i965: Rework gl_TessLevel*[] handling to use NIR compact arrays.

Treating everything as scalar arrays allows us to drop a bunch of
special case input/output munging all throughout the backend.
Instead, we just need to remap the TessLevel components to the
appropriate patch URB header locations in remap_patch_urb_offsets().

We also switch to treating the TES input versions of these as ordinary
shader inputs rather than system values, as remap_patch_urb_offsets()
just makes everything work out without special handling.

This regresses one Piglit test:
arb_tessellation_shader-large-uniforms/GL_TESS_CONTROL_SHADER-array-at-limit

The compiler starts promoting the constant arrays assigned to gl_TessLevel*
to uniform arrays.  Since the shader also has a uniform array that uses
the maximum number of uniform components, this puts it over the uniform
component limit enforced by the linker.  This is arguably a bug in the
constant array promotion code (it should avoid pushing us over limits),
but is unlikely to penalize any real application.

Signed-off-by: Kenneth Graunke <[email protected]>
Reviewed-by: Jason Ekstrand <[email protected]>

10 files changed, 92 insertions, 364 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_context.c b/src/mesa/drivers/dri/i965/brw_context.c
index f939463539c..84d34c0f6a0 100644
--- a/src/mesa/drivers/dri/i965/brw_context.c
+++ b/src/mesa/drivers/dri/i965/brw_context.c
@@ -673,7 +673,7 @@ brw_initialize_context_constants(struct brw_context *brw)
    if (brw->gen >= 5 || brw->is_g4x)
       ctx->Const.MaxClipPlanes = 8;
 
-   ctx->Const.LowerTessLevel = true;
+   ctx->Const.GLSLTessLevelsAsInputs = true;
    ctx->Const.LowerTCSPatchVerticesIn = brw->gen >= 8;
    ctx->Const.LowerTESPatchVerticesIn = true;
    ctx->Const.PrimitiveRestartForPatches = true;
diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
index 2ed843bd03d..8f745dff440 100644
--- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
@@ -2520,78 +2520,7 @@ fs_visitor::nir_emit_tcs_intrinsic(const fs_builder &bld,
          bld.MOV(patch_handle,
                  retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD));
 
-         if (imm_offset == 0) {
-            /* This is a read of gl_TessLevelInner[], which lives in the
-             * Patch URB header.  The layout depends on the domain.
-             */
-            dst.type = BRW_REGISTER_TYPE_F;
-            switch (tcs_key->tes_primitive_mode) {
-            case GL_QUADS: {
-               /* DWords 3-2 (reversed) */
-               fs_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_F, 4);
-
-               inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, tmp, patch_handle);
-               inst->offset = 0;
-               inst->mlen = 1;
-               inst->size_written = 4 * REG_SIZE;
-
-               /* dst.xy = tmp.wz */
-               bld.MOV(dst,                 offset(tmp, bld, 3));
-               bld.MOV(offset(dst, bld, 1), offset(tmp, bld, 2));
-               break;
-            }
-            case GL_TRIANGLES:
-               /* DWord 4; hardcode offset = 1 and size_written = REG_SIZE */
-               inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, dst, patch_handle);
-               inst->offset = 1;
-               inst->mlen = 1;
-               inst->size_written = REG_SIZE;
-               break;
-            case GL_ISOLINES:
-               /* All channels are undefined. */
-               break;
-            default:
-               unreachable("Bogus tessellation domain");
-            }
-         } else if (imm_offset == 1) {
-            /* This is a read of gl_TessLevelOuter[], which lives in the
-             * Patch URB header.  The layout depends on the domain.
-             */
-            dst.type = BRW_REGISTER_TYPE_F;
-
-            fs_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_F, 4);
-            inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, tmp, patch_handle);
-            inst->offset = 1;
-            inst->mlen = 1;
-            inst->size_written = 4 * REG_SIZE;
-
-            /* Reswizzle: WZYX */
-            fs_reg srcs[4] = {
-               offset(tmp, bld, 3),
-               offset(tmp, bld, 2),
-               offset(tmp, bld, 1),
-               offset(tmp, bld, 0),
-            };
-
-            unsigned num_components;
-            switch (tcs_key->tes_primitive_mode) {
-            case GL_QUADS:
-               num_components = 4;
-               break;
-            case GL_TRIANGLES:
-               num_components = 3;
-               break;
-            case GL_ISOLINES:
-               /* Isolines are not reversed; swizzle .zw -> .xy */
-               srcs[0] = offset(tmp, bld, 2);
-               srcs[1] = offset(tmp, bld, 3);
-               num_components = 2;
-               break;
-            default:
-               unreachable("Bogus tessellation domain");
-            }
-            bld.LOAD_PAYLOAD(dst, srcs, num_components, 0);
-         } else {
+         {
             if (first_component != 0) {
                unsigned read_components =
                   instr->num_components + first_component;
@@ -2656,55 +2585,6 @@ fs_visitor::nir_emit_tcs_intrinsic(const fs_builder &bld,
 
       if (indirect_offset.file != BAD_FILE) {
          srcs[header_regs++] = indirect_offset;
-      } else if (!is_passthrough_shader) {
-         if (imm_offset == 0) {
-            value.type = BRW_REGISTER_TYPE_F;
-
-            mask &= (1 << tesslevel_inner_components(tcs_key->tes_primitive_mode)) - 1;
-
-            /* This is a write to gl_TessLevelInner[], which lives in the
-             * Patch URB header.  The layout depends on the domain.
-             */
-            switch (tcs_key->tes_primitive_mode) {
-            case GL_QUADS:
-               /* gl_TessLevelInner[].xy lives at DWords 3-2 (reversed).
-                * We use an XXYX swizzle to reverse put .xy in the .wz
-                * channels, and use a .zw writemask.
-                */
-               mask = writemask_for_backwards_vector(mask);
-               swiz = BRW_SWIZZLE4(0, 0, 1, 0);
-               break;
-            case GL_TRIANGLES:
-               /* gl_TessLevelInner[].x lives at DWord 4, so we set the
-                * writemask to X and bump the URB offset by 1.
-                */
-               imm_offset = 1;
-               break;
-            case GL_ISOLINES:
-               /* Skip; gl_TessLevelInner[] doesn't exist for isolines. */
-               return;
-            default:
-               unreachable("Bogus tessellation domain");
-            }
-         } else if (imm_offset == 1) {
-            /* This is a write to gl_TessLevelOuter[] which lives in the
-             * Patch URB Header at DWords 4-7.  However, it's reversed, so
-             * instead of .xyzw we have .wzyx.
-             */
-            value.type = BRW_REGISTER_TYPE_F;
-
-            mask &= (1 << tesslevel_outer_components(tcs_key->tes_primitive_mode)) - 1;
-
-            if (tcs_key->tes_primitive_mode == GL_ISOLINES) {
-               /* Isolines .xy should be stored in .zw, in order. */
-               swiz = BRW_SWIZZLE4(0, 0, 0, 1);
-               mask <<= 2;
-            } else {
-               /* Other domains are reversed; store .wzyx instead of .xyzw */
-               swiz = BRW_SWIZZLE_WZYX;
-               mask = writemask_for_backwards_vector(mask);
-            }
-         }
       }
 
       if (mask == 0)
@@ -2851,48 +2731,6 @@ fs_visitor::nir_emit_tes_intrinsic(const fs_builder &bld,
       }
       break;
 
-   case nir_intrinsic_load_tess_level_outer:
-      /* When the TES reads gl_TessLevelOuter, we ensure that the patch header
-       * appears as a push-model input.  So, we can simply use the ATTR file
-       * rather than issuing URB read messages.  The data is stored in the
-       * high DWords in reverse order - DWord 7 contains .x, DWord 6 contains
-       * .y, and so on.
-       */
-      switch (tes_prog_data->domain) {
-      case BRW_TESS_DOMAIN_QUAD:
-         for (unsigned i = 0; i < 4; i++)
-            bld.MOV(offset(dest, bld, i), component(fs_reg(ATTR, 0), 7 - i));
-         break;
-      case BRW_TESS_DOMAIN_TRI:
-         for (unsigned i = 0; i < 3; i++)
-            bld.MOV(offset(dest, bld, i), component(fs_reg(ATTR, 0), 7 - i));
-         break;
-      case BRW_TESS_DOMAIN_ISOLINE:
-         for (unsigned i = 0; i < 2; i++)
-            bld.MOV(offset(dest, bld, i), component(fs_reg(ATTR, 0), 6 + i));
-         break;
-      }
-      break;
-
-   case nir_intrinsic_load_tess_level_inner:
-      /* When the TES reads gl_TessLevelInner, we ensure that the patch header
-       * appears as a push-model input.  So, we can simply use the ATTR file
-       * rather than issuing URB read messages.
-       */
-      switch (tes_prog_data->domain) {
-      case BRW_TESS_DOMAIN_QUAD:
-         bld.MOV(dest, component(fs_reg(ATTR, 0), 3));
-         bld.MOV(offset(dest, bld, 1), component(fs_reg(ATTR, 0), 2));
-         break;
-      case BRW_TESS_DOMAIN_TRI:
-         bld.MOV(dest, component(fs_reg(ATTR, 0), 4));
-         break;
-      case BRW_TESS_DOMAIN_ISOLINE:
-         /* ignore - value is undefined */
-         break;
-      }
-      break;
-
    case nir_intrinsic_load_input:
    case nir_intrinsic_load_per_vertex_input: {
       fs_reg indirect_offset = get_indirect_offset(instr);
diff --git a/src/mesa/drivers/dri/i965/brw_nir.c b/src/mesa/drivers/dri/i965/brw_nir.c
index 6f37e97a86f..46eeb1723b4 100644
--- a/src/mesa/drivers/dri/i965/brw_nir.c
+++ b/src/mesa/drivers/dri/i965/brw_nir.c
@@ -141,9 +141,68 @@ remap_inputs_with_vue_map(nir_block *block, const struct brw_vue_map *vue_map)
 }
 
 static bool
+remap_tess_levels(nir_builder *b, nir_intrinsic_instr *intr,
+                  GLenum primitive_mode)
+{
+   const int location = nir_intrinsic_base(intr);
+   const unsigned component = nir_intrinsic_component(intr);
+   bool out_of_bounds;
+
+   if (location == VARYING_SLOT_TESS_LEVEL_INNER) {
+      switch (primitive_mode) {
+      case GL_QUADS:
+         /* gl_TessLevelInner[0..1] lives at DWords 3-2 (reversed). */
+         nir_intrinsic_set_base(intr, 0);
+         nir_intrinsic_set_component(intr, 3 - component);
+         out_of_bounds = false;
+         break;
+      case GL_TRIANGLES:
+         /* gl_TessLevelInner[0] lives at DWord 4. */
+         nir_intrinsic_set_base(intr, 1);
+         out_of_bounds = component > 0;
+         break;
+      case GL_ISOLINES:
+         out_of_bounds = true;
+         break;
+      default:
+         unreachable("Bogus tessellation domain");
+      }
+   } else if (location == VARYING_SLOT_TESS_LEVEL_OUTER) {
+      if (primitive_mode == GL_ISOLINES) {
+         /* gl_TessLevelOuter[0..1] lives at DWords 6-7 (in order). */
+         nir_intrinsic_set_base(intr, 1);
+         nir_intrinsic_set_component(intr, 2 + nir_intrinsic_component(intr));
+         out_of_bounds = component > 1;
+      } else {
+         /* Triangles use DWords 7-5 (reversed); Quads use 7-4 (reversed) */
+         nir_intrinsic_set_base(intr, 1);
+         nir_intrinsic_set_component(intr, 3 - nir_intrinsic_component(intr));
+         out_of_bounds = component == 3 && primitive_mode == GL_TRIANGLES;
+      }
+   } else {
+      return false;
+   }
+
+   if (out_of_bounds) {
+      if (nir_intrinsic_infos[intr->intrinsic].has_dest) {
+         b->cursor = nir_before_instr(&intr->instr);
+         nir_ssa_def *undef = nir_ssa_undef(b, 1, 32);
+         nir_ssa_def_rewrite_uses(&intr->dest.ssa, nir_src_for_ssa(undef));
+      }
+      nir_instr_remove(&intr->instr);
+   }
+
+   return true;
+}
+
+static bool
 remap_patch_urb_offsets(nir_block *block, nir_builder *b,
-                        const struct brw_vue_map *vue_map)
+                        const struct brw_vue_map *vue_map,
+                        GLenum tes_primitive_mode)
 {
+   const bool is_passthrough_tcs = b->shader->info->name &&
+      strcmp(b->shader->info->name, "passthrough") == 0;
+
    nir_foreach_instr_safe(instr, block) {
       if (instr->type != nir_instr_type_intrinsic)
          continue;
@@ -154,6 +213,11 @@ remap_patch_urb_offsets(nir_block *block, nir_builder *b,
 
       if ((stage == MESA_SHADER_TESS_CTRL && is_output(intrin)) ||
           (stage == MESA_SHADER_TESS_EVAL && is_input(intrin))) {
+
+         if (!is_passthrough_tcs &&
+             remap_tess_levels(b, intrin, tes_primitive_mode))
+            continue;
+
          int vue_slot = vue_map->varying_to_slot[intrin->const_index[0]];
          assert(vue_slot != -1);
          intrin->const_index[0] = vue_slot;
@@ -273,7 +337,8 @@ brw_nir_lower_tes_inputs(nir_shader *nir, const struct brw_vue_map *vue_map)
          nir_builder b;
          nir_builder_init(&b, function->impl);
          nir_foreach_block(block, function->impl) {
-            remap_patch_urb_offsets(block, &b, vue_map);
+            remap_patch_urb_offsets(block, &b, vue_map,
+                                    nir->info->tes.primitive_mode);
          }
       }
    }
@@ -341,7 +406,8 @@ brw_nir_lower_vue_outputs(nir_shader *nir,
 }
 
 void
-brw_nir_lower_tcs_outputs(nir_shader *nir, const struct brw_vue_map *vue_map)
+brw_nir_lower_tcs_outputs(nir_shader *nir, const struct brw_vue_map *vue_map,
+                          GLenum tes_primitive_mode)
 {
    nir_foreach_variable(var, &nir->outputs) {
       var->data.driver_location = var->data.location;
@@ -359,7 +425,7 @@ brw_nir_lower_tcs_outputs(nir_shader *nir, const struct brw_vue_map *vue_map)
          nir_builder b;
          nir_builder_init(&b, function->impl);
          nir_foreach_block(block, function->impl) {
-            remap_patch_urb_offsets(block, &b, vue_map);
+            remap_patch_urb_offsets(block, &b, vue_map, tes_primitive_mode);
          }
       }
    }
diff --git a/src/mesa/drivers/dri/i965/brw_nir.h b/src/mesa/drivers/dri/i965/brw_nir.h
index a7189c12165..f713d47b40e 100644
--- a/src/mesa/drivers/dri/i965/brw_nir.h
+++ b/src/mesa/drivers/dri/i965/brw_nir.h
@@ -109,7 +109,8 @@ void brw_nir_lower_fs_inputs(nir_shader *nir, struct brw_vue_map *vue_map,
                              const struct gen_device_info *devinfo,
                              const struct brw_wm_prog_key *key);
 void brw_nir_lower_vue_outputs(nir_shader *nir, bool is_scalar);
-void brw_nir_lower_tcs_outputs(nir_shader *nir, const struct brw_vue_map *vue);
+void brw_nir_lower_tcs_outputs(nir_shader *nir, const struct brw_vue_map *vue,
+                               GLenum tes_primitive_mode);
 void brw_nir_lower_fs_outputs(nir_shader *nir);
 void brw_nir_lower_cs_shared(nir_shader *nir);
 
diff --git a/src/mesa/drivers/dri/i965/brw_nir_tcs_workarounds.c b/src/mesa/drivers/dri/i965/brw_nir_tcs_workarounds.c
index 154be3b1125..a85f493c704 100644
--- a/src/mesa/drivers/dri/i965/brw_nir_tcs_workarounds.c
+++ b/src/mesa/drivers/dri/i965/brw_nir_tcs_workarounds.c
@@ -73,7 +73,7 @@
  */
 
 static inline nir_ssa_def *
-load_output(nir_builder *b, int num_components, int offset)
+load_output(nir_builder *b, int num_components, int offset, int component)
 {
    nir_intrinsic_instr *load =
       nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_output);
@@ -81,6 +81,7 @@ load_output(nir_builder *b, int num_components, int offset)
    load->num_components = num_components;
    load->src[0] = nir_src_for_ssa(nir_imm_int(b, 0));
    nir_intrinsic_set_base(load, offset);
+   nir_intrinsic_set_component(load, component);
 
    nir_builder_instr_insert(b, &load->instr);
 
@@ -92,8 +93,8 @@ emit_quads_workaround(nir_builder *b, nir_block *block)
 {
    b->cursor = nir_after_block_before_jump(block);
 
-   nir_ssa_def *inner = load_output(b, 2, 0);
-   nir_ssa_def *outer = load_output(b, 4, 1);
+   nir_ssa_def *inner = load_output(b, 2, 0, 2);
+   nir_ssa_def *outer = load_output(b, 4, 1, 0);
 
    nir_ssa_def *any_greater_than_1 =
        nir_ior(b, nir_bany(b, nir_flt(b, nir_imm_float(b, 1.0f), outer)),
@@ -113,6 +114,7 @@ emit_quads_workaround(nir_builder *b, nir_block *block)
       nir_intrinsic_instr_create(b->shader, nir_intrinsic_store_output);
    store->num_components = 2;
    nir_intrinsic_set_write_mask(store, WRITEMASK_XY);
+   nir_intrinsic_set_component(store, 2);
    store->src[0] = nir_src_for_ssa(inner);
    store->src[1] = nir_src_for_ssa(nir_imm_int(b, 0));
    nir_builder_instr_insert(b, &store->instr);
diff --git a/src/mesa/drivers/dri/i965/brw_shader.cpp b/src/mesa/drivers/dri/i965/brw_shader.cpp
index af8a4e501f7..6c4f5f8de5e 100644
--- a/src/mesa/drivers/dri/i965/brw_shader.cpp
+++ b/src/mesa/drivers/dri/i965/brw_shader.cpp
@@ -648,53 +648,6 @@ get_atomic_counter_op(nir_intrinsic_op op)
    }
 }
 
-unsigned
-tesslevel_outer_components(GLenum tes_primitive_mode)
-{
-   switch (tes_primitive_mode) {
-   case GL_QUADS:
-      return 4;
-   case GL_TRIANGLES:
-      return 3;
-   case GL_ISOLINES:
-      return 2;
-   default:
-      unreachable("Bogus tessellation domain");
-   }
-   return 0;
-}
-
-unsigned
-tesslevel_inner_components(GLenum tes_primitive_mode)
-{
-   switch (tes_primitive_mode) {
-   case GL_QUADS:
-      return 2;
-   case GL_TRIANGLES:
-      return 1;
-   case GL_ISOLINES:
-      return 0;
-   default:
-      unreachable("Bogus tessellation domain");
-   }
-   return 0;
-}
-
-/**
- * Given a normal .xyzw writemask, convert it to a writemask for a vector
- * that's stored backwards, i.e. .wzyx.
- */
-unsigned
-writemask_for_backwards_vector(unsigned mask)
-{
-   unsigned new_mask = 0;
-
-   for (int i = 0; i < 4; i++)
-      new_mask |= ((mask >> i) & 1) << (3 - i);
-
-   return new_mask;
-}
-
 backend_shader::backend_shader(const struct brw_compiler *compiler,
                                void *log_data,
                                void *mem_ctx,
@@ -712,8 +665,6 @@ backend_shader::backend_shader(const struct brw_compiler *compiler,
    debug_enabled = INTEL_DEBUG & intel_debug_flag_for_shader_stage(stage);
    stage_name = _mesa_shader_stage_to_string(stage);
    stage_abbrev = _mesa_shader_stage_to_abbrev(stage);
-   is_passthrough_shader =
-      nir->info->name && strcmp(nir->info->name, "passthrough") == 0;
 }
 
 bool
@@ -1394,17 +1345,7 @@ brw_compile_tes(const struct brw_compiler *compiler,
 
    /* URB entry sizes are stored as a multiple of 64 bytes. */
    prog_data->base.urb_entry_size = ALIGN(output_size_bytes, 64) / 64;
-
-   bool need_patch_header = nir->info->system_values_read &
-      (BITFIELD64_BIT(SYSTEM_VALUE_TESS_LEVEL_OUTER) |
-       BITFIELD64_BIT(SYSTEM_VALUE_TESS_LEVEL_INNER));
-
-   /* The TES will pull most inputs using URB read messages.
-    *
-    * However, we push the patch header for TessLevel factors when required,
-    * as it's a tiny amount of extra data.
-    */
-   prog_data->base.urb_read_length = need_patch_header ? 1 : 0;
+   prog_data->base.urb_read_length = 0;
 
    if (unlikely(INTEL_DEBUG & DEBUG_TES)) {
       fprintf(stderr, "TES Input ");
diff --git a/src/mesa/drivers/dri/i965/brw_shader.h b/src/mesa/drivers/dri/i965/brw_shader.h
index 9366d91c74a..60f498f0ebb 100644
--- a/src/mesa/drivers/dri/i965/brw_shader.h
+++ b/src/mesa/drivers/dri/i965/brw_shader.h
@@ -217,7 +217,6 @@ public:
    bool debug_enabled;
    const char *stage_name;
    const char *stage_abbrev;
-   bool is_passthrough_shader;
 
    brw::simple_allocator alloc;
 
@@ -291,10 +290,6 @@ bool brw_cs_precompile(struct gl_context *ctx, struct gl_program *prog);
 GLboolean brw_link_shader(struct gl_context *ctx, struct gl_shader_program *prog);
 struct gl_linked_shader *brw_new_shader(gl_shader_stage stage);
 
-unsigned tesslevel_outer_components(GLenum tes_primitive_mode);
-unsigned tesslevel_inner_components(GLenum tes_primitive_mode);
-unsigned writemask_for_backwards_vector(unsigned mask);
-
 unsigned get_atomic_counter_op(nir_intrinsic_op op);
 
 #ifdef __cplusplus
diff --git a/src/mesa/drivers/dri/i965/brw_tcs.c b/src/mesa/drivers/dri/i965/brw_tcs.c
index 70571774a1e..b6251d52f60 100644
--- a/src/mesa/drivers/dri/i965/brw_tcs.c
+++ b/src/mesa/drivers/dri/i965/brw_tcs.c
@@ -51,7 +51,8 @@ create_passthrough_tcs(void *mem_ctx, const struct brw_compiler *compiler,
    nir_ssa_def *invoc_id =
       nir_load_system_value(&b, nir_intrinsic_load_invocation_id, 0);
 
-   nir->info->inputs_read = key->outputs_written;
+   nir->info->inputs_read = key->outputs_written &
+      ~(VARYING_BIT_TESS_LEVEL_INNER | VARYING_BIT_TESS_LEVEL_OUTER);
    nir->info->outputs_written = key->outputs_written;
    nir->info->tcs.vertices_out = key->input_vertices;
    nir->info->name = ralloc_strdup(nir, "passthrough");
@@ -81,7 +82,7 @@ create_passthrough_tcs(void *mem_ctx, const struct brw_compiler *compiler,
    }
 
    /* Copy inputs to outputs. */
-   uint64_t varyings = key->outputs_written;
+   uint64_t varyings = nir->info->inputs_read;
 
    while (varyings != 0) {
       const int varying = ffsll(varyings) - 1;
diff --git a/src/mesa/drivers/dri/i965/brw_tes.c b/src/mesa/drivers/dri/i965/brw_tes.c
index 124df66ef9b..06862df8346 100644
--- a/src/mesa/drivers/dri/i965/brw_tes.c
+++ b/src/mesa/drivers/dri/i965/brw_tes.c
@@ -235,16 +235,12 @@ brw_tes_populate_key(struct brw_context *brw,
     */
    if (tcp) {
       struct gl_program *tcp_prog = &tcp->program;
-      per_vertex_slots |= tcp_prog->info.outputs_written;
+      per_vertex_slots |= tcp_prog->info.outputs_written &
+         ~(VARYING_BIT_TESS_LEVEL_INNER | VARYING_BIT_TESS_LEVEL_OUTER);
       per_patch_slots |= tcp_prog->info.patch_outputs_written;
    }
 
-   /* Ignore gl_TessLevelInner/Outer - we treat them as system values,
-    * not inputs, and they're always present in the URB entry regardless
-    * of whether or not we read them.
-    */
-   key->inputs_read = per_vertex_slots &
-      ~(VARYING_BIT_TESS_LEVEL_INNER | VARYING_BIT_TESS_LEVEL_OUTER);
+   key->inputs_read = per_vertex_slots;
    key->patch_inputs_read = per_patch_slots;
 
    /* _NEW_TEXTURE */
@@ -299,14 +295,11 @@ brw_tes_precompile(struct gl_context *ctx,
    if (shader_prog->_LinkedShaders[MESA_SHADER_TESS_CTRL]) {
       struct gl_program *tcp =
          shader_prog->_LinkedShaders[MESA_SHADER_TESS_CTRL]->Program;
-      key.inputs_read |= tcp->nir->info->outputs_written;
+      key.inputs_read |= tcp->nir->info->outputs_written &
+         ~(VARYING_BIT_TESS_LEVEL_INNER | VARYING_BIT_TESS_LEVEL_OUTER);
       key.patch_inputs_read |= tcp->nir->info->patch_outputs_written;
    }
 
-   /* Ignore gl_TessLevelInner/Outer - they're system values. */
-   key.inputs_read &= ~(VARYING_BIT_TESS_LEVEL_INNER |
-                        VARYING_BIT_TESS_LEVEL_OUTER);
-
    brw_setup_tex_for_precompile(brw, &key.tex, prog);
 
    success = brw_codegen_tes_prog(brw, btep, &key);
diff --git a/src/mesa/drivers/dri/i965/brw_vec4_tcs.cpp b/src/mesa/drivers/dri/i965/brw_vec4_tcs.cpp
index bfa22458f0e..9ef3dc04665 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4_tcs.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4_tcs.cpp
@@ -319,62 +319,8 @@ vec4_tcs_visitor::nir_emit_intrinsic(nir_intrinsic_instr *instr)
       dst_reg dst = get_nir_dest(instr->dest, BRW_REGISTER_TYPE_D);
       dst.writemask = brw_writemask_for_size(instr->num_components);
 
-      if (imm_offset == 0 && indirect_offset.file == BAD_FILE) {
-         dst.type = BRW_REGISTER_TYPE_F;
-
-         /* This is a read of gl_TessLevelInner[], which lives in the
-          * Patch URB header.  The layout depends on the domain.
-          */
-         switch (key->tes_primitive_mode) {
-         case GL_QUADS: {
-            /* DWords 3-2 (reversed); use offset 0 and WZYX swizzle. */
-            dst_reg tmp(this, glsl_type::vec4_type);
-            emit_output_urb_read(tmp, 0, 0, src_reg());
-            emit(MOV(writemask(dst, WRITEMASK_XY),
-                     swizzle(src_reg(tmp), BRW_SWIZZLE_WZYX)));
-            break;
-         }
-         case GL_TRIANGLES:
-            /* DWord 4; use offset 1 but normal swizzle/writemask. */
-            emit_output_urb_read(writemask(dst, WRITEMASK_X), 1, 0,
-                                 src_reg());
-            break;
-         case GL_ISOLINES:
-            /* All channels are undefined. */
-            return;
-         default:
-            unreachable("Bogus tessellation domain");
-         }
-      } else if (imm_offset == 1 && indirect_offset.file == BAD_FILE) {
-         dst.type = BRW_REGISTER_TYPE_F;
-         unsigned swiz = BRW_SWIZZLE_WZYX;
-
-         /* This is a read of gl_TessLevelOuter[], which lives in the
-          * high 4 DWords of the Patch URB header, in reverse order.
-          */
-         switch (key->tes_primitive_mode) {
-         case GL_QUADS:
-            dst.writemask = WRITEMASK_XYZW;
-            break;
-         case GL_TRIANGLES:
-            dst.writemask = WRITEMASK_XYZ;
-            break;
-         case GL_ISOLINES:
-            /* Isolines are not reversed; swizzle .zw -> .xy */
-            swiz = BRW_SWIZZLE_ZWZW;
-            dst.writemask = WRITEMASK_XY;
-            return;
-         default:
-            unreachable("Bogus tessellation domain");
-         }
-
-         dst_reg tmp(this, glsl_type::vec4_type);
-         emit_output_urb_read(tmp, 1, 0, src_reg());
-         emit(MOV(dst, swizzle(src_reg(tmp), swiz)));
-      } else {
-         emit_output_urb_read(dst, imm_offset, nir_intrinsic_component(instr),
-                              indirect_offset);
-      }
+      emit_output_urb_read(dst, imm_offset, nir_intrinsic_component(instr),
+                           indirect_offset);
       break;
    }
    case nir_intrinsic_store_output:
@@ -386,62 +332,6 @@ vec4_tcs_visitor::nir_emit_intrinsic(nir_intrinsic_instr *instr)
       src_reg indirect_offset = get_indirect_offset(instr);
       unsigned imm_offset = instr->const_index[0];
 
-      /* The passthrough shader writes the whole patch header as two vec4s;
-       * skip all the gl_TessLevelInner/Outer swizzling.
-       */
-      if (indirect_offset.file == BAD_FILE && !is_passthrough_shader) {
-         if (imm_offset == 0) {
-            value.type = BRW_REGISTER_TYPE_F;
-
-            mask &=
-               (1 << tesslevel_inner_components(key->tes_primitive_mode)) - 1;
-
-            /* This is a write to gl_TessLevelInner[], which lives in the
-             * Patch URB header.  The layout depends on the domain.
-             */
-            switch (key->tes_primitive_mode) {
-            case GL_QUADS:
-               /* gl_TessLevelInner[].xy lives at DWords 3-2 (reversed).
-                * We use an XXYX swizzle to reverse put .xy in the .wz
-                * channels, and use a .zw writemask.
-                */
-               swiz = BRW_SWIZZLE4(0, 0, 1, 0);
-               mask = writemask_for_backwards_vector(mask);
-               break;
-            case GL_TRIANGLES:
-               /* gl_TessLevelInner[].x lives at DWord 4, so we set the
-                * writemask to X and bump the URB offset by 1.
-                */
-               imm_offset = 1;
-               break;
-            case GL_ISOLINES:
-               /* Skip; gl_TessLevelInner[] doesn't exist for isolines. */
-               return;
-            default:
-               unreachable("Bogus tessellation domain");
-            }
-         } else if (imm_offset == 1) {
-            value.type = BRW_REGISTER_TYPE_F;
-
-            mask &=
-               (1 << tesslevel_outer_components(key->tes_primitive_mode)) - 1;
-
-            /* This is a write to gl_TessLevelOuter[] which lives in the
-             * Patch URB Header at DWords 4-7.  However, it's reversed, so
-             * instead of .xyzw we have .wzyx.
-             */
-            if (key->tes_primitive_mode == GL_ISOLINES) {
-               /* Isolines .xy should be stored in .zw, in order. */
-               swiz = BRW_SWIZZLE4(0, 0, 0, 1);
-               mask <<= 2;
-            } else {
-               /* Other domains are reversed; store .wzyx instead of .xyzw. */
-               swiz = BRW_SWIZZLE_WZYX;
-               mask = writemask_for_backwards_vector(mask);
-            }
-         }
-      }
-
       unsigned first_component = nir_intrinsic_component(instr);
       if (first_component) {
          if (nir_src_bit_size(instr->src[0]) == 64)
@@ -522,7 +412,8 @@ brw_compile_tcs(const struct brw_compiler *compiler,
 
    nir = brw_nir_apply_sampler_key(nir, compiler, &key->tex, is_scalar);
    brw_nir_lower_vue_inputs(nir, is_scalar, &input_vue_map);
-   brw_nir_lower_tcs_outputs(nir, &vue_prog_data->vue_map);
+   brw_nir_lower_tcs_outputs(nir, &vue_prog_data->vue_map,
+                             key->tes_primitive_mode);
    if (key->quads_workaround)
       brw_nir_apply_tcs_quads_workaround(nir);