C++ ralloc_free函数代码示例

static void
vc4_screen_destroy(struct pipe_screen *pscreen)
{
        vc4_bufmgr_destroy(pscreen);
        ralloc_free(pscreen);
}

bool
brwCreateContext(int api,
	         const struct gl_config *mesaVis,
		 __DRIcontext *driContextPriv,
                 unsigned major_version,
                 unsigned minor_version,
                 uint32_t flags,
                 unsigned *error,
	         void *sharedContextPrivate)
{
   __DRIscreen *sPriv = driContextPriv->driScreenPriv;
   struct intel_screen *screen = sPriv->driverPrivate;
   struct dd_function_table functions;

   struct brw_context *brw = rzalloc(NULL, struct brw_context);
   if (!brw) {
      printf("%s: failed to alloc context\n", __FUNCTION__);
      *error = __DRI_CTX_ERROR_NO_MEMORY;
      return false;
   }

   /* brwInitVtbl needs to know the chipset generation so that it can set the
    * right pointers.
    */
   brw->gen = screen->gen;

   brwInitVtbl( brw );

   brwInitDriverFunctions(screen, &functions);

   struct gl_context *ctx = &brw->ctx;

   if (!intelInitContext( brw, api, major_version, minor_version,
                          mesaVis, driContextPriv,
			  sharedContextPrivate, &functions,
			  error)) {
      ralloc_free(brw);
      return false;
   }

   brw_initialize_context_constants(brw);

   /* Reinitialize the context point state.  It depends on ctx->Const values. */
   _mesa_init_point(ctx);

   if (brw->gen >= 6) {
      /* Create a new hardware context.  Using a hardware context means that
       * our GPU state will be saved/restored on context switch, allowing us
       * to assume that the GPU is in the same state we left it in.
       *
       * This is required for transform feedback buffer offsets, query objects,
       * and also allows us to reduce how much state we have to emit.
       */
      brw->hw_ctx = drm_intel_gem_context_create(brw->bufmgr);

      if (!brw->hw_ctx) {
         fprintf(stderr, "Gen6+ requires Kernel 3.6 or later.\n");
         ralloc_free(brw);
         return false;
      }
   }

   brw_init_surface_formats(brw);

   /* Initialize swrast, tnl driver tables: */
   TNLcontext *tnl = TNL_CONTEXT(ctx);
   if (tnl)
      tnl->Driver.RunPipeline = _tnl_run_pipeline;

   ctx->DriverFlags.NewTransformFeedback = BRW_NEW_TRANSFORM_FEEDBACK;
   ctx->DriverFlags.NewRasterizerDiscard = BRW_NEW_RASTERIZER_DISCARD;
   ctx->DriverFlags.NewUniformBuffer = BRW_NEW_UNIFORM_BUFFER;

   if (brw->is_g4x || brw->gen >= 5) {
      brw->CMD_VF_STATISTICS = GM45_3DSTATE_VF_STATISTICS;
      brw->CMD_PIPELINE_SELECT = CMD_PIPELINE_SELECT_GM45;
      brw->has_surface_tile_offset = true;
      if (brw->gen < 6)
	  brw->has_compr4 = true;
      brw->has_aa_line_parameters = true;
      brw->has_pln = true;
  } else {
      brw->CMD_VF_STATISTICS = GEN4_3DSTATE_VF_STATISTICS;
      brw->CMD_PIPELINE_SELECT = CMD_PIPELINE_SELECT_965;
   }

   /* WM maximum threads is number of EUs times number of threads per EU. */
   assert(brw->gen <= 7);

   if (brw->is_haswell) {
      if (brw->gt == 1) {
	 brw->max_wm_threads = 102;
	 brw->max_vs_threads = 70;
	 brw->urb.size = 128;
	 brw->urb.max_vs_entries = 640;
	 brw->urb.max_gs_entries = 256;
      } else if (brw->gt == 2) {
	 brw->max_wm_threads = 204;
	 brw->max_vs_threads = 280;
	 brw->urb.size = 256;
//.........这里部分代码省略.........

static bool
brw_codegen_gs_prog(struct brw_context *brw,
                    struct brw_program *gp,
                    struct brw_gs_prog_key *key)
{
   struct brw_compiler *compiler = brw->screen->compiler;
   const struct gen_device_info *devinfo = &brw->screen->devinfo;
   struct brw_stage_state *stage_state = &brw->gs.base;
   struct brw_gs_prog_data prog_data;
   bool start_busy = false;
   double start_time = 0;

   memset(&prog_data, 0, sizeof(prog_data));

   assign_gs_binding_table_offsets(devinfo, &gp->program, &prog_data);

   /* Allocate the references to the uniforms that will end up in the
    * prog_data associated with the compiled program, and which will be freed
    * by the state cache.
    *
    * Note: param_count needs to be num_uniform_components * 4, since we add
    * padding around uniform values below vec4 size, so the worst case is that
    * every uniform is a float which gets padded to the size of a vec4.
    */
   int param_count = gp->program.nir->num_uniforms / 4;

   prog_data.base.base.param =
      rzalloc_array(NULL, const gl_constant_value *, param_count);
   prog_data.base.base.pull_param =
      rzalloc_array(NULL, const gl_constant_value *, param_count);
   prog_data.base.base.image_param =
      rzalloc_array(NULL, struct brw_image_param,
                    gp->program.info.num_images);
   prog_data.base.base.nr_params = param_count;
   prog_data.base.base.nr_image_params = gp->program.info.num_images;

   brw_nir_setup_glsl_uniforms(gp->program.nir, &gp->program,
                               &prog_data.base.base,
                               compiler->scalar_stage[MESA_SHADER_GEOMETRY]);
   brw_nir_analyze_ubo_ranges(compiler, gp->program.nir,
                              prog_data.base.base.ubo_ranges);

   uint64_t outputs_written = gp->program.info.outputs_written;

   brw_compute_vue_map(devinfo,
                       &prog_data.base.vue_map, outputs_written,
                       gp->program.info.separate_shader);

   int st_index = -1;
   if (INTEL_DEBUG & DEBUG_SHADER_TIME)
      st_index = brw_get_shader_time_index(brw, &gp->program, ST_GS, true);

   if (unlikely(brw->perf_debug)) {
      start_busy = brw->batch.last_bo && brw_bo_busy(brw->batch.last_bo);
      start_time = get_time();
   }

   void *mem_ctx = ralloc_context(NULL);
   unsigned program_size;
   char *error_str;
   const unsigned *program =
      brw_compile_gs(brw->screen->compiler, brw, mem_ctx, key,
                     &prog_data, gp->program.nir, &gp->program,
                     st_index, &program_size, &error_str);
   if (program == NULL) {
      ralloc_strcat(&gp->program.sh.data->InfoLog, error_str);
      _mesa_problem(NULL, "Failed to compile geometry shader: %s\n", error_str);

      ralloc_free(mem_ctx);
      return false;
   }

   if (unlikely(brw->perf_debug)) {
      if (gp->compiled_once) {
         brw_gs_debug_recompile(brw, &gp->program, key);
      }
      if (start_busy && !brw_bo_busy(brw->batch.last_bo)) {
         perf_debug("GS compile took %.03f ms and stalled the GPU\n",
                    (get_time() - start_time) * 1000);
      }
      gp->compiled_once = true;
   }

   /* Scratch space is used for register spilling */
   brw_alloc_stage_scratch(brw, stage_state,
                           prog_data.base.base.total_scratch,
                           devinfo->max_gs_threads);

   brw_upload_cache(&brw->cache, BRW_CACHE_GS_PROG,
                    key, sizeof(*key),
                    program, program_size,
                    &prog_data, sizeof(prog_data),
                    &stage_state->prog_offset, &brw->gs.base.prog_data);
   ralloc_free(mem_ctx);

   return true;
}

	~has_recursion_visitor()
	{
		hash_table_dtor(this->function_hash);
		ralloc_free(this->mem_ctx);
	}

/**
 * All Mesa program -> GPU code generation goes through this function.
 * Depending on the instructions used (i.e. flow control instructions)
 * we'll use one of two code generators.
 */
bool do_wm_prog(struct brw_context *brw,
		struct gl_shader_program *prog,
		struct brw_fragment_program *fp,
		struct brw_wm_prog_key *key)
{
   struct brw_wm_compile *c;
   const GLuint *program;
   struct gl_shader *fs = NULL;
   GLuint program_size;

   if (prog)
      fs = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];

   c = rzalloc(NULL, struct brw_wm_compile);

   /* Allocate the references to the uniforms that will end up in the
    * prog_data associated with the compiled program, and which will be freed
    * by the state cache.
    */
   int param_count;
   if (fs) {
      param_count = fs->num_uniform_components;
   } else {
      param_count = fp->program.Base.Parameters->NumParameters * 4;
   }
   /* The backend also sometimes adds params for texture size. */
   param_count += 2 * BRW_MAX_TEX_UNIT;
   c->prog_data.param = rzalloc_array(NULL, const float *, param_count);
   c->prog_data.pull_param = rzalloc_array(NULL, const float *, param_count);

   memcpy(&c->key, key, sizeof(*key));

   c->prog_data.barycentric_interp_modes =
      brw_compute_barycentric_interp_modes(brw, c->key.flat_shade,
                                           &fp->program);

   program = brw_wm_fs_emit(brw, c, &fp->program, prog, &program_size);
   if (program == NULL)
      return false;

   /* Scratch space is used for register spilling */
   if (c->last_scratch) {
      perf_debug("Fragment shader triggered register spilling.  "
                 "Try reducing the number of live scalar values to "
                 "improve performance.\n");

      c->prog_data.total_scratch = brw_get_scratch_size(c->last_scratch);

      brw_get_scratch_bo(brw, &brw->wm.scratch_bo,
			 c->prog_data.total_scratch * brw->max_wm_threads);
   }

   if (unlikely(INTEL_DEBUG & DEBUG_WM))
      fprintf(stderr, "\n");

   brw_upload_cache(&brw->cache, BRW_WM_PROG,
		    &c->key, sizeof(c->key),
		    program, program_size,
		    &c->prog_data, sizeof(c->prog_data),
		    &brw->wm.prog_offset, &brw->wm.prog_data);

   ralloc_free(c);

   return true;
}

static bool
constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
{
   nir_const_value src[NIR_MAX_VEC_COMPONENTS];

   if (!instr->dest.dest.is_ssa)
      return false;

   /* In the case that any outputs/inputs have unsized types, then we need to
    * guess the bit-size. In this case, the validator ensures that all
    * bit-sizes match so we can just take the bit-size from first
    * output/input with an unsized type. If all the outputs/inputs are sized
    * then we don't need to guess the bit-size at all because the code we
    * generate for constant opcodes in this case already knows the sizes of
    * the types involved and does not need the provided bit-size for anything
    * (although it still requires to receive a valid bit-size).
    */
   unsigned bit_size = 0;
   if (!nir_alu_type_get_type_size(nir_op_infos[instr->op].output_type))
      bit_size = instr->dest.dest.ssa.bit_size;

   for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
      if (!instr->src[i].src.is_ssa)
         return false;

      if (bit_size == 0 &&
          !nir_alu_type_get_type_size(nir_op_infos[instr->op].input_sizes[i])) {
         bit_size = instr->src[i].src.ssa->bit_size;
      }

      nir_instr *src_instr = instr->src[i].src.ssa->parent_instr;

      if (src_instr->type != nir_instr_type_load_const)
         return false;
      nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr);

      for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i);
           j++) {
         switch(load_const->def.bit_size) {
         case 64:
            src[i].u64[j] = load_const->value.u64[instr->src[i].swizzle[j]];
            break;
         case 32:
            src[i].u32[j] = load_const->value.u32[instr->src[i].swizzle[j]];
            break;
         case 16:
            src[i].u16[j] = load_const->value.u16[instr->src[i].swizzle[j]];
            break;
         case 8:
            src[i].u8[j] = load_const->value.u8[instr->src[i].swizzle[j]];
            break;
         default:
            unreachable("Invalid bit size");
         }
      }

      /* We shouldn't have any source modifiers in the optimization loop. */
      assert(!instr->src[i].abs && !instr->src[i].negate);
   }

   if (bit_size == 0)
      bit_size = 32;

   /* We shouldn't have any saturate modifiers in the optimization loop. */
   assert(!instr->dest.saturate);

   nir_const_value dest =
      nir_eval_const_opcode(instr->op, instr->dest.dest.ssa.num_components,
                            bit_size, src);

   nir_load_const_instr *new_instr =
      nir_load_const_instr_create(mem_ctx,
                                  instr->dest.dest.ssa.num_components,
                                  instr->dest.dest.ssa.bit_size);

   new_instr->value = dest;

   nir_instr_insert_before(&instr->instr, &new_instr->instr);

   nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa,
                            nir_src_for_ssa(&new_instr->def));

   nir_instr_remove(&instr->instr);
   ralloc_free(instr);

   return true;
}

brw_blorp_const_color_program::~brw_blorp_const_color_program()
{
   ralloc_free(mem_ctx);
}

 ~get_sampler_name()
 {
    ralloc_free(this->mem_ctx);
 }

	~string_buffer()
	{
		ralloc_free(m_Ptr);
	}

//.........这里部分代码省略.........
    *   512 bytes for varyings (a varying component is 4 bytes and
    *             gl_MaxGeometryOutputComponents = 128)
    *    16 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
    *             bytes)
    *    16 bytes overhead for gl_Position (we allocate it a slot in the VUE
    *             even if it's not used)
    *    32 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
    *             whenever clip planes are enabled, even if the shader doesn't
    *             write to gl_ClipDistance)
    *    16 bytes overhead since the VUE size must be a multiple of 32 bytes
    *             (see below)--this causes up to 1 VUE slot to be wasted
    *   400 bytes available for varying packing overhead
    *
    * Worst-case varying packing overhead is 3/4 of a varying slot (12 bytes)
    * per interpolation type, so this is plenty.
    *
    */
   unsigned output_vertex_size_bytes = c.prog_data.base.vue_map.num_slots * 16;
   assert(output_vertex_size_bytes <= GEN7_MAX_GS_OUTPUT_VERTEX_SIZE_BYTES);
   c.prog_data.output_vertex_size_hwords =
      ALIGN(output_vertex_size_bytes, 32) / 32;

   /* Compute URB entry size.  The maximum allowed URB entry size is 32k.
    * That divides up as follows:
    *
    *     64 bytes for the control data header (cut indices or StreamID bits)
    *   4096 bytes for varyings (a varying component is 4 bytes and
    *              gl_MaxGeometryTotalOutputComponents = 1024)
    *   4096 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
    *              bytes/vertex and gl_MaxGeometryOutputVertices is 256)
    *   4096 bytes overhead for gl_Position (we allocate it a slot in the VUE
    *              even if it's not used)
    *   8192 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
    *              whenever clip planes are enabled, even if the shader doesn't
    *              write to gl_ClipDistance)
    *   4096 bytes overhead since the VUE size must be a multiple of 32
    *              bytes (see above)--this causes up to 1 VUE slot to be wasted
    *   8128 bytes available for varying packing overhead
    *
    * Worst-case varying packing overhead is 3/4 of a varying slot per
    * interpolation type, which works out to 3072 bytes, so this would allow
    * us to accommodate 2 interpolation types without any danger of running
    * out of URB space.
    *
    * In practice, the risk of running out of URB space is very small, since
    * the above figures are all worst-case, and most of them scale with the
    * number of output vertices.  So we'll just calculate the amount of space
    * we need, and if it's too large, fail to compile.
    */
   unsigned output_size_bytes =
      c.prog_data.output_vertex_size_hwords * 32 * gp->program.VerticesOut;
   output_size_bytes += 32 * c.prog_data.control_data_header_size_hwords;

   assert(output_size_bytes >= 1);
   if (output_size_bytes > GEN7_MAX_GS_URB_ENTRY_SIZE_BYTES)
      return false;

   /* URB entry sizes are stored as a multiple of 64 bytes. */
   c.prog_data.base.urb_entry_size = ALIGN(output_size_bytes, 64) / 64;

   c.prog_data.output_topology = prim_to_hw_prim[gp->program.OutputType];

   brw_compute_vue_map(brw, &c.input_vue_map, c.key.input_varyings);

   /* GS inputs are read from the VUE 256 bits (2 vec4's) at a time, so we
    * need to program a URB read length of ceiling(num_slots / 2).
    */
   c.prog_data.base.urb_read_length = (c.input_vue_map.num_slots + 1) / 2;

   void *mem_ctx = ralloc_context(NULL);
   unsigned program_size;
   const unsigned *program =
      brw_gs_emit(brw, prog, &c, mem_ctx, &program_size);
   if (program == NULL) {
      ralloc_free(mem_ctx);
      return false;
   }

   /* Scratch space is used for register spilling */
   if (c.base.last_scratch) {
      perf_debug("Geometry shader triggered register spilling.  "
                 "Try reducing the number of live vec4 values to "
                 "improve performance.\n");

      c.prog_data.base.total_scratch
         = brw_get_scratch_size(c.base.last_scratch*REG_SIZE);

      brw_get_scratch_bo(brw, &stage_state->scratch_bo,
			 c.prog_data.base.total_scratch * brw->max_gs_threads);
   }

   brw_upload_cache(&brw->cache, BRW_GS_PROG,
                    &c.key, sizeof(c.key),
                    program, program_size,
                    &c.prog_data, sizeof(c.prog_data),
                    &stage_state->prog_offset, &brw->gs.prog_data);
   ralloc_free(mem_ctx);

   return true;
}

void
brw_wm_clear_compile(struct brw_context *brw,
                     struct brw_wm_compile *c)
{
   ralloc_free(c);
}

/**
 * All Mesa program -> GPU code generation goes through this function.
 * Depending on the instructions used (i.e. flow control instructions)
 * we'll use one of two code generators.
 */
bool do_wm_prog(struct brw_context *brw,
		struct gl_shader_program *prog,
		struct brw_fragment_program *fp,
		struct brw_wm_prog_key *key)
{
   struct gl_context *ctx = &brw->ctx;
   void *mem_ctx = ralloc_context(NULL);
   struct brw_wm_prog_data prog_data;
   const GLuint *program;
   struct gl_shader *fs = NULL;
   GLuint program_size;

   if (prog)
      fs = prog->_LinkedShaders[MESA_SHADER_FRAGMENT];

   memset(&prog_data, 0, sizeof(prog_data));

   /* Allocate the references to the uniforms that will end up in the
    * prog_data associated with the compiled program, and which will be freed
    * by the state cache.
    */
   int param_count;
   if (fs) {
      param_count = fs->num_uniform_components;
   } else {
      param_count = fp->program.Base.Parameters->NumParameters * 4;
   }
   /* The backend also sometimes adds params for texture size. */
   param_count += 2 * ctx->Const.Program[MESA_SHADER_FRAGMENT].MaxTextureImageUnits;
   prog_data.base.param =
      rzalloc_array(NULL, const gl_constant_value *, param_count);
   prog_data.base.pull_param =
      rzalloc_array(NULL, const gl_constant_value *, param_count);
   prog_data.base.nr_params = param_count;

   prog_data.barycentric_interp_modes =
      brw_compute_barycentric_interp_modes(brw, key->flat_shade,
                                           key->persample_shading,
                                           &fp->program);

   program = brw_wm_fs_emit(brw, mem_ctx, key, &prog_data,
                            &fp->program, prog, &program_size);
   if (program == NULL) {
      ralloc_free(mem_ctx);
      return false;
   }

   if (prog_data.total_scratch) {
      brw_get_scratch_bo(brw, &brw->wm.base.scratch_bo,
			 prog_data.total_scratch * brw->max_wm_threads);
   }

   if (unlikely(INTEL_DEBUG & DEBUG_WM))
      fprintf(stderr, "\n");

   brw_upload_cache(&brw->cache, BRW_WM_PROG,
		    key, sizeof(struct brw_wm_prog_key),
		    program, program_size,
		    &prog_data, sizeof(prog_data),
		    &brw->wm.base.prog_offset, &brw->wm.prog_data);

   ralloc_free(mem_ctx);

   return true;
}

static void compile_gs_prog( struct brw_context *brw,
			     struct brw_gs_prog_key *key )
{
   struct brw_gs_compile c;
   const GLuint *program;
   void *mem_ctx;
   GLuint program_size;

   memset(&c, 0, sizeof(c));

   c.key = *key;
   c.vue_map = brw->vs.prog_data->base.vue_map;
   c.nr_regs = (c.vue_map.num_slots + 1)/2;

   mem_ctx = ralloc_context(NULL);

   /* Begin the compilation:
    */
   brw_init_compile(brw, &c.func, mem_ctx);

   c.func.single_program_flow = 1;

   /* For some reason the thread is spawned with only 4 channels
    * unmasked.
    */
   brw_set_mask_control(&c.func, BRW_MASK_DISABLE);

   if (brw->gen >= 6) {
      unsigned num_verts;
      bool check_edge_flag;
      /* On Sandybridge, we use the GS for implementing transform feedback
       * (called "Stream Out" in the PRM).
       */
      switch (key->primitive) {
      case _3DPRIM_POINTLIST:
         num_verts = 1;
         check_edge_flag = false;
	 break;
      case _3DPRIM_LINELIST:
      case _3DPRIM_LINESTRIP:
      case _3DPRIM_LINELOOP:
         num_verts = 2;
         check_edge_flag = false;
	 break;
      case _3DPRIM_TRILIST:
      case _3DPRIM_TRIFAN:
      case _3DPRIM_TRISTRIP:
      case _3DPRIM_RECTLIST:
	 num_verts = 3;
         check_edge_flag = false;
         break;
      case _3DPRIM_QUADLIST:
      case _3DPRIM_QUADSTRIP:
      case _3DPRIM_POLYGON:
         num_verts = 3;
         check_edge_flag = true;
         break;
      default:
	 assert(!"Unexpected primitive type in Gen6 SOL program.");
	 return;
      }
      gen6_sol_program(&c, key, num_verts, check_edge_flag);
   } else {
      /* On Gen4-5, we use the GS to decompose certain types of primitives.
       * Note that primitives which don't require a GS program have already
       * been weeded out by now.
       */
      switch (key->primitive) {
      case _3DPRIM_QUADLIST:
	 brw_gs_quads( &c, key );
	 break;
      case _3DPRIM_QUADSTRIP:
	 brw_gs_quad_strip( &c, key );
	 break;
      case _3DPRIM_LINELOOP:
	 brw_gs_lines( &c );
	 break;
      default:
	 ralloc_free(mem_ctx);
	 return;
      }
   }

   /* get the program
    */
   program = brw_get_program(&c.func, &program_size);

   if (unlikely(INTEL_DEBUG & DEBUG_GS)) {
      int i;

      printf("gs:\n");
      for (i = 0; i < program_size / sizeof(struct brw_instruction); i++)
	 brw_disasm(stdout, &((struct brw_instruction *)program)[i],
		    brw->gen);
      printf("\n");
    }

   brw_upload_cache(&brw->cache, BRW_GS_PROG,
		    &c.key, sizeof(c.key),
		    program, program_size,
//.........这里部分代码省略.........

//.........这里部分代码省略.........

      if (dst_is_msaa) {
         arb_sample_shading_extension_string = "#extension GL_ARB_sample_shading : enable";
         sample_resolve = ralloc_asprintf(mem_ctx, "   out_color = texelFetch(texSampler, i%s(texCoords), gl_SampleID);", texcoord_type);
         merge_function = "";
      } else {
         int i;
         int step;

         if (src_datatype == GL_INT || src_datatype == GL_UNSIGNED_INT) {
            merge_function =
               "gvec4 merge(gvec4 a, gvec4 b) { return (a >> gvec4(1)) + (b >> gvec4(1)) + (a & b & gvec4(1)); }\n";
         } else {
            /* The divide will happen at the end for floats. */
            merge_function =
               "vec4 merge(vec4 a, vec4 b) { return (a + b); }\n";
         }

         arb_sample_shading_extension_string = "";

         /* We're assuming power of two samples for this resolution procedure.
          *
          * To avoid losing any floating point precision if the samples all
          * happen to have the same value, we merge pairs of values at a time
          * (so the floating point exponent just gets increased), rather than
          * doing a naive sum and dividing.
          */
         assert(_mesa_is_pow_two(samples));
         /* Fetch each individual sample. */
         sample_resolve = rzalloc_size(mem_ctx, 1);
         for (i = 0; i < samples; i++) {
            ralloc_asprintf_append(&sample_resolve,
                                   "   gvec4 sample_1_%d = texelFetch(texSampler, i%s(texCoords), %d);\n",
                                   i, texcoord_type, i);
         }
         /* Now, merge each pair of samples, then merge each pair of those,
          * etc.
          */
         for (step = 2; step <= samples; step *= 2) {
            for (i = 0; i < samples; i += step) {
               ralloc_asprintf_append(&sample_resolve,
                                      "   gvec4 sample_%d_%d = merge(sample_%d_%d, sample_%d_%d);\n",
                                      step, i,
                                      step / 2, i,
                                      step / 2, i + step / 2);
            }
         }

         /* Scale the final result. */
         if (src_datatype == GL_UNSIGNED_INT || src_datatype == GL_INT) {
            ralloc_asprintf_append(&sample_resolve,
                                   "   out_color = sample_%d_0;\n",
                                   samples);
         } else {
            ralloc_asprintf_append(&sample_resolve,
                                   "   gl_FragColor = sample_%d_0 / %f;\n",
                                   samples, (float)samples);
         }
      }

      vs_source = ralloc_asprintf(mem_ctx,
                                  "#version 130\n"
                                  "in vec2 position;\n"
                                  "in %s textureCoords;\n"
                                  "out %s texCoords;\n"
                                  "void main()\n"
                                  "{\n"
                                  "   texCoords = textureCoords;\n"
                                  "   gl_Position = vec4(position, 0.0, 1.0);\n"
                                  "}\n",
                                  texcoord_type,
                                  texcoord_type);
      fs_source = ralloc_asprintf(mem_ctx,
                                  "#version 130\n"
                                  "#extension GL_ARB_texture_multisample : enable\n"
                                  "%s\n"
                                  "#define gvec4 %svec4\n"
                                  "uniform %ssampler2DMS%s texSampler;\n"
                                  "in %s texCoords;\n"
                                  "out gvec4 out_color;\n"
                                  "\n"
                                  "%s" /* merge_function */
                                  "void main()\n"
                                  "{\n"
                                  "%s\n" /* sample_resolve */
                                  "}\n",
                                  arb_sample_shading_extension_string,
                                  vec4_prefix,
                                  vec4_prefix,
                                  sampler_array_suffix,
                                  texcoord_type,
                                  merge_function,
                                  sample_resolve);
   }

   _mesa_meta_compile_and_link_program(ctx, vs_source, fs_source, name,
                                       &blit->msaa_shaders[shader_index]);

   ralloc_free(mem_ctx);
}

void
program_resource_visitor::process(ir_variable *var)
{
    const glsl_type *t = var->type;
    const bool row_major =
        var->data.matrix_layout == GLSL_MATRIX_LAYOUT_ROW_MAJOR;

    /* false is always passed for the row_major parameter to the other
     * processing functions because no information is available to do
     * otherwise.  See the warning in linker.h.
     */

    /* Only strdup the name if we actually will need to modify it. */
    if (var->data.from_named_ifc_block_array) {
        /* lower_named_interface_blocks created this variable by lowering an
         * interface block array to an array variable.  For example if the
         * original source code was:
         *
         *     out Blk { vec4 bar } foo[3];
         *
         * Then the variable is now:
         *
         *     out vec4 bar[3];
         *
         * We need to visit each array element using the names constructed like
         * so:
         *
         *     Blk[0].bar
         *     Blk[1].bar
         *     Blk[2].bar
         */
        assert(t->is_array());
        const glsl_type *ifc_type = var->get_interface_type();
        char *name = ralloc_strdup(NULL, ifc_type->name);
        size_t name_length = strlen(name);
        for (unsigned i = 0; i < t->length; i++) {
            size_t new_length = name_length;
            ralloc_asprintf_rewrite_tail(&name, &new_length, "[%u].%s", i,
                                         var->name);
            /* Note: row_major is only meaningful for uniform blocks, and
             * lowering is only applied to non-uniform interface blocks, so we
             * can safely pass false for row_major.
             */
            recursion(var->type, &name, new_length, row_major, NULL, false);
        }
        ralloc_free(name);
    } else if (var->data.from_named_ifc_block_nonarray) {
        /* lower_named_interface_blocks created this variable by lowering a
         * named interface block (non-array) to an ordinary variable.  For
         * example if the original source code was:
         *
         *     out Blk { vec4 bar } foo;
         *
         * Then the variable is now:
         *
         *     out vec4 bar;
         *
         * We need to visit this variable using the name:
         *
         *     Blk.bar
         */
        const glsl_type *ifc_type = var->get_interface_type();
        char *name = ralloc_asprintf(NULL, "%s.%s", ifc_type->name, var->name);
        /* Note: row_major is only meaningful for uniform blocks, and lowering
         * is only applied to non-uniform interface blocks, so we can safely
         * pass false for row_major.
         */
        recursion(var->type, &name, strlen(name), row_major, NULL, false);
        ralloc_free(name);
    } else if (t->without_array()->is_record()) {
        char *name = ralloc_strdup(NULL, var->name);
        recursion(var->type, &name, strlen(name), row_major, NULL, false);
        ralloc_free(name);
    } else if (t->is_interface()) {
        char *name = ralloc_strdup(NULL, var->type->name);
        recursion(var->type, &name, strlen(name), row_major, NULL, false);
        ralloc_free(name);
    } else if (t->is_array() && t->fields.array->is_interface()) {
        char *name = ralloc_strdup(NULL, var->type->fields.array->name);
        recursion(var->type, &name, strlen(name), row_major, NULL, false);
        ralloc_free(name);
    } else {
        this->visit_field(t, var->name, row_major, NULL, false);
    }
}