[Mesa-dev] [PATCH v2 1/3] nir: Add a pass to lower vector phi nodes to scalar phi nodes

[Eric Anholt]

Jason Ekstrand <jason at jlekstrand.net> writes:
> diff --git a/src/glsl/nir/nir_lower_phis_to_scalar.c b/src/glsl/nir/nir_lower_phis_to_scalar.c
> new file mode 100644
> index 0000000..9f901d6
> --- /dev/null
> +++ b/src/glsl/nir/nir_lower_phis_to_scalar.c
> @@ -0,0 +1,238 @@
> +/*
> + * Copyright © 2014 Intel Corporation

Possibly update your copyright date?

> + *
> + * 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)
> + *
> + */
> +
> +#include "nir.h"
> +
> +/*
> + * Implements common subexpression elimination
> + */

Stale comment.

> +struct lower_phis_to_scalar_state {
> +   void *mem_ctx;
> +   void *dead_ctx;
> +
> +   /* Hash table marking which phi nodes are scalarizable.  The key is
> +    * pointers to phi instructions and the entry is either NULL for not
> +    * scalarizable or non-null for scalarizable.
> +    */
> +   struct hash_table *phi_table;
> +};
> +
> +/* Determines if the given phi node should be lowered.  The only phi nodes
> + * we will scalarize at the moment are those where all of the sources are
> + * scalarizable.
> + */
> +static bool
> +should_lower_phi(nir_phi_instr *phi, struct lower_phis_to_scalar_state *state)
> +{
> +   /* Already scalar */
> +   if (phi->dest.ssa.num_components == 1)
> +      return false;
> +
> +   struct hash_entry *entry = _mesa_hash_table_search(state->phi_table, phi);
> +   if (entry)
> +      return entry->data != NULL;
> +
> +   nir_foreach_phi_src(phi, src) {
> +      /* Don't know what to do with non-ssa sources */
> +      if (!src->src.is_ssa)
> +         return false;
> +      nir_instr *src_instr = src->src.ssa->parent_instr;
> +      switch (src_instr->type) {
> +      case nir_instr_type_alu: {
> +         nir_alu_instr *src_alu = nir_instr_as_alu(src_instr);
> +
> +         /* ALU operations with output_size == 0 should be scalarized.  We
> +          * will also see a bunch of vecN operations from scalarizing ALU
> +          * operations and, since they can easily be copy-propagated, they
> +          * are ok too.
> +          */
> +         return nir_op_infos[src_alu->op].output_size == 0 ||
> +                src_alu->op != nir_op_vec2 ||
> +                src_alu->op != nir_op_vec3 ||
> +                src_alu->op != nir_op_vec4;
> +      }
> +
> +      case nir_instr_type_phi: {
> +         nir_phi_instr *src_phi = nir_instr_as_phi(src_instr);
> +
> +         /* Insert an entry and mark it as scalarizable for now. That way
> +          * we don't recurse forever and a cycle in the depencence graph
> +          * won't automatically make us fail to scalarize.
> +          */
> +         entry = _mesa_hash_table_insert(state->phi_table, src_phi, (void *)1);

I expect "(void *)1" will give compiler warnings generally.  "(void
*)(uintptr_t)1" is the usual workaround.

Useful comment, though!

> +         bool scalarizable = should_lower_phi(src_phi, state);
> +         entry->data = (void *)scalarizable;

No need to cast to void *, since not C++.

> +static bool
> +lower_phis_to_scalar_block(nir_block *block, void *void_state)
> +{
> +   struct lower_phis_to_scalar_state *state = void_state;
> +
> +   /* Find the last phi node in the block */
> +   nir_phi_instr *last_phi = NULL;
> +   nir_foreach_instr(block, instr) {
> +      if (instr->type != nir_instr_type_phi)
> +         break;
> +
> +      last_phi = nir_instr_as_phi(instr);
> +   }
> +
> +   /* We have to handle the phi nodes in their own pass due to the way
> +    * we're modifying the linked list of instructions.
> +    */
> +   nir_foreach_instr_safe(block, instr) {
> +      if (instr->type != nir_instr_type_phi)
> +         break;
> +
> +      nir_phi_instr *phi = nir_instr_as_phi(instr);
> +
> +      if (!should_lower_phi(phi, state))
> +         continue;
> +
> +      /* Create a vecN operation to combine the results.  Most of these
> +       * will be redundant, but copy propagation should clean them up for
> +       * us.  No need to add the complexity here.
> +       */
> +      nir_op vec_op;
> +      switch (phi->dest.ssa.num_components) {
> +      case 2: vec_op = nir_op_vec2; break;
> +      case 3: vec_op = nir_op_vec3; break;
> +      case 4: vec_op = nir_op_vec4; break;
> +      default: unreachable("Invalid number of components");
> +      }
> +
> +      nir_alu_instr *vec = nir_alu_instr_create(state->mem_ctx, vec_op);
> +      vec->dest.dest.is_ssa = true;
> +      nir_ssa_def_init(&vec->instr, &vec->dest.dest.ssa,
> +                       phi->dest.ssa.num_components, NULL);

nir_ssa_dest_init() instead? (throughout the file)

> +      vec->dest.write_mask = (1 << phi->dest.ssa.num_components) - 1;
> +
> +      for (unsigned i = 0; i < phi->dest.ssa.num_components; i++) {
> +         nir_phi_instr *new_phi = nir_phi_instr_create(state->mem_ctx);
> +         new_phi->dest.is_ssa = true;
> +         nir_ssa_def_init(&new_phi->instr, &new_phi->dest.ssa, 1, NULL);
> +
> +         vec->src[i].src.is_ssa = true;
> +         vec->src[i].src.ssa = &new_phi->dest.ssa;
> +
> +         nir_foreach_phi_src(phi, src) {
> +            /* We need to insert a mov to grab the i'th component of src */
> +            nir_alu_instr *mov = nir_alu_instr_create(state->mem_ctx,
> +                                                      nir_op_imov);
> +            mov->dest.dest.is_ssa = true;
> +            nir_ssa_def_init(&mov->instr, &mov->dest.dest.ssa, 1, NULL);
> +            mov->dest.write_mask = 1;
> +            mov->src[0].src = nir_src_copy(src->src, state->mem_ctx);
  > +            mov->src[0].swizzle[0] = i;
> +
> +            /* Insert at the end of the predecessor but before the jump */
> +            nir_instr *pred_last_instr = nir_block_last_instr(src->pred);
> +            if (pred_last_instr && pred_last_instr->type == nir_instr_type_jump)
> +               nir_instr_insert_before(pred_last_instr, &mov->instr);
> +            else
> +               nir_instr_insert_after_block(src->pred, &mov->instr);

Couldn't you just insert your mov right after the def of the src?  I'm
confused.

> +
> +            nir_phi_src *new_src = ralloc(state->mem_ctx, nir_phi_src);
> +            new_src->pred = src->pred;
> +            new_src->src.is_ssa = true;
> +            new_src->src.ssa = &mov->dest.dest.ssa;

new_src->src = nir_src_for_ssa(&mov->dest.dest.ssa);
(and elsewhere in the file)

> +
> +            exec_list_push_tail(&new_phi->srcs, &new_src->node);
> +         }
> +
> +         nir_instr_insert_before(&phi->instr, &new_phi->instr);
> +      }
> +
> +      nir_instr_insert_after(&last_phi->instr, &vec->instr);
> +
> +      nir_ssa_def_rewrite_uses(&phi->dest.ssa,
> +                               nir_src_for_ssa(&vec->dest.dest.ssa),
> +                               state->mem_ctx);
> +
> +      ralloc_steal(state->dead_ctx, phi);
> +      nir_instr_remove(&phi->instr);
> +
> +      /* We're using the safe iterator and inserting all the newly
> +       * scalarized phi nodes before their non-scalarized version so that's
> +       * ok.  However, we are also inserting vec operations after all of
> +       * the last phi node so once we get here, we can't trust even the
> +       * safe iterator to stop properly.  We have to break manually.
> +       */
> +      if (instr == &last_phi->instr)
> +         break;
> +   }
> +
> +   return true;
> +}
> +
> +static void
> +lower_phis_to_scalar_impl(nir_function_impl *impl)
> +{
> +   struct lower_phis_to_scalar_state state;
> +
> +   state.mem_ctx = ralloc_parent(impl);
> +   state.dead_ctx = ralloc_context(NULL);
> +   state.phi_table = _mesa_hash_table_create(state.dead_ctx, _mesa_hash_pointer,
> +                                             _mesa_key_pointer_equal);
> +
> +   nir_foreach_block(impl, lower_phis_to_scalar_block, &state);
> +
> +   nir_metadata_preserve(impl, nir_metadata_block_index |
> +                               nir_metadata_dominance);
> +
> +   ralloc_free(state.dead_ctx);
> +}

I just notied: When do I need to all nir_metadata_preserve()?  I didn't
in my pass.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 818 bytes
Desc: not available
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20150126/c578af89/attachment.sig>