[Mesa-dev] [PATCH 3/4] nir: take cross-thread operations into account into a few places

[Connor Abbott]

These optimizations happened to work with derivatives, but they won't
with upcoming shader_ballot and group_vote instructions.

Signed-off-by: Connor Abbott <cwabbott0 at gmail.com>
---
 src/compiler/nir/nir_instr_set.c           | 12 ++++++++++++
 src/compiler/nir/nir_opt_peephole_select.c | 11 +++++++++++
 2 files changed, 23 insertions(+)

diff --git a/src/compiler/nir/nir_instr_set.c b/src/compiler/nir/nir_instr_set.c
index 9cb9ed4..74309d6 100644
--- a/src/compiler/nir/nir_instr_set.c
+++ b/src/compiler/nir/nir_instr_set.c
@@ -178,6 +178,18 @@ hash_instr(const void *data)
    const nir_instr *instr = data;
    uint32_t hash = _mesa_fnv32_1a_offset_bias;
 
+   /*
+    * If an instruction is cross-thread but not convergent, hash its block. If
+    * an instruction is convergent, then we can always replace one invocation
+    * with another since every invocation is guaranteed convergent. But not so
+    * for non-convergent instructions, since different invocations may be
+    * called with different execution maskes and therefore have different
+    * results.
+    */
+   if (nir_instr_is_cross_thread(instr) && !nir_instr_is_convergent(instr)) {
+      HASH(hash, instr->block);
+   }
+
    switch (instr->type) {
    case nir_instr_type_alu:
       hash = hash_alu(hash, nir_instr_as_alu(instr));
diff --git a/src/compiler/nir/nir_opt_peephole_select.c b/src/compiler/nir/nir_opt_peephole_select.c
index 4ca4f80..ce41781 100644
--- a/src/compiler/nir/nir_opt_peephole_select.c
+++ b/src/compiler/nir/nir_opt_peephole_select.c
@@ -61,6 +61,17 @@ static bool
 block_check_for_allowed_instrs(nir_block *block, unsigned *count, bool alu_ok)
 {
    nir_foreach_instr(instr, block) {
+      if (nir_instr_is_cross_thread(instr) && !nir_instr_is_convergent(instr)) {
+         /* If the instruction is cross-thread, then we can't execute it
+          * conditionally when we would've executed it unconditionally before,
+          * except when the condition is uniform. If the instruction is
+          * convergent, though, we're already guaranteed that the entire
+          * region is convergent (including the condition) so we can go ahead.
+          *
+          * TODO: allow when the if-condition is uniform
+          */
+         return false;
+      }
       switch (instr->type) {
       case nir_instr_type_intrinsic: {
          nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
-- 
2.9.3