jay/lower_scoreboard: be the sole emitter of SYNC

this gets closer to something we can schedule and avoids some pointless syncs.

Totals from 491 (18.55% of 2647) affected shaders:
Instrs: 602994 -> 602946 (-0.01%)
CodeSize: 9063888 -> 9015904 (-0.53%)

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41398>

src/intel/compiler/jay/jay_from_nir.c

@@ -681,8 +681,6 @@ jay_emit_memory_barrier(struct nir_to_jay_state *nj, nir_intrinsic_instr *intr)
 {
    nir_variable_mode modes = nir_intrinsic_memory_modes(intr);
 
-   jay_SYNC(&nj->bld, TGL_SYNC_ALLWR);
-
    if (modes & nir_var_image) {
       emit_lsc_fence(nj, intr, BRW_SFID_TGM);
       assert(!nj->nir->info.use_lowered_image_to_global && "fix common code");
@@ -703,10 +701,8 @@ jay_emit_memory_barrier(struct nir_to_jay_state *nj, nir_intrinsic_instr *intr)
 }
 
 static void
-jay_emit_signal_barrier(struct nir_to_jay_state *nj)
+jay_emit_signal_barrier(jay_builder *b, struct nir_to_jay_state *nj)
 {
-   jay_builder *b = &nj->bld;
-
    /* Signal barrier / Active threads only (BSpec 72052).
     *
     * Source 0 is the number of subgroups in [31:24], which comes from the u0.2
@@ -729,8 +725,6 @@ jay_emit_signal_barrier(struct nir_to_jay_state *nj)
    jay_SEND(b, .sfid = BRW_SFID_MESSAGE_GATEWAY,
             .msg_desc = BRW_MESSAGE_GATEWAY_SFID_BARRIER_MSG, .srcs = &zipped,
             .nr_srcs = 1, .type = JAY_TYPE_U32, .uniform = true);
-
-   jay_SYNC(b, TGL_SYNC_BAR);
 }
 
 static void
@@ -1194,24 +1188,22 @@ jay_emit_intrinsic(struct nir_to_jay_state *nj, nir_intrinsic_instr *intr)
       break;
    }
 
-   case nir_intrinsic_barrier:
+   case nir_intrinsic_barrier: {
+      jay_SCHEDULE_BARRIER(b);
+
       if (nir_intrinsic_memory_scope(intr) != SCOPE_NONE) {
          jay_emit_memory_barrier(nj, intr);
       }
 
-      if (cs) {
-         if (nir_intrinsic_execution_scope(intr) == SCOPE_WORKGROUP) {
-            if (jay_workgroup_is_one_subgroup(b, nj->nir)) {
-               // XXX: when we have a scheduler, jay_SCHEDULE_BARRIER(b);
-            } else {
-               jay_emit_signal_barrier(nj);
-               s->prog_data->cs.uses_barrier = true;
-            }
-         }
-      } else {
-         // XXX: when we have a scheduler, jay_SCHEDULE_BARRIER(b);
+      if ((cs && nir_intrinsic_execution_scope(intr) == SCOPE_WORKGROUP) &&
+          !jay_workgroup_is_one_subgroup(b, nj->nir)) {
+
+         jay_emit_signal_barrier(b, nj);
+         s->prog_data->cs.uses_barrier = true;
       }
+
       break;
+   }
 
    case nir_intrinsic_begin_invocation_interlock:
    case nir_intrinsic_end_invocation_interlock:

src/intel/compiler/jay/jay_lower_scoreboard.c

@@ -80,6 +80,16 @@ lower_send_local(jay_function *func, jay_block *block)
          }
       }
 
+      /* Lower ordering barriers */
+      if (I->op == JAY_OPCODE_SCHEDULE_BARRIER) {
+         if (busy) {
+            jay_SYNC(&b, TGL_SYNC_ALLWR);
+            busy = 0;
+         }
+
+         jay_remove_instruction(I);
+      }
+
       if (I->op == JAY_OPCODE_SEND && !jay_send_eot(I)) {
          unsigned sbid = (roundrobin++) % NUM_TOKENS;
          jay_set_send_sbid(I, sbid);
@@ -97,6 +107,12 @@ lower_send_local(jay_function *func, jay_block *block)
             struct gpr_range src = def_to_gpr(func, I, I->src[s]);
             BITSET_SET_COUNT(tokens[sbid].reading, src.base, src.width);
          }
+
+         /* Barriers are non-EOT gateway messages. Insert the needed SYNC */
+         if (jay_send_sfid(I) == BRW_SFID_MESSAGE_GATEWAY) {
+            b.cursor = jay_after_inst(I);
+            jay_SYNC(&b, TGL_SYNC_BAR);
+         }
       }
    }
 

src/intel/compiler/jay/jay_opcodes.py

@@ -129,6 +129,7 @@ op('shr', 2, 'u32 u64 u16 s16 s32 s64', Props.CMOD | Props.NEGATE0)
 
 op('quad_swizzle', 1, 'u1 u32', 0, ['enum jay_quad_swizzle swizzle'])
 op('sync', 0, None, Props.NO_DEST, ['enum tgl_sync_function op'])
+op('schedule_barrier', 0, None, Props.NO_DEST)
 
 for n in ['brd', 'illegal', 'goto', 'join', 'if', 'else',
           'endif', 'while', 'break', 'cont', 'call', 'calla', 'jmpi', 'ret',

src/intel/compiler/jay/jay_validate.c

@@ -250,6 +250,8 @@ validate_inst(struct validate_state *validate, jay_inst *I)
 
    if (I->op == JAY_OPCODE_SEL) {
       CHECK(jay_is_flag(I->src[2]) && "SEL src[2] (selector) must be a flag");
+   } else if (I->op == JAY_OPCODE_SYNC) {
+      CHECK(validate->post_ra && "SYNC does not exist while scheduling");
    }
 }