diff --git a/src/intel/compiler/jay/jay_lower_scoreboard.c b/src/intel/compiler/jay/jay_lower_scoreboard.c
index 252fdad5500..06244e17252 100644
--- a/src/intel/compiler/jay/jay_lower_scoreboard.c
+++ b/src/intel/compiler/jay/jay_lower_scoreboard.c
@@ -285,11 +285,16 @@ lower_regdist_local(jay_function *func, jay_block *block, u32_per_pipe *access)
       };
 
       /* Fold the immediate preceding SYNC.nop into this instruction, allowing
-       * us to wait on both ALU and a SEND in the same annotation.
+       * us to wait on both ALU and a SEND in the same annotation. We cannot do
+       * this safely in the presence of predication or SIMD splitting that could
+       * cause any part of the instruction to get shot down, skipping the sync
+       * for future instructions (at least not without more tricky logic).
        */
       if (last_sync &&
           jay_sync_op(last_sync) == TGL_SYNC_NOP &&
           I->dep.mode == TGL_SBID_NULL &&
+          !I->predication &&
+          !jay_simd_split(func->shader, I) &&
           (I->dep.regdist == 0 ||
            inferred_sync_pipe(func->shader->devinfo, I) == I->dep.pipe)) {
 
diff --git a/src/intel/compiler/jay/jay_to_binary.c b/src/intel/compiler/jay/jay_to_binary.c
index 935ae4d2727..01ce1395bc6 100644
--- a/src/intel/compiler/jay/jay_to_binary.c
+++ b/src/intel/compiler/jay/jay_to_binary.c
@@ -246,13 +246,15 @@ emit(struct brw_codegen *p,
    unsigned exec_size = jay_simd_width_physical(f->shader, I);
    // jay_print_inst(stdout, (jay_inst *) I);
 
-   /* Fix up SWSB dependencies for SIMD split instructions. The latter
-    * instructions do not need to redundantly wait on an SBID but might
-    * replicate their regdists.
-    */
+   /* Replicate the SWSB regdist for SIMD split instructions if needed */
    struct tgl_swsb dep =
       simd_offs && !I->replicate_dep ? tgl_swsb_null() : I->dep;
-   dep.mode = simd_offs ? TGL_SBID_NULL : dep.mode;
+
+   /* We do not allow SBID dependencies on SIMD split instructions since
+    * individual groups could get shot down. This would require more tracking
+    * and is unclear whether it's beneficial.
+    */
+   assert(simd_offs == 0 || I->dep.mode == TGL_SBID_NULL);
 
    if (I->decrement_dep) {
       unsigned delta = simd_offs * jay_macro_length(I);