nir: remove dead glsl variables before lowering io.
[mesa.git] / src / mesa / drivers / dri / i965 / brw_nir.c
index 83921891d1c72c5bd86cad0f8ffb637d622100ee..9274f2e1616d6d240a8b327fed6907674a7bcbed 100644 (file)
 
 #include "brw_nir.h"
 #include "brw_shader.h"
+#include "compiler/glsl_types.h"
 #include "compiler/nir/glsl_to_nir.h"
 #include "compiler/nir/nir_builder.h"
 #include "program/prog_to_nir.h"
+#include "program/prog_parameter.h"
 
 static bool
 is_input(nir_intrinsic_instr *intrin)
@@ -54,25 +56,19 @@ is_output(nir_intrinsic_instr *intrin)
  * we don't know what part of a compound variable is accessed, we allocate
  * storage for the entire thing.
  */
-struct add_const_offset_to_base_params {
-   nir_builder b;
-   nir_variable_mode mode;
-};
 
 static bool
-add_const_offset_to_base_block(nir_block *block, void *closure)
+add_const_offset_to_base_block(nir_block *block, nir_builder *b,
+                               nir_variable_mode mode)
 {
-   struct add_const_offset_to_base_params *params = closure;
-   nir_builder *b = &params->b;
-
-   nir_foreach_instr_safe(block, instr) {
+   nir_foreach_instr_safe(instr, block) {
       if (instr->type != nir_instr_type_intrinsic)
          continue;
 
       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 
-      if ((params->mode == nir_var_shader_in && is_input(intrin)) ||
-          (params->mode == nir_var_shader_out && is_output(intrin))) {
+      if ((mode == nir_var_shader_in && is_input(intrin)) ||
+          (mode == nir_var_shader_out && is_output(intrin))) {
          nir_src *offset = nir_get_io_offset_src(intrin);
          nir_const_value *const_offset = nir_src_as_const_value(*offset);
 
@@ -90,22 +86,21 @@ add_const_offset_to_base_block(nir_block *block, void *closure)
 static void
 add_const_offset_to_base(nir_shader *nir, nir_variable_mode mode)
 {
-   struct add_const_offset_to_base_params params = { .mode = mode };
-
-   nir_foreach_function(nir, f) {
+   nir_foreach_function(f, nir) {
       if (f->impl) {
-         nir_builder_init(&params.b, f->impl);
-         nir_foreach_block(f->impl, add_const_offset_to_base_block, &params);
+         nir_builder b;
+         nir_builder_init(&b, f->impl);
+         nir_foreach_block(block, f->impl) {
+            add_const_offset_to_base_block(block, &b, mode);
+         }
       }
    }
 }
 
 static bool
-remap_vs_attrs(nir_block *block, void *closure)
+remap_vs_attrs(nir_block *block, struct nir_shader_info *nir_info)
 {
-   GLbitfield64 inputs_read = *((GLbitfield64 *) closure);
-
-   nir_foreach_instr(block, instr) {
+   nir_foreach_instr(instr, block) {
       if (instr->type != nir_instr_type_intrinsic)
          continue;
 
@@ -118,20 +113,20 @@ remap_vs_attrs(nir_block *block, void *closure)
           * before it and counting the bits.
           */
          int attr = intrin->const_index[0];
-         int slot = _mesa_bitcount_64(inputs_read & BITFIELD64_MASK(attr));
-
-         intrin->const_index[0] = 4 * slot;
+         int slot = _mesa_bitcount_64(nir_info->inputs_read &
+                                      BITFIELD64_MASK(attr));
+         int dslot = _mesa_bitcount_64(nir_info->double_inputs_read &
+                                       BITFIELD64_MASK(attr));
+         intrin->const_index[0] = 4 * (slot + dslot);
       }
    }
    return true;
 }
 
 static bool
-remap_inputs_with_vue_map(nir_block *block, void *closure)
+remap_inputs_with_vue_map(nir_block *block, const struct brw_vue_map *vue_map)
 {
-   const struct brw_vue_map *vue_map = closure;
-
-   nir_foreach_instr(block, instr) {
+   nir_foreach_instr(instr, block) {
       if (instr->type != nir_instr_type_intrinsic)
          continue;
 
@@ -147,27 +142,21 @@ remap_inputs_with_vue_map(nir_block *block, void *closure)
    return true;
 }
 
-struct remap_patch_urb_offsets_state {
-   nir_builder b;
-   const struct brw_vue_map *vue_map;
-};
-
 static bool
-remap_patch_urb_offsets(nir_block *block, void *closure)
+remap_patch_urb_offsets(nir_block *block, nir_builder *b,
+                        const struct brw_vue_map *vue_map)
 {
-   struct remap_patch_urb_offsets_state *state = closure;
-
-   nir_foreach_instr_safe(block, instr) {
+   nir_foreach_instr_safe(instr, block) {
       if (instr->type != nir_instr_type_intrinsic)
          continue;
 
       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
 
-      gl_shader_stage stage = state->b.shader->stage;
+      gl_shader_stage stage = b->shader->stage;
 
       if ((stage == MESA_SHADER_TESS_CTRL && is_output(intrin)) ||
           (stage == MESA_SHADER_TESS_EVAL && is_input(intrin))) {
-         int vue_slot = state->vue_map->varying_to_slot[intrin->const_index[0]];
+         int vue_slot = vue_map->varying_to_slot[intrin->const_index[0]];
          assert(vue_slot != -1);
          intrin->const_index[0] = vue_slot;
 
@@ -176,22 +165,22 @@ remap_patch_urb_offsets(nir_block *block, void *closure)
             nir_const_value *const_vertex = nir_src_as_const_value(*vertex);
             if (const_vertex) {
                intrin->const_index[0] += const_vertex->u32[0] *
-                                         state->vue_map->num_per_vertex_slots;
+                                         vue_map->num_per_vertex_slots;
             } else {
-               state->b.cursor = nir_before_instr(&intrin->instr);
+               b->cursor = nir_before_instr(&intrin->instr);
 
                /* Multiply by the number of per-vertex slots. */
                nir_ssa_def *vertex_offset =
-                  nir_imul(&state->b,
-                           nir_ssa_for_src(&state->b, *vertex, 1),
-                           nir_imm_int(&state->b,
-                                       state->vue_map->num_per_vertex_slots));
+                  nir_imul(b,
+                           nir_ssa_for_src(b, *vertex, 1),
+                           nir_imm_int(b,
+                                       vue_map->num_per_vertex_slots));
 
                /* Add it to the existing offset */
                nir_src *offset = nir_get_io_offset_src(intrin);
                nir_ssa_def *total_offset =
-                  nir_iadd(&state->b, vertex_offset,
-                           nir_ssa_for_src(&state->b, *offset, 1));
+                  nir_iadd(b, vertex_offset,
+                           nir_ssa_for_src(b, *offset, 1));
 
                nir_instr_rewrite_src(&intrin->instr, offset,
                                      nir_src_for_ssa(total_offset));
@@ -214,11 +203,11 @@ brw_nir_lower_vs_inputs(nir_shader *nir,
       var->data.driver_location = var->data.location;
    }
 
-   /* Now use nir_lower_io to walk dereference chains.  Attribute arrays
-    * are loaded as one vec4 per element (or matrix column), so we use
-    * type_size_vec4 here.
+   /* Now use nir_lower_io to walk dereference chains.  Attribute arrays are
+    * loaded as one vec4 or dvec4 per element (or matrix column), depending on
+    * whether it is a double-precision type or not.
     */
-   nir_lower_io(nir, nir_var_shader_in, type_size_vec4);
+   nir_lower_io(nir, nir_var_shader_in, type_size_vs_input);
 
    /* This pass needs actual constants */
    nir_opt_constant_folding(nir);
@@ -229,17 +218,13 @@ brw_nir_lower_vs_inputs(nir_shader *nir,
                                        vs_attrib_wa_flags);
 
    if (is_scalar) {
-      /* Finally, translate VERT_ATTRIB_* values into the actual registers.
-       *
-       * Note that we can use nir->info.inputs_read instead of
-       * key->inputs_read since the two are identical aside from Gen4-5
-       * edge flag differences.
-       */
-      GLbitfield64 inputs_read = nir->info.inputs_read;
+      /* Finally, translate VERT_ATTRIB_* values into the actual registers. */
 
-      nir_foreach_function(nir, function) {
+      nir_foreach_function(function, nir) {
          if (function->impl) {
-            nir_foreach_block(function->impl, remap_vs_attrs, &inputs_read);
+            nir_foreach_block(block, function->impl) {
+               remap_vs_attrs(block, &nir->info);
+            }
          }
       }
    }
@@ -262,10 +247,11 @@ brw_nir_lower_vue_inputs(nir_shader *nir, bool is_scalar,
 
       add_const_offset_to_base(nir, nir_var_shader_in);
 
-      nir_foreach_function(nir, function) {
+      nir_foreach_function(function, nir) {
          if (function->impl) {
-            nir_foreach_block(function->impl, remap_inputs_with_vue_map,
-                              (void *) vue_map);
+            nir_foreach_block(block, function->impl) {
+               remap_inputs_with_vue_map(block, vue_map);
+            }
          }
       }
    }
@@ -274,9 +260,6 @@ brw_nir_lower_vue_inputs(nir_shader *nir, bool is_scalar,
 void
 brw_nir_lower_tes_inputs(nir_shader *nir, const struct brw_vue_map *vue_map)
 {
-   struct remap_patch_urb_offsets_state state;
-   state.vue_map = vue_map;
-
    foreach_list_typed(nir_variable, var, node, &nir->inputs) {
       var->data.driver_location = var->data.location;
    }
@@ -288,10 +271,13 @@ brw_nir_lower_tes_inputs(nir_shader *nir, const struct brw_vue_map *vue_map)
 
    add_const_offset_to_base(nir, nir_var_shader_in);
 
-   nir_foreach_function(nir, function) {
+   nir_foreach_function(function, nir) {
       if (function->impl) {
-         nir_builder_init(&state.b, function->impl);
-         nir_foreach_block(function->impl, remap_patch_urb_offsets, &state);
+         nir_builder b;
+         nir_builder_init(&b, function->impl);
+         nir_foreach_block(block, function->impl) {
+            remap_patch_urb_offsets(block, &b, vue_map);
+         }
       }
    }
 }
@@ -321,9 +307,6 @@ brw_nir_lower_vue_outputs(nir_shader *nir,
 void
 brw_nir_lower_tcs_outputs(nir_shader *nir, const struct brw_vue_map *vue_map)
 {
-   struct remap_patch_urb_offsets_state state;
-   state.vue_map = vue_map;
-
    nir_foreach_variable(var, &nir->outputs) {
       var->data.driver_location = var->data.location;
    }
@@ -335,10 +318,13 @@ brw_nir_lower_tcs_outputs(nir_shader *nir, const struct brw_vue_map *vue_map)
 
    add_const_offset_to_base(nir, nir_var_shader_out);
 
-   nir_foreach_function(nir, function) {
+   nir_foreach_function(function, nir) {
       if (function->impl) {
-         nir_builder_init(&state.b, function->impl);
-         nir_foreach_block(function->impl, remap_patch_urb_offsets, &state);
+         nir_builder b;
+         nir_builder_init(&b, function->impl);
+         nir_foreach_block(block, function->impl) {
+            remap_patch_urb_offsets(block, &b, vue_map);
+         }
       }
    }
 }
@@ -422,6 +408,16 @@ nir_optimize(nir_shader *nir, bool is_scalar)
       OPT(nir_opt_dead_cf);
       OPT(nir_opt_remove_phis);
       OPT(nir_opt_undef);
+      OPT_V(nir_lower_doubles, nir_lower_drcp |
+                               nir_lower_dsqrt |
+                               nir_lower_drsq |
+                               nir_lower_dtrunc |
+                               nir_lower_dfloor |
+                               nir_lower_dceil |
+                               nir_lower_dfract |
+                               nir_lower_dround_even |
+                               nir_lower_dmod);
+      OPT_V(nir_lower_double_pack);
    } while (progress);
 
    return nir;
@@ -437,14 +433,19 @@ nir_optimize(nir_shader *nir, bool is_scalar)
  * is_scalar = true to scalarize everything prior to code gen.
  */
 nir_shader *
-brw_preprocess_nir(nir_shader *nir, bool is_scalar)
+brw_preprocess_nir(const struct brw_compiler *compiler, nir_shader *nir)
 {
    bool progress; /* Written by OPT and OPT_V */
    (void)progress;
 
+   const bool is_scalar = compiler->scalar_stage[nir->stage];
+
    if (nir->stage == MESA_SHADER_GEOMETRY)
       OPT(nir_lower_gs_intrinsics);
 
+   if (compiler->precise_trig)
+      OPT(brw_nir_apply_trig_workarounds);
+
    static const nir_lower_tex_options tex_options = {
       .lower_txp = ~0,
    };
@@ -508,7 +509,7 @@ brw_postprocess_nir(nir_shader *nir,
 
    if (unlikely(debug_enabled)) {
       /* Re-index SSA defs so we print more sensible numbers. */
-      nir_foreach_function(nir, function) {
+      nir_foreach_function(function, nir) {
          if (function->impl)
             nir_index_ssa_defs(function->impl);
       }
@@ -560,6 +561,10 @@ brw_create_nir(struct brw_context *brw,
    /* First, lower the GLSL IR or Mesa IR to NIR */
    if (shader_prog) {
       nir = glsl_to_nir(shader_prog, stage, options);
+      nir_remove_dead_variables(nir, nir_var_shader_in | nir_var_shader_out);
+      OPT_V(nir_lower_io_to_temporaries,
+            nir_shader_get_entrypoint(nir),
+            true, false);
    } else {
       nir = prog_to_nir(prog, options);
       OPT_V(nir_convert_to_ssa); /* turn registers into SSA */
@@ -568,7 +573,19 @@ brw_create_nir(struct brw_context *brw,
 
    (void)progress;
 
-   nir = brw_preprocess_nir(nir, is_scalar);
+   nir = brw_preprocess_nir(brw->intelScreen->compiler, nir);
+
+   if (stage == MESA_SHADER_FRAGMENT) {
+      static const struct nir_lower_wpos_ytransform_options wpos_options = {
+         .state_tokens = {STATE_INTERNAL, STATE_FB_WPOS_Y_TRANSFORM, 0, 0, 0},
+         .fs_coord_pixel_center_integer = 1,
+         .fs_coord_origin_upper_left = 1,
+      };
+      _mesa_add_state_reference(prog->Parameters,
+                                (gl_state_index *) wpos_options.state_tokens);
+
+      OPT(nir_lower_wpos_ytransform, &wpos_options);
+   }
 
    OPT(nir_lower_system_values);
    OPT_V(brw_nir_lower_uniforms, is_scalar);