nir: Add a lowering pass to lower memcpy

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6713>

src/compiler/Makefile.sources

@@ -275,6 +275,7 @@ NIR_FILES = \
 	nir/nir_lower_io_to_scalar.c \
 	nir/nir_lower_io_to_vector.c \
 	nir/nir_lower_mediump_outputs.c \
+	nir/nir_lower_memcpy.c \
 	nir/nir_lower_memory_model.c \
 	nir/nir_lower_non_uniform_access.c \
 	nir/nir_lower_packing.c \

src/compiler/nir/meson.build

@@ -155,6 +155,7 @@ files_libnir = files(
   'nir_lower_io_to_scalar.c',
   'nir_lower_io_to_vector.c',
   'nir_lower_mediump_outputs.c',
+  'nir_lower_memcpy.c',
   'nir_lower_memory_model.c',
   'nir_lower_non_uniform_access.c',
   'nir_lower_packing.c',

src/compiler/nir/nir.h

@@ -4089,6 +4089,8 @@ void nir_lower_deref_copy_instr(struct nir_builder *b,
                                 nir_intrinsic_instr *copy);
 bool nir_lower_var_copies(nir_shader *shader);
 
+bool nir_lower_memcpy(nir_shader *shader);
+
 void nir_fixup_deref_modes(nir_shader *shader);
 
 bool nir_lower_global_vars_to_local(nir_shader *shader);

src/compiler/nir/nir_lower_memcpy.c

@@ -0,0 +1,195 @@
+/*
+ * Copyright © 2020 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.
+ */
+
+#include "nir_builder.h"
+
+#include <string.h>
+
+/** Returns the type to use for a copy of the given size.
+ *
+ * The actual type doesn't matter here all that much as we're just going to do
+ * a load/store on it and never any arithmetic.
+ */
+static const struct glsl_type *
+copy_type_for_byte_size(unsigned size)
+{
+   switch (size) {
+   case 1:  return glsl_vector_type(GLSL_TYPE_UINT8, 1);
+   case 2:  return glsl_vector_type(GLSL_TYPE_UINT16, 1);
+   case 4:  return glsl_vector_type(GLSL_TYPE_UINT, 1);
+   case 8:  return glsl_vector_type(GLSL_TYPE_UINT, 2);
+   case 16: return glsl_vector_type(GLSL_TYPE_UINT, 4);
+   default:
+      unreachable("Unsupported size");
+   }
+}
+
+static nir_ssa_def *
+memcpy_load_deref_elem(nir_builder *b, nir_deref_instr *parent,
+                       nir_ssa_def *index)
+{
+   nir_deref_instr *deref;
+
+   index = nir_i2i(b, index, nir_dest_bit_size(parent->dest));
+   assert(parent->deref_type == nir_deref_type_cast);
+   deref = nir_build_deref_ptr_as_array(b, parent, index);
+
+   return nir_load_deref(b, deref);
+}
+
+static nir_ssa_def *
+memcpy_load_deref_elem_imm(nir_builder *b, nir_deref_instr *parent,
+                           uint64_t index)
+{
+   nir_ssa_def *idx = nir_imm_intN_t(b, index, parent->dest.ssa.bit_size);
+   return memcpy_load_deref_elem(b, parent, idx);
+}
+
+static void
+memcpy_store_deref_elem(nir_builder *b, nir_deref_instr *parent,
+                        nir_ssa_def *index, nir_ssa_def *value)
+{
+   nir_deref_instr *deref;
+
+   index = nir_i2i(b, index, nir_dest_bit_size(parent->dest));
+   assert(parent->deref_type == nir_deref_type_cast);
+   deref = nir_build_deref_ptr_as_array(b, parent, index);
+   nir_store_deref(b, deref, value, 1);
+}
+
+static void
+memcpy_store_deref_elem_imm(nir_builder *b, nir_deref_instr *parent,
+                            uint64_t index, nir_ssa_def *value)
+{
+   nir_ssa_def *idx = nir_imm_intN_t(b, index, parent->dest.ssa.bit_size);
+   memcpy_store_deref_elem(b, parent, idx, value);
+}
+
+static bool
+lower_memcpy_impl(nir_function_impl *impl)
+{
+   nir_builder b;
+   nir_builder_init(&b, impl);
+
+   bool found_const_memcpy = false, found_non_const_memcpy = false;
+
+   nir_foreach_block_safe(block, impl) {
+      nir_foreach_instr_safe(instr, block) {
+         if (instr->type != nir_instr_type_intrinsic)
+            continue;
+
+         nir_intrinsic_instr *cpy = nir_instr_as_intrinsic(instr);
+         if (cpy->intrinsic != nir_intrinsic_memcpy_deref)
+            continue;
+
+         b.cursor = nir_instr_remove(&cpy->instr);
+
+         nir_deref_instr *dst = nir_src_as_deref(cpy->src[0]);
+         nir_deref_instr *src = nir_src_as_deref(cpy->src[1]);
+         if (nir_src_is_const(cpy->src[2])) {
+            found_const_memcpy = true;
+            uint64_t size = nir_src_as_uint(cpy->src[2]);
+            uint64_t offset = 0;
+            while (offset < size) {
+               uint64_t remaining = offset - size;
+               /* For our chunk size, we choose the largest power-of-two that
+                * divides size with a maximum of 16B (a vec4).
+                */
+               unsigned copy_size = 1u << MIN2(ffsll(remaining) - 1, 4);
+               const struct glsl_type *copy_type =
+                  copy_type_for_byte_size(copy_size);
+
+               nir_deref_instr *copy_dst =
+                  nir_build_deref_cast(&b, &dst->dest.ssa, dst->mode,
+                                       copy_type, copy_size);
+               nir_deref_instr *copy_src =
+                  nir_build_deref_cast(&b, &src->dest.ssa, src->mode,
+                                       copy_type, copy_size);
+
+               uint64_t index = offset / copy_size;
+               nir_ssa_def *value =
+                  memcpy_load_deref_elem_imm(&b, copy_src, index);
+               memcpy_store_deref_elem_imm(&b, copy_dst, index, value);
+               offset += copy_size;
+            }
+         } else {
+            found_non_const_memcpy = true;
+            assert(cpy->src[2].is_ssa);
+            nir_ssa_def *size = cpy->src[2].ssa;
+
+            /* In this case, we don't have any idea what the size is so we
+             * emit a loop which copies one byte at a time.
+             */
+            nir_deref_instr *copy_dst =
+               nir_build_deref_cast(&b, &dst->dest.ssa, dst->mode,
+                                    glsl_uint8_t_type(), 1);
+            nir_deref_instr *copy_src =
+               nir_build_deref_cast(&b, &src->dest.ssa, src->mode,
+                                    glsl_uint8_t_type(), 1);
+
+            nir_variable *i = nir_local_variable_create(impl,
+               glsl_uintN_t_type(size->bit_size), NULL);
+            nir_store_var(&b, i, nir_imm_intN_t(&b, 0, size->bit_size), ~0);
+            nir_push_loop(&b);
+            {
+               nir_ssa_def *index = nir_load_var(&b, i);
+               nir_push_if(&b, nir_uge(&b, index, size));
+               {
+                  nir_jump(&b, nir_jump_break);
+               }
+               nir_pop_if(&b, NULL);
+
+               nir_ssa_def *value =
+                  memcpy_load_deref_elem(&b, copy_src, index);
+               memcpy_store_deref_elem(&b, copy_dst, index, value);
+               nir_store_var(&b, i, nir_iadd_imm(&b, index, 1),  ~0);
+            }
+            nir_pop_loop(&b, NULL);
+         }
+      }
+   }
+
+   if (found_non_const_memcpy) {
+      nir_metadata_preserve(impl, nir_metadata_none);
+   } else if (found_const_memcpy) {
+      nir_metadata_preserve(impl, nir_metadata_block_index |
+                                  nir_metadata_dominance);
+   } else {
+      nir_metadata_preserve(impl, nir_metadata_all);
+   }
+
+   return found_const_memcpy || found_non_const_memcpy;
+}
+
+bool
+nir_lower_memcpy(nir_shader *shader)
+{
+   bool progress = false;
+
+   nir_foreach_function(function, shader) {
+      if (function->impl && lower_memcpy_impl(function->impl))
+         progress = true;
+   }
+
+   return progress;
+}