diff --git a/src/intel/compiler/jay/jay_builder.h b/src/intel/compiler/jay/jay_builder.h
index da22cff6c55..5b3f6ee582d 100644
--- a/src/intel/compiler/jay/jay_builder.h
+++ b/src/intel/compiler/jay/jay_builder.h
@@ -462,6 +462,7 @@ struct jayb_send_params {
    enum jay_type src_type[2];
    unsigned nr_srcs;
    uint32_t ex_desc_imm;
+   int split; /**< explicit split point */
    bool eot;
    bool check_tdr;
    bool uniform;
@@ -538,15 +539,9 @@ _jay_SEND(jay_builder *b, const struct jayb_send_params p)
       I->src[2] = p.nr_srcs > 0 ? p.srcs[0] : jay_null();
       I->src[3] = p.nr_srcs > 1 ? p.srcs[1] : jay_null();
    } else {
-      /* Otherwise, we need to pick a point to split at.
-       *
-       * Heuristic: don't split render targer writes becuase RA gets confused
-       * with the EOT requirements. Split everything else in half.
-       *
-       * TODO: Come up with a better heuristic.
-       */
+      /* Otherwise, we need to pick a point to split at. */
       assert(info->type_0 == info->type_1);
-      unsigned split = !p.check_tdr ? (p.nr_srcs / 2) : p.nr_srcs;
+      unsigned split = p.split > 0 ? p.split : p.nr_srcs / 2;
       I->src[2] = jay_collect_vectors(b, &p.srcs[0], split);
       I->src[3] = jay_collect_vectors(b, &p.srcs[split], p.nr_srcs - split);
    }
diff --git a/src/intel/compiler/jay/jay_from_nir.c b/src/intel/compiler/jay/jay_from_nir.c
index c88c4b6918b..ab233b769a6 100644
--- a/src/intel/compiler/jay/jay_from_nir.c
+++ b/src/intel/compiler/jay/jay_from_nir.c
@@ -804,6 +804,7 @@ jay_emit_fb_write(jay_builder *b, nir_intrinsic_instr *intr)
    jay_def srcs[4 + 16 + 4 + 1 + 16];
 
    unsigned len = 0;
+   int split = -1;
 
    if (!jay_is_null(src0_alpha))
       srcs[len++] = jay_as_gpr(b, src0_alpha);
@@ -816,7 +817,16 @@ jay_emit_fb_write(jay_builder *b, nir_intrinsic_instr *intr)
    if (!jay_is_null(depth))
       srcs[len++] = jay_as_gpr(b, depth);
 
-   assert(jay_is_null(stencil) && "TODO: stencil");
+   if (!jay_is_null(stencil)) {
+      jay_def packed = jay_alloc_def(b, UGPR, jay_ugpr_per_grf(b->shader));
+      jay_BYTE_PACK(b, packed, jay_as_gpr(b, stencil));
+
+      /* Split send before stencil */
+      split = len;
+
+      for (unsigned i = 0; i < jay_num_values(packed); i++)
+         srcs[len++] = jay_extract(packed, i);
+   }
 
    /* Optimize out unconditional discards (probably should do this in NIR) */
    if (nir_src_is_const(intr->src[5]) && nir_src_as_bool(intr->src[5])) {
@@ -824,10 +834,15 @@ jay_emit_fb_write(jay_builder *b, nir_intrinsic_instr *intr)
          srcs[i] = jay_UNDEF_u32(b);
    }
 
+   /* Our current send splitting heuristic is bad, override it. */
+   if (split == -1) {
+      split = len;
+   }
+
    jay_inst *send =
       jay_SEND(b, .sfid = BRW_SFID_RENDER_CACHE, .check_tdr = true,
                .msg_desc = desc | (ex_desc << 32), .srcs = srcs, .nr_srcs = len,
-               .type = JAY_TYPE_U32, .eot = last);
+               .type = JAY_TYPE_U32, .eot = last, .split = split);
 
    /* Handle the disable predicate. It is logically inverted. */
    if (!nir_src_is_const(intr->src[5]) || nir_src_as_bool(intr->src[5])) {
diff --git a/src/intel/compiler/jay/jay_nir.c b/src/intel/compiler/jay/jay_nir.c
index dfaffb04a25..e3d4b588616 100644
--- a/src/intel/compiler/jay/jay_nir.c
+++ b/src/intel/compiler/jay/jay_nir.c
@@ -151,7 +151,6 @@ collect_fragment_output(nir_builder *b, nir_intrinsic_instr *intr, void *ctx_)
    } else if (loc == FRAG_RESULT_DEPTH) {
       out = &ctx->depth;
    } else if (loc == FRAG_RESULT_STENCIL) {
-      UNREACHABLE("todo");
       out = &ctx->stencil;
    } else if (loc == FRAG_RESULT_SAMPLE_MASK) {
       UNREACHABLE("todo");
diff --git a/src/intel/compiler/jay/jay_opcodes.py b/src/intel/compiler/jay/jay_opcodes.py
index 80bacd8442b..548557ae3c5 100644
--- a/src/intel/compiler/jay/jay_opcodes.py
+++ b/src/intel/compiler/jay/jay_opcodes.py
@@ -166,6 +166,12 @@ op('extract_byte_per_8lanes', 2, 'u32')
 op('shr_odd_subspans_by_4', 1, 'u16')
 op('and_u32_u16', 2, 'u32')
 
+# Copy the first byte of each lane, treating the destination as if it were
+# effectively JAY_STRIDE_1 (which doesn't exist).  Because the destination
+# doesn't follow proper lane alignments, this should not write to GPRs.
+# This is used for stencil outputs in render target write messages.
+op('byte_pack', 1, 'u32')
+
 # Pixel coord calculations. expand_quad replicates out the per-2x2 values from
 # its source g0.[10...13] and - in the case of SIMD32 - g1.[10...13] into a
 # per-lane value. Then offset_packed_pixel_coords adds the appropriate packed
diff --git a/src/intel/compiler/jay/jay_opt_propagate.c b/src/intel/compiler/jay/jay_opt_propagate.c
index 3718e4b83db..cf37c8369f6 100644
--- a/src/intel/compiler/jay/jay_opt_propagate.c
+++ b/src/intel/compiler/jay/jay_opt_propagate.c
@@ -153,7 +153,9 @@ propagate_forwards(jay_function *f)
       }
 
       /* Don't propagate into phis yet - TODO: File awareness */
-      if (I->op == JAY_OPCODE_PHI_SRC || I->op == JAY_OPCODE_SEND)
+      if (I->op == JAY_OPCODE_PHI_SRC ||
+          I->op == JAY_OPCODE_SEND ||
+          I->op == JAY_OPCODE_BYTE_PACK)
          continue;
 
       jay_foreach_ssa_src(I, s) {
diff --git a/src/intel/compiler/jay/jay_to_binary.c b/src/intel/compiler/jay/jay_to_binary.c
index 2aa6408b033..6fdfef4780c 100644
--- a/src/intel/compiler/jay/jay_to_binary.c
+++ b/src/intel/compiler/jay/jay_to_binary.c
@@ -488,6 +488,11 @@ emit(struct brw_codegen *p,
       brw_MOV(p, dst, stride(retype(SRC(simd_offs), BRW_TYPE_UB), 1, 8, 0));
       break;
 
+   case JAY_OPCODE_BYTE_PACK:
+      brw_MOV(p, stride(retype(dst, BRW_TYPE_UB), 1, 1, 0),
+              stride(retype(SRC(0), BRW_TYPE_UB), 4, 1, 0));
+      break;
+
    case JAY_OPCODE_SHR_ODD_SUBSPANS_BY_4:
       brw_SHR(p, dst, SRC(0), brw_imm_uv(0x44440000));
       break;