[Mesa-dev] [PATCH 2/3] nir: Add a pass to lower outputs to temporary variables

[Ian Romanick]

On 09/08/2015 04:51 PM, Jason Ekstrand wrote:
> This pass can be used as a helper for NIR producers so they don't have to
> worry about creating the temporaries themselves.
> ---
>  src/glsl/Makefile.sources                       |   1 +
>  src/glsl/nir/nir.h                              |   2 +
>  src/glsl/nir/nir_lower_outputs_to_temporaries.c | 131 ++++++++++++++++++++++++
>  3 files changed, 134 insertions(+)
>  create mode 100644 src/glsl/nir/nir_lower_outputs_to_temporaries.c
> 
> diff --git a/src/glsl/Makefile.sources b/src/glsl/Makefile.sources
> index c422303..6c2854f 100644
> --- a/src/glsl/Makefile.sources
> +++ b/src/glsl/Makefile.sources
> @@ -40,6 +40,7 @@ NIR_FILES = \
>  	nir/nir_lower_locals_to_regs.c \
>  	nir/nir_lower_idiv.c \
>  	nir/nir_lower_io.c \
> +	nir/nir_lower_outputs_to_temporaries.c \
>  	nir/nir_lower_phis_to_scalar.c \
>  	nir/nir_lower_samplers.cpp \
>  	nir/nir_lower_system_values.c \
> diff --git a/src/glsl/nir/nir.h b/src/glsl/nir/nir.h
> index c4db005..3b9d504 100644
> --- a/src/glsl/nir/nir.h
> +++ b/src/glsl/nir/nir.h
> @@ -1786,6 +1786,8 @@ void nir_lower_global_vars_to_local(nir_shader *shader);
>  
>  void nir_lower_locals_to_regs(nir_shader *shader);
>  
> +void nir_lower_outputs_to_temporaries(nir_shader *shader);
> +
>  void nir_assign_var_locations(struct exec_list *var_list,
>                                unsigned *size,
>                                int (*type_size)(const struct glsl_type *));
> diff --git a/src/glsl/nir/nir_lower_outputs_to_temporaries.c b/src/glsl/nir/nir_lower_outputs_to_temporaries.c
> new file mode 100644
> index 0000000..dd84ed4
> --- /dev/null
> +++ b/src/glsl/nir/nir_lower_outputs_to_temporaries.c
> @@ -0,0 +1,131 @@
> +/*
> + * Copyright © 2015 Intel Corporation
> + *
> + * Permission is hereby granted, free of charge, to any person obtaining a
> + * copy of this software and associated documentation files (the "Software"),
> + * to deal in the Software without restriction, including without limitation
> + * the rights to use, copy, modify, merge, publish, distribute, sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom the
> + * Software is furnished to do so, subject to the following conditions:
> + *
> + * The above copyright notice and this permission notice (including the next
> + * paragraph) shall be included in all copies or substantial portions of the
> + * Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
> + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
> + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
> + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
> + * IN THE SOFTWARE.
> + *
> + * Authors:
> + *    Jason Ekstrand (jason at jlekstrand.net)

GIT will keep track of this for us. :) We stopped putting by-lines in
the code a long, long time ago.  I sent out some patches a couple weeks
ago to remove all the remaining by-lines from src/drivers/dri/i965.

Assuming that there are no shader-db regressions, this series is

Reviewed-by: Ian Romanick <ian.d.romanick at intel.com>

> + *
> + */
> +
> +/*
> + * Implements a pass that lowers output variables to a temporary plus an
> + * output variable with a single copy at each exit point of the shader.
> + * This way the output variable is only ever written.
> + *
> + * Because valid NIR requires that output variables are never read, this
> + * pass is more of a helper for NIR producers and must be run before the
> + * shader is ever validated.
> + */
> +
> +#include "nir.h"
> +
> +struct lower_outputs_state {
> +   nir_shader *shader;
> +   struct exec_list old_outputs;
> +};
> +
> +static void
> +emit_output_copies(nir_cursor cursor, struct lower_outputs_state *state)
> +{
> +   assert(exec_list_length(&state->shader->outputs) ==
> +          exec_list_length(&state->old_outputs));
> +
> +   foreach_two_lists(out_node, &state->shader->outputs,
> +                     temp_node, &state->old_outputs) {
> +      nir_variable *output = exec_node_data(nir_variable, out_node, node);
> +      nir_variable *temp = exec_node_data(nir_variable, temp_node, node);
> +
> +      nir_intrinsic_instr *copy =
> +         nir_intrinsic_instr_create(state->shader, nir_intrinsic_copy_var);
> +      copy->variables[0] = nir_deref_var_create(copy, output);
> +      copy->variables[1] = nir_deref_var_create(copy, temp);
> +
> +      nir_instr_insert(cursor, &copy->instr);
> +   }
> +}
> +
> +static bool
> +emit_output_copies_block(nir_block *block, void *state)
> +{
> +   nir_foreach_instr(block, instr) {
> +      if (instr->type != nir_instr_type_intrinsic)
> +         continue;
> +
> +      nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
> +      if (intrin->intrinsic == nir_intrinsic_emit_vertex)
> +         emit_output_copies(nir_before_instr(&intrin->instr), state);
> +   }
> +
> +   return true;
> +}
> +
> +void
> +nir_lower_outputs_to_temporaries(nir_shader *shader)
> +{
> +   struct lower_outputs_state state;
> +
> +   state.shader = shader;
> +   exec_list_move_nodes_to(&shader->outputs, &state.old_outputs);
> +
> +   /* Walk over all of the outputs turn each output into a temporary and
> +    * make a new variable for the actual output.
> +    */
> +   foreach_list_typed(nir_variable, var, node, &state.old_outputs) {
> +      nir_variable *output = ralloc(shader, nir_variable);
> +      memcpy(output, var, sizeof *output);
> +
> +      /* The orignal is now the temporary */
> +      nir_variable *temp = var;
> +
> +      /* Give the output a new name with @out-temp appended */
> +      temp->name = ralloc_asprintf(var, "%s at out-temp", output->name);
> +      temp->data.mode = nir_var_global;
> +      temp->constant_initializer = NULL;
> +
> +      exec_list_push_tail(&shader->outputs, &output->node);
> +   }
> +
> +   nir_foreach_overload(shader, overload) {
> +      if (overload->impl == NULL)
> +         continue;
> +
> +      if (shader->stage == MESA_SHADER_GEOMETRY) {
> +         /* For geometry shaders, we have to emit the output copies right
> +          * before each EmitVertex call.
> +          */
> +         nir_foreach_block(overload->impl, emit_output_copies_block, &state);
> +      } else if (strcmp(overload->function->name, "main") == 0) {
> +         /* For all other shader types, we need to do the copies right before
> +          * the jumps to the end block.
> +          */
> +         struct set_entry *block_entry;
> +         set_foreach(overload->impl->end_block->predecessors, block_entry) {
> +            struct nir_block *block = (void *)block_entry->key;
> +            emit_output_copies(nir_after_block_before_jump(block), &state);
> +         }
> +      }
> +
> +      nir_metadata_preserve(overload->impl, nir_metadata_block_index |
> +                                            nir_metadata_dominance);
> +   }
> +
> +   exec_list_append(&shader->globals, &state.old_outputs);
> +}