From cb45120556938070a623b056867545301bce8cac Mon Sep 17 00:00:00 2001
From: Connor Abbott <cwabbott0@gmail.com>
Date: Fri, 17 Dec 2021 19:48:49 +0100
Subject: [PATCH] ir3: Use (ss) for instructions writing shared regs

The blob uses *both* nops and (ss). It turns out that in some rare cases
the hardware does take more than 6 cycles, at least for movmsk, but
adding nops is unnecessary. I believe the extra nops are only there due
to the immaturity of the blob's implementation of subgroup ops, so we
don't have to copy them - just handle shared reg producers the same as
SFU instructions.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14246>
---
 src/freedreno/ir3/ir3.h          | 12 +++++++++++-
 src/freedreno/ir3/ir3_delay.c    |  2 +-
 src/freedreno/ir3/ir3_legalize.c |  5 +++++
 3 files changed, 17 insertions(+), 2 deletions(-)

diff --git a/src/freedreno/ir3/ir3.h b/src/freedreno/ir3/ir3.h
index e1d6399..932f85a 100644
--- a/src/freedreno/ir3/ir3.h
+++ b/src/freedreno/ir3/ir3.h
@@ -1670,6 +1670,10 @@ is_local_mem_load(struct ir3_instruction *instr)
 static inline bool
 is_ss_producer(struct ir3_instruction *instr)
 {
+   foreach_dst (dst, instr) {
+      if (dst->flags & IR3_REG_SHARED)
+         return true;
+   }
    return is_sfu(instr) || is_local_mem_load(instr);
 }
 
@@ -1687,7 +1691,13 @@ soft_ss_delay(struct ir3_instruction *instr)
     * and so on. Not quite sure where it tapers out (ie. how many warps share an
     * SFU unit). But 10 seems like a reasonable # to choose:
     */
-   return 10;
+   if (is_sfu(instr) || is_local_mem_load(instr))
+      return 10;
+
+   /* The blob adds 6 nops between shared producers and consumers, and before we
+    * used (ss) this was sufficient in most cases.
+    */
+   return 6;
 }
 
 static inline bool
diff --git a/src/freedreno/ir3/ir3_delay.c b/src/freedreno/ir3/ir3_delay.c
index 83730b5..054f4c8 100644
--- a/src/freedreno/ir3/ir3_delay.c
+++ b/src/freedreno/ir3/ir3_delay.c
@@ -76,7 +76,7 @@ ir3_delayslots(struct ir3_instruction *assigner,
 
    /* assigner must be alu: */
    if (is_flow(consumer) || is_sfu(consumer) || is_tex(consumer) ||
-       is_mem(consumer) || (assigner->dsts[0]->flags & IR3_REG_SHARED)) {
+       is_mem(consumer)) {
       return 6;
    } else {
       /* In mergedregs mode, there is an extra 2-cycle penalty when half of
diff --git a/src/freedreno/ir3/ir3_legalize.c b/src/freedreno/ir3/ir3_legalize.c
index bf8906f..f46312d 100644
--- a/src/freedreno/ir3/ir3_legalize.c
+++ b/src/freedreno/ir3/ir3_legalize.c
@@ -255,6 +255,11 @@ legalize_block(struct ir3_legalize_ctx *ctx, struct ir3_block *block)
       if (is_sfu(n))
          regmask_set(&state->needs_ss, n->dsts[0]);
 
+      foreach_dst (dst, n) {
+         if (dst->flags & IR3_REG_SHARED)
+            regmask_set(&state->needs_ss, dst);
+      }
+
       if (is_tex_or_prefetch(n)) {
          regmask_set(&state->needs_sy, n->dsts[0]);
          if (n->opc == OPC_META_TEX_PREFETCH)
-- 
2.7.4