fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-02 05:48:07 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

ir3/sched: Don't schedule collect early

Browse source 
I don't think there was ever a good reason to do this, but when we start
folding constants/immediates into collect, this can become actively
harmful.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9842>
This commit is contained in:
Connor Abbott 2021-03-11 15:00:15 +01:00 • committed by Emma Anholt
parent 27593cb241
commit 8b15c2f30c
1 changed files with 10 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

11
src/freedreno/ir3/ir3_sched.c
View file

@ -746,7 +746,16 @@ choose_instr_prio(struct ir3_sched_ctx *ctx, struct ir3_sched_notes *notes)
struct ir3_sched_node *chosen = NULL;
foreach_sched_node (n, &ctx->dag->heads) {
if (!is_meta(n->instr))
/*
* - phi nodes and inputs must be scheduled first
* - split should be scheduled first, so that the vector value is
* killed as soon as possible. RA cannot split up the vector and
* reuse components that have been killed until it's been killed.
* - collect, on the other hand, should be treated as a "normal"
* instruction, and may add to register pressure if its sources are
* part of another vector or immediates.
*/
if (!is_meta(n->instr) || n->instr->opc == OPC_META_COLLECT)
continue;
if (!chosen || (chosen->max_delay < n->max_delay))