, clear the 'len' bit
when 'mask' is chosen, and clear both bits when neither is valid. Maybe
this encoding could also be applied to the loop vectoriser.
Do you want me to make that change?
--
Christopher Bazley
Staff Software Engineer, GNU Tools Team.
Arm Ltd, 110 Fulbourn Road, Cambridge, CB1 9NJ, UK.
http://www.arm.com/
