Merge remote-tracking branch 'public/master' into vulkan
authorJason Ekstrand <jason.ekstrand@intel.com>
Thu, 14 Apr 2016 03:25:39 +0000 (20:25 -0700)
committerJason Ekstrand <jason.ekstrand@intel.com>
Thu, 14 Apr 2016 03:25:39 +0000 (20:25 -0700)
14 files changed:
1  2 
configure.ac
src/compiler/Makefile.nir.am
src/compiler/Makefile.sources
src/compiler/nir/glsl_to_nir.cpp
src/compiler/nir/nir.h
src/compiler/nir/nir_lower_io.c
src/compiler/nir/nir_opt_algebraic.py
src/compiler/nir/nir_print.c
src/compiler/nir/spirv/spirv_to_nir.c
src/intel/vulkan/anv_pipeline.c
src/mesa/drivers/dri/i965/Makefile.sources
src/mesa/drivers/dri/i965/brw_fs_nir.cpp
src/mesa/drivers/dri/i965/brw_vec4_nir.cpp
src/mesa/main/mtypes.h

diff --cc configure.ac
Simple merge
index 0000000000000000000000000000000000000000,e1acf8a29a2e4b3433f0aedce1ed0140906a949e..45a9c809bf883c232cfa50ea41acacd0c75619ce
mode 000000,100644..100644
--- /dev/null
@@@ -1,0 -1,93 +1,94 @@@
+ #
+ # Copyright © 2012 Jon TURNEY
+ # Copyright (C) 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.
+ noinst_LTLIBRARIES += nir/libnir.la
+ nir_libnir_la_CPPFLAGS = \
+       $(AM_CPPFLAGS) \
+       -I$(top_builddir)/src/compiler/nir \
+       -I$(top_srcdir)/src/compiler/nir
+ nir_libnir_la_LIBADD = \
+       libcompiler.la
+ nir_libnir_la_SOURCES =                                       \
+       $(NIR_FILES)                                    \
++      $(SPIRV_FILES)                                  \
+       $(NIR_GENERATED_FILES)
+ PYTHON_GEN = $(AM_V_GEN)$(PYTHON2) $(PYTHON_FLAGS)
+ nir/nir_builder_opcodes.h: nir/nir_opcodes.py nir/nir_builder_opcodes_h.py
+       $(MKDIR_GEN)
+       $(PYTHON_GEN) $(srcdir)/nir/nir_builder_opcodes_h.py > $@ || ($(RM) $@; false)
+ nir/nir_constant_expressions.c: nir/nir_opcodes.py nir/nir_constant_expressions.py
+       $(MKDIR_GEN)
+       $(PYTHON_GEN) $(srcdir)/nir/nir_constant_expressions.py > $@ || ($(RM) $@; false)
+ nir/nir_opcodes.h: nir/nir_opcodes.py nir/nir_opcodes_h.py
+       $(MKDIR_GEN)
+       $(PYTHON_GEN) $(srcdir)/nir/nir_opcodes_h.py > $@ || ($(RM) $@; false)
+ nir/nir_opcodes.c: nir/nir_opcodes.py nir/nir_opcodes_c.py
+       $(MKDIR_GEN)
+       $(PYTHON_GEN) $(srcdir)/nir/nir_opcodes_c.py > $@ || ($(RM) $@; false)
+ nir/nir_opt_algebraic.c: nir/nir_opt_algebraic.py nir/nir_algebraic.py
+       $(MKDIR_GEN)
+       $(PYTHON_GEN) $(srcdir)/nir/nir_opt_algebraic.py > $@ || ($(RM) $@; false)
+ check_PROGRAMS += nir/tests/control_flow_tests
+ nir_tests_control_flow_tests_CPPFLAGS = \
+       $(AM_CPPFLAGS) \
+       -I$(top_builddir)/src/compiler/nir \
+       -I$(top_srcdir)/src/compiler/nir
+ nir_tests_control_flow_tests_SOURCES =                        \
+       nir/tests/control_flow_tests.cpp
+ nir_tests_control_flow_tests_CFLAGS =                 \
+       $(PTHREAD_CFLAGS)
+ nir_tests_control_flow_tests_LDADD =                  \
+       $(top_builddir)/src/gtest/libgtest.la           \
+       nir/libnir.la   \
+       $(top_builddir)/src/util/libmesautil.la         \
+       $(PTHREAD_LIBS)
+ TESTS += nir/tests/control_flow_tests
+ BUILT_SOURCES += $(NIR_GENERATED_FILES)
+ CLEANFILES += $(NIR_GENERATED_FILES)
+ EXTRA_DIST += \
+       nir/nir_algebraic.py                            \
+       nir/nir_builder_opcodes_h.py                    \
+       nir/nir_constant_expressions.py                 \
+       nir/nir_opcodes.py                              \
+       nir/nir_opcodes_c.py                            \
+       nir/nir_opcodes_h.py                            \
+       nir/nir_opt_algebraic.py                        \
+       nir/tests
Simple merge
Simple merge
index 8e45cba5a160493b3c41d28f4f7cf1f8330e5a2c,4a66e8b0d3e31fff979167cf70066604d5dc9755..c3a334312399163c5e648b1e8457bc34df0bf44b
@@@ -2214,9 -2191,8 +2215,10 @@@ bool nir_lower_locals_to_regs(nir_shade
  
  void nir_lower_outputs_to_temporaries(nir_shader *shader,
                                        nir_function *entrypoint);
+ void nir_shader_gather_info(nir_shader *shader, nir_function_impl *entrypoint);
  
 +void nir_shader_gather_info(nir_shader *shader, nir_function_impl *entrypoint);
 +
  void nir_assign_var_locations(struct exec_list *var_list,
                                unsigned *size,
                                int (*type_size)(const struct glsl_type *));
Simple merge
Simple merge
Simple merge
index 948454494fa327407bb84b7c0b9dfd551ba0afe8,0000000000000000000000000000000000000000..99514b49650344018a9a65c8c214fdb2d3a3d044
mode 100644,000000..100644
--- /dev/null
@@@ -1,2710 -1,0 +1,2710 @@@
-             nir_load_const_instr_create(b->shader, num_components);
 +/*
 + * 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@jlekstrand.net)
 + *
 + */
 +
 +#include "vtn_private.h"
 +#include "nir/nir_vla.h"
 +#include "nir/nir_control_flow.h"
 +#include "nir/nir_constant_expressions.h"
 +
 +static struct vtn_ssa_value *
 +vtn_undef_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
 +{
 +   struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
 +   val->type = type;
 +
 +   if (glsl_type_is_vector_or_scalar(type)) {
 +      unsigned num_components = glsl_get_vector_elements(val->type);
 +      unsigned bit_size = glsl_get_bit_size(glsl_get_base_type(val->type));
 +      val->def = nir_ssa_undef(&b->nb, num_components, bit_size);
 +   } else {
 +      unsigned elems = glsl_get_length(val->type);
 +      val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
 +      if (glsl_type_is_matrix(type)) {
 +         const struct glsl_type *elem_type =
 +            glsl_vector_type(glsl_get_base_type(type),
 +                             glsl_get_vector_elements(type));
 +
 +         for (unsigned i = 0; i < elems; i++)
 +            val->elems[i] = vtn_undef_ssa_value(b, elem_type);
 +      } else if (glsl_type_is_array(type)) {
 +         const struct glsl_type *elem_type = glsl_get_array_element(type);
 +         for (unsigned i = 0; i < elems; i++)
 +            val->elems[i] = vtn_undef_ssa_value(b, elem_type);
 +      } else {
 +         for (unsigned i = 0; i < elems; i++) {
 +            const struct glsl_type *elem_type = glsl_get_struct_field(type, i);
 +            val->elems[i] = vtn_undef_ssa_value(b, elem_type);
 +         }
 +      }
 +   }
 +
 +   return val;
 +}
 +
 +static struct vtn_ssa_value *
 +vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
 +                    const struct glsl_type *type)
 +{
 +   struct hash_entry *entry = _mesa_hash_table_search(b->const_table, constant);
 +
 +   if (entry)
 +      return entry->data;
 +
 +   struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
 +   val->type = type;
 +
 +   switch (glsl_get_base_type(type)) {
 +   case GLSL_TYPE_INT:
 +   case GLSL_TYPE_UINT:
 +   case GLSL_TYPE_BOOL:
 +   case GLSL_TYPE_FLOAT:
 +   case GLSL_TYPE_DOUBLE:
 +      if (glsl_type_is_vector_or_scalar(type)) {
 +         unsigned num_components = glsl_get_vector_elements(val->type);
 +         nir_load_const_instr *load =
-                nir_load_const_instr_create(b->shader, rows);
++            nir_load_const_instr_create(b->shader, num_components, 32);
 +
 +         for (unsigned i = 0; i < num_components; i++)
 +            load->value.u32[i] = constant->value.u[i];
 +
 +         nir_instr_insert_before_cf_list(&b->impl->body, &load->instr);
 +         val->def = &load->def;
 +      } else {
 +         assert(glsl_type_is_matrix(type));
 +         unsigned rows = glsl_get_vector_elements(val->type);
 +         unsigned columns = glsl_get_matrix_columns(val->type);
 +         val->elems = ralloc_array(b, struct vtn_ssa_value *, columns);
 +
 +         for (unsigned i = 0; i < columns; i++) {
 +            struct vtn_ssa_value *col_val = rzalloc(b, struct vtn_ssa_value);
 +            col_val->type = glsl_get_column_type(val->type);
 +            nir_load_const_instr *load =
++               nir_load_const_instr_create(b->shader, rows, 32);
 +
 +            for (unsigned j = 0; j < rows; j++)
 +               load->value.u32[j] = constant->value.u[rows * i + j];
 +
 +            nir_instr_insert_before_cf_list(&b->impl->body, &load->instr);
 +            col_val->def = &load->def;
 +
 +            val->elems[i] = col_val;
 +         }
 +      }
 +      break;
 +
 +   case GLSL_TYPE_ARRAY: {
 +      unsigned elems = glsl_get_length(val->type);
 +      val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
 +      const struct glsl_type *elem_type = glsl_get_array_element(val->type);
 +      for (unsigned i = 0; i < elems; i++)
 +         val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
 +                                             elem_type);
 +      break;
 +   }
 +
 +   case GLSL_TYPE_STRUCT: {
 +      unsigned elems = glsl_get_length(val->type);
 +      val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
 +      for (unsigned i = 0; i < elems; i++) {
 +         const struct glsl_type *elem_type =
 +            glsl_get_struct_field(val->type, i);
 +         val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
 +                                             elem_type);
 +      }
 +      break;
 +   }
 +
 +   default:
 +      unreachable("bad constant type");
 +   }
 +
 +   return val;
 +}
 +
 +struct vtn_ssa_value *
 +vtn_ssa_value(struct vtn_builder *b, uint32_t value_id)
 +{
 +   struct vtn_value *val = vtn_untyped_value(b, value_id);
 +   switch (val->value_type) {
 +   case vtn_value_type_undef:
 +      return vtn_undef_ssa_value(b, val->type->type);
 +
 +   case vtn_value_type_constant:
 +      return vtn_const_ssa_value(b, val->constant, val->const_type);
 +
 +   case vtn_value_type_ssa:
 +      return val->ssa;
 +
 +   case vtn_value_type_access_chain:
 +      /* This is needed for function parameters */
 +      return vtn_variable_load(b, val->access_chain);
 +
 +   default:
 +      unreachable("Invalid type for an SSA value");
 +   }
 +}
 +
 +static char *
 +vtn_string_literal(struct vtn_builder *b, const uint32_t *words,
 +                   unsigned word_count, unsigned *words_used)
 +{
 +   char *dup = ralloc_strndup(b, (char *)words, word_count * sizeof(*words));
 +   if (words_used) {
 +      /* Ammount of space taken by the string (including the null) */
 +      unsigned len = strlen(dup) + 1;
 +      *words_used = DIV_ROUND_UP(len, sizeof(*words));
 +   }
 +   return dup;
 +}
 +
 +const uint32_t *
 +vtn_foreach_instruction(struct vtn_builder *b, const uint32_t *start,
 +                        const uint32_t *end, vtn_instruction_handler handler)
 +{
 +   b->file = NULL;
 +   b->line = -1;
 +   b->col = -1;
 +
 +   const uint32_t *w = start;
 +   while (w < end) {
 +      SpvOp opcode = w[0] & SpvOpCodeMask;
 +      unsigned count = w[0] >> SpvWordCountShift;
 +      assert(count >= 1 && w + count <= end);
 +
 +      switch (opcode) {
 +      case SpvOpNop:
 +         break; /* Do nothing */
 +
 +      case SpvOpLine:
 +         b->file = vtn_value(b, w[1], vtn_value_type_string)->str;
 +         b->line = w[2];
 +         b->col = w[3];
 +         break;
 +
 +      case SpvOpNoLine:
 +         b->file = NULL;
 +         b->line = -1;
 +         b->col = -1;
 +         break;
 +
 +      default:
 +         if (!handler(b, opcode, w, count))
 +            return w;
 +         break;
 +      }
 +
 +      w += count;
 +   }
 +   assert(w == end);
 +   return w;
 +}
 +
 +static void
 +vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
 +                     const uint32_t *w, unsigned count)
 +{
 +   switch (opcode) {
 +   case SpvOpExtInstImport: {
 +      struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_extension);
 +      if (strcmp((const char *)&w[2], "GLSL.std.450") == 0) {
 +         val->ext_handler = vtn_handle_glsl450_instruction;
 +      } else {
 +         assert(!"Unsupported extension");
 +      }
 +      break;
 +   }
 +
 +   case SpvOpExtInst: {
 +      struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
 +      bool handled = val->ext_handler(b, w[4], w, count);
 +      (void)handled;
 +      assert(handled);
 +      break;
 +   }
 +
 +   default:
 +      unreachable("Unhandled opcode");
 +   }
 +}
 +
 +static void
 +_foreach_decoration_helper(struct vtn_builder *b,
 +                           struct vtn_value *base_value,
 +                           int parent_member,
 +                           struct vtn_value *value,
 +                           vtn_decoration_foreach_cb cb, void *data)
 +{
 +   for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
 +      int member;
 +      if (dec->scope == VTN_DEC_DECORATION) {
 +         member = parent_member;
 +      } else if (dec->scope >= VTN_DEC_STRUCT_MEMBER0) {
 +         assert(parent_member == -1);
 +         member = dec->scope - VTN_DEC_STRUCT_MEMBER0;
 +      } else {
 +         /* Not a decoration */
 +         continue;
 +      }
 +
 +      if (dec->group) {
 +         assert(dec->group->value_type == vtn_value_type_decoration_group);
 +         _foreach_decoration_helper(b, base_value, member, dec->group,
 +                                    cb, data);
 +      } else {
 +         cb(b, base_value, member, dec, data);
 +      }
 +   }
 +}
 +
 +/** Iterates (recursively if needed) over all of the decorations on a value
 + *
 + * This function iterates over all of the decorations applied to a given
 + * value.  If it encounters a decoration group, it recurses into the group
 + * and iterates over all of those decorations as well.
 + */
 +void
 +vtn_foreach_decoration(struct vtn_builder *b, struct vtn_value *value,
 +                       vtn_decoration_foreach_cb cb, void *data)
 +{
 +   _foreach_decoration_helper(b, value, -1, value, cb, data);
 +}
 +
 +void
 +vtn_foreach_execution_mode(struct vtn_builder *b, struct vtn_value *value,
 +                           vtn_execution_mode_foreach_cb cb, void *data)
 +{
 +   for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
 +      if (dec->scope != VTN_DEC_EXECUTION_MODE)
 +         continue;
 +
 +      assert(dec->group == NULL);
 +      cb(b, value, dec, data);
 +   }
 +}
 +
 +static void
 +vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
 +                      const uint32_t *w, unsigned count)
 +{
 +   const uint32_t *w_end = w + count;
 +   const uint32_t target = w[1];
 +   w += 2;
 +
 +   switch (opcode) {
 +   case SpvOpDecorationGroup:
 +      vtn_push_value(b, target, vtn_value_type_decoration_group);
 +      break;
 +
 +   case SpvOpDecorate:
 +   case SpvOpMemberDecorate:
 +   case SpvOpExecutionMode: {
 +      struct vtn_value *val = &b->values[target];
 +
 +      struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
 +      switch (opcode) {
 +      case SpvOpDecorate:
 +         dec->scope = VTN_DEC_DECORATION;
 +         break;
 +      case SpvOpMemberDecorate:
 +         dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(w++);
 +         break;
 +      case SpvOpExecutionMode:
 +         dec->scope = VTN_DEC_EXECUTION_MODE;
 +         break;
 +      default:
 +         unreachable("Invalid decoration opcode");
 +      }
 +      dec->decoration = *(w++);
 +      dec->literals = w;
 +
 +      /* Link into the list */
 +      dec->next = val->decoration;
 +      val->decoration = dec;
 +      break;
 +   }
 +
 +   case SpvOpGroupMemberDecorate:
 +   case SpvOpGroupDecorate: {
 +      struct vtn_value *group =
 +         vtn_value(b, target, vtn_value_type_decoration_group);
 +
 +      for (; w < w_end; w++) {
 +         struct vtn_value *val = vtn_untyped_value(b, *w);
 +         struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
 +
 +         dec->group = group;
 +         if (opcode == SpvOpGroupDecorate) {
 +            dec->scope = VTN_DEC_DECORATION;
 +         } else {
 +            dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(++w);
 +         }
 +
 +         /* Link into the list */
 +         dec->next = val->decoration;
 +         val->decoration = dec;
 +      }
 +      break;
 +   }
 +
 +   default:
 +      unreachable("Unhandled opcode");
 +   }
 +}
 +
 +struct member_decoration_ctx {
 +   unsigned num_fields;
 +   struct glsl_struct_field *fields;
 +   struct vtn_type *type;
 +};
 +
 +/* does a shallow copy of a vtn_type */
 +
 +static struct vtn_type *
 +vtn_type_copy(struct vtn_builder *b, struct vtn_type *src)
 +{
 +   struct vtn_type *dest = ralloc(b, struct vtn_type);
 +   dest->type = src->type;
 +   dest->is_builtin = src->is_builtin;
 +   if (src->is_builtin)
 +      dest->builtin = src->builtin;
 +
 +   if (!glsl_type_is_scalar(src->type)) {
 +      switch (glsl_get_base_type(src->type)) {
 +      case GLSL_TYPE_INT:
 +      case GLSL_TYPE_UINT:
 +      case GLSL_TYPE_BOOL:
 +      case GLSL_TYPE_FLOAT:
 +      case GLSL_TYPE_DOUBLE:
 +      case GLSL_TYPE_ARRAY:
 +         dest->row_major = src->row_major;
 +         dest->stride = src->stride;
 +         dest->array_element = src->array_element;
 +         break;
 +
 +      case GLSL_TYPE_STRUCT: {
 +         unsigned elems = glsl_get_length(src->type);
 +
 +         dest->members = ralloc_array(b, struct vtn_type *, elems);
 +         memcpy(dest->members, src->members, elems * sizeof(struct vtn_type *));
 +
 +         dest->offsets = ralloc_array(b, unsigned, elems);
 +         memcpy(dest->offsets, src->offsets, elems * sizeof(unsigned));
 +         break;
 +      }
 +
 +      default:
 +         unreachable("unhandled type");
 +      }
 +   }
 +
 +   return dest;
 +}
 +
 +static struct vtn_type *
 +mutable_matrix_member(struct vtn_builder *b, struct vtn_type *type, int member)
 +{
 +   type->members[member] = vtn_type_copy(b, type->members[member]);
 +   type = type->members[member];
 +
 +   /* We may have an array of matrices.... Oh, joy! */
 +   while (glsl_type_is_array(type->type)) {
 +      type->array_element = vtn_type_copy(b, type->array_element);
 +      type = type->array_element;
 +   }
 +
 +   assert(glsl_type_is_matrix(type->type));
 +
 +   return type;
 +}
 +
 +static void
 +struct_member_decoration_cb(struct vtn_builder *b,
 +                            struct vtn_value *val, int member,
 +                            const struct vtn_decoration *dec, void *void_ctx)
 +{
 +   struct member_decoration_ctx *ctx = void_ctx;
 +
 +   if (member < 0)
 +      return;
 +
 +   assert(member < ctx->num_fields);
 +
 +   switch (dec->decoration) {
 +   case SpvDecorationRelaxedPrecision:
 +      break; /* FIXME: Do nothing with this for now. */
 +   case SpvDecorationNoPerspective:
 +      ctx->fields[member].interpolation = INTERP_QUALIFIER_NOPERSPECTIVE;
 +      break;
 +   case SpvDecorationFlat:
 +      ctx->fields[member].interpolation = INTERP_QUALIFIER_FLAT;
 +      break;
 +   case SpvDecorationCentroid:
 +      ctx->fields[member].centroid = true;
 +      break;
 +   case SpvDecorationSample:
 +      ctx->fields[member].sample = true;
 +      break;
 +   case SpvDecorationLocation:
 +      ctx->fields[member].location = dec->literals[0];
 +      break;
 +   case SpvDecorationBuiltIn:
 +      ctx->type->members[member] = vtn_type_copy(b, ctx->type->members[member]);
 +      ctx->type->members[member]->is_builtin = true;
 +      ctx->type->members[member]->builtin = dec->literals[0];
 +      ctx->type->builtin_block = true;
 +      break;
 +   case SpvDecorationOffset:
 +      ctx->type->offsets[member] = dec->literals[0];
 +      break;
 +   case SpvDecorationMatrixStride:
 +      mutable_matrix_member(b, ctx->type, member)->stride = dec->literals[0];
 +      break;
 +   case SpvDecorationColMajor:
 +      break; /* Nothing to do here.  Column-major is the default. */
 +   case SpvDecorationRowMajor:
 +      mutable_matrix_member(b, ctx->type, member)->row_major = true;
 +      break;
 +   default:
 +      unreachable("Unhandled member decoration");
 +   }
 +}
 +
 +static void
 +type_decoration_cb(struct vtn_builder *b,
 +                   struct vtn_value *val, int member,
 +                    const struct vtn_decoration *dec, void *ctx)
 +{
 +   struct vtn_type *type = val->type;
 +
 +   if (member != -1)
 +      return;
 +
 +   switch (dec->decoration) {
 +   case SpvDecorationArrayStride:
 +      type->stride = dec->literals[0];
 +      break;
 +   case SpvDecorationBlock:
 +      type->block = true;
 +      break;
 +   case SpvDecorationBufferBlock:
 +      type->buffer_block = true;
 +      break;
 +   case SpvDecorationGLSLShared:
 +   case SpvDecorationGLSLPacked:
 +      /* Ignore these, since we get explicit offsets anyways */
 +      break;
 +
 +   case SpvDecorationStream:
 +      assert(dec->literals[0] == 0);
 +      break;
 +
 +   default:
 +      unreachable("Unhandled type decoration");
 +   }
 +}
 +
 +static unsigned
 +translate_image_format(SpvImageFormat format)
 +{
 +   switch (format) {
 +   case SpvImageFormatUnknown:      return 0;      /* GL_NONE */
 +   case SpvImageFormatRgba32f:      return 0x8814; /* GL_RGBA32F */
 +   case SpvImageFormatRgba16f:      return 0x881A; /* GL_RGBA16F */
 +   case SpvImageFormatR32f:         return 0x822E; /* GL_R32F */
 +   case SpvImageFormatRgba8:        return 0x8058; /* GL_RGBA8 */
 +   case SpvImageFormatRgba8Snorm:   return 0x8F97; /* GL_RGBA8_SNORM */
 +   case SpvImageFormatRg32f:        return 0x8230; /* GL_RG32F */
 +   case SpvImageFormatRg16f:        return 0x822F; /* GL_RG16F */
 +   case SpvImageFormatR11fG11fB10f: return 0x8C3A; /* GL_R11F_G11F_B10F */
 +   case SpvImageFormatR16f:         return 0x822D; /* GL_R16F */
 +   case SpvImageFormatRgba16:       return 0x805B; /* GL_RGBA16 */
 +   case SpvImageFormatRgb10A2:      return 0x8059; /* GL_RGB10_A2 */
 +   case SpvImageFormatRg16:         return 0x822C; /* GL_RG16 */
 +   case SpvImageFormatRg8:          return 0x822B; /* GL_RG8 */
 +   case SpvImageFormatR16:          return 0x822A; /* GL_R16 */
 +   case SpvImageFormatR8:           return 0x8229; /* GL_R8 */
 +   case SpvImageFormatRgba16Snorm:  return 0x8F9B; /* GL_RGBA16_SNORM */
 +   case SpvImageFormatRg16Snorm:    return 0x8F99; /* GL_RG16_SNORM */
 +   case SpvImageFormatRg8Snorm:     return 0x8F95; /* GL_RG8_SNORM */
 +   case SpvImageFormatR16Snorm:     return 0x8F98; /* GL_R16_SNORM */
 +   case SpvImageFormatR8Snorm:      return 0x8F94; /* GL_R8_SNORM */
 +   case SpvImageFormatRgba32i:      return 0x8D82; /* GL_RGBA32I */
 +   case SpvImageFormatRgba16i:      return 0x8D88; /* GL_RGBA16I */
 +   case SpvImageFormatRgba8i:       return 0x8D8E; /* GL_RGBA8I */
 +   case SpvImageFormatR32i:         return 0x8235; /* GL_R32I */
 +   case SpvImageFormatRg32i:        return 0x823B; /* GL_RG32I */
 +   case SpvImageFormatRg16i:        return 0x8239; /* GL_RG16I */
 +   case SpvImageFormatRg8i:         return 0x8237; /* GL_RG8I */
 +   case SpvImageFormatR16i:         return 0x8233; /* GL_R16I */
 +   case SpvImageFormatR8i:          return 0x8231; /* GL_R8I */
 +   case SpvImageFormatRgba32ui:     return 0x8D70; /* GL_RGBA32UI */
 +   case SpvImageFormatRgba16ui:     return 0x8D76; /* GL_RGBA16UI */
 +   case SpvImageFormatRgba8ui:      return 0x8D7C; /* GL_RGBA8UI */
 +   case SpvImageFormatR32ui:        return 0x8236; /* GL_R32UI */
 +   case SpvImageFormatRgb10a2ui:    return 0x906F; /* GL_RGB10_A2UI */
 +   case SpvImageFormatRg32ui:       return 0x823C; /* GL_RG32UI */
 +   case SpvImageFormatRg16ui:       return 0x823A; /* GL_RG16UI */
 +   case SpvImageFormatRg8ui:        return 0x8238; /* GL_RG8UI */
 +   case SpvImageFormatR16ui:        return 0x823A; /* GL_RG16UI */
 +   case SpvImageFormatR8ui:         return 0x8232; /* GL_R8UI */
 +   default:
 +      assert(!"Invalid image format");
 +      return 0;
 +   }
 +}
 +
 +static void
 +vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
 +                const uint32_t *w, unsigned count)
 +{
 +   struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_type);
 +
 +   val->type = rzalloc(b, struct vtn_type);
 +   val->type->is_builtin = false;
 +   val->type->val = val;
 +
 +   switch (opcode) {
 +   case SpvOpTypeVoid:
 +      val->type->type = glsl_void_type();
 +      break;
 +   case SpvOpTypeBool:
 +      val->type->type = glsl_bool_type();
 +      break;
 +   case SpvOpTypeInt: {
 +      const bool signedness = w[3];
 +      val->type->type = (signedness ? glsl_int_type() : glsl_uint_type());
 +      break;
 +   }
 +   case SpvOpTypeFloat:
 +      val->type->type = glsl_float_type();
 +      break;
 +
 +   case SpvOpTypeVector: {
 +      struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
 +      unsigned elems = w[3];
 +
 +      assert(glsl_type_is_scalar(base->type));
 +      val->type->type = glsl_vector_type(glsl_get_base_type(base->type), elems);
 +
 +      /* Vectors implicitly have sizeof(base_type) stride.  For now, this
 +       * is always 4 bytes.  This will have to change if we want to start
 +       * supporting doubles or half-floats.
 +       */
 +      val->type->stride = 4;
 +      val->type->array_element = base;
 +      break;
 +   }
 +
 +   case SpvOpTypeMatrix: {
 +      struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type;
 +      unsigned columns = w[3];
 +
 +      assert(glsl_type_is_vector(base->type));
 +      val->type->type = glsl_matrix_type(glsl_get_base_type(base->type),
 +                                         glsl_get_vector_elements(base->type),
 +                                         columns);
 +      assert(!glsl_type_is_error(val->type->type));
 +      val->type->array_element = base;
 +      val->type->row_major = false;
 +      val->type->stride = 0;
 +      break;
 +   }
 +
 +   case SpvOpTypeRuntimeArray:
 +   case SpvOpTypeArray: {
 +      struct vtn_type *array_element =
 +         vtn_value(b, w[2], vtn_value_type_type)->type;
 +
 +      unsigned length;
 +      if (opcode == SpvOpTypeRuntimeArray) {
 +         /* A length of 0 is used to denote unsized arrays */
 +         length = 0;
 +      } else {
 +         length =
 +            vtn_value(b, w[3], vtn_value_type_constant)->constant->value.u[0];
 +      }
 +
 +      val->type->type = glsl_array_type(array_element->type, length);
 +      val->type->array_element = array_element;
 +      val->type->stride = 0;
 +      break;
 +   }
 +
 +   case SpvOpTypeStruct: {
 +      unsigned num_fields = count - 2;
 +      val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
 +      val->type->offsets = ralloc_array(b, unsigned, num_fields);
 +
 +      NIR_VLA(struct glsl_struct_field, fields, count);
 +      for (unsigned i = 0; i < num_fields; i++) {
 +         val->type->members[i] =
 +            vtn_value(b, w[i + 2], vtn_value_type_type)->type;
 +         fields[i] = (struct glsl_struct_field) {
 +            .type = val->type->members[i]->type,
 +            .name = ralloc_asprintf(b, "field%d", i),
 +            .location = -1,
 +         };
 +      }
 +
 +      struct member_decoration_ctx ctx = {
 +         .num_fields = num_fields,
 +         .fields = fields,
 +         .type = val->type
 +      };
 +
 +      vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
 +
 +      const char *name = val->name ? val->name : "struct";
 +
 +      val->type->type = glsl_struct_type(fields, num_fields, name);
 +      break;
 +   }
 +
 +   case SpvOpTypeFunction: {
 +      const struct glsl_type *return_type =
 +         vtn_value(b, w[2], vtn_value_type_type)->type->type;
 +      NIR_VLA(struct glsl_function_param, params, count - 3);
 +      for (unsigned i = 0; i < count - 3; i++) {
 +         params[i].type = vtn_value(b, w[i + 3], vtn_value_type_type)->type->type;
 +
 +         /* FIXME: */
 +         params[i].in = true;
 +         params[i].out = true;
 +      }
 +      val->type->type = glsl_function_type(return_type, params, count - 3);
 +      break;
 +   }
 +
 +   case SpvOpTypePointer:
 +      /* FIXME:  For now, we'll just do the really lame thing and return
 +       * the same type.  The validator should ensure that the proper number
 +       * of dereferences happen
 +       */
 +      val->type = vtn_value(b, w[3], vtn_value_type_type)->type;
 +      break;
 +
 +   case SpvOpTypeImage: {
 +      const struct glsl_type *sampled_type =
 +         vtn_value(b, w[2], vtn_value_type_type)->type->type;
 +
 +      assert(glsl_type_is_vector_or_scalar(sampled_type));
 +
 +      enum glsl_sampler_dim dim;
 +      switch ((SpvDim)w[3]) {
 +      case SpvDim1D:       dim = GLSL_SAMPLER_DIM_1D;    break;
 +      case SpvDim2D:       dim = GLSL_SAMPLER_DIM_2D;    break;
 +      case SpvDim3D:       dim = GLSL_SAMPLER_DIM_3D;    break;
 +      case SpvDimCube:     dim = GLSL_SAMPLER_DIM_CUBE;  break;
 +      case SpvDimRect:     dim = GLSL_SAMPLER_DIM_RECT;  break;
 +      case SpvDimBuffer:   dim = GLSL_SAMPLER_DIM_BUF;   break;
 +      default:
 +         unreachable("Invalid SPIR-V Sampler dimension");
 +      }
 +
 +      bool is_shadow = w[4];
 +      bool is_array = w[5];
 +      bool multisampled = w[6];
 +      unsigned sampled = w[7];
 +      SpvImageFormat format = w[8];
 +
 +      if (count > 9)
 +         val->type->access_qualifier = w[9];
 +      else
 +         val->type->access_qualifier = SpvAccessQualifierReadWrite;
 +
 +      if (multisampled) {
 +         assert(dim == GLSL_SAMPLER_DIM_2D);
 +         dim = GLSL_SAMPLER_DIM_MS;
 +      }
 +
 +      val->type->image_format = translate_image_format(format);
 +
 +      if (sampled == 1) {
 +         val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
 +                                             glsl_get_base_type(sampled_type));
 +      } else if (sampled == 2) {
 +         assert(format);
 +         assert(!is_shadow);
 +         val->type->type = glsl_image_type(dim, is_array,
 +                                           glsl_get_base_type(sampled_type));
 +      } else {
 +         assert(!"We need to know if the image will be sampled");
 +      }
 +      break;
 +   }
 +
 +   case SpvOpTypeSampledImage:
 +      val->type = vtn_value(b, w[2], vtn_value_type_type)->type;
 +      break;
 +
 +   case SpvOpTypeSampler:
 +      /* The actual sampler type here doesn't really matter.  It gets
 +       * thrown away the moment you combine it with an image.  What really
 +       * matters is that it's a sampler type as opposed to an integer type
 +       * so the backend knows what to do.
 +       */
 +      val->type->type = glsl_bare_sampler_type();
 +      break;
 +
 +   case SpvOpTypeOpaque:
 +   case SpvOpTypeEvent:
 +   case SpvOpTypeDeviceEvent:
 +   case SpvOpTypeReserveId:
 +   case SpvOpTypeQueue:
 +   case SpvOpTypePipe:
 +   default:
 +      unreachable("Unhandled opcode");
 +   }
 +
 +   vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
 +}
 +
 +static nir_constant *
 +vtn_null_constant(struct vtn_builder *b, const struct glsl_type *type)
 +{
 +   nir_constant *c = rzalloc(b, nir_constant);
 +
 +   switch (glsl_get_base_type(type)) {
 +   case GLSL_TYPE_INT:
 +   case GLSL_TYPE_UINT:
 +   case GLSL_TYPE_BOOL:
 +   case GLSL_TYPE_FLOAT:
 +   case GLSL_TYPE_DOUBLE:
 +      /* Nothing to do here.  It's already initialized to zero */
 +      break;
 +
 +   case GLSL_TYPE_ARRAY:
 +      assert(glsl_get_length(type) > 0);
 +      c->num_elements = glsl_get_length(type);
 +      c->elements = ralloc_array(b, nir_constant *, c->num_elements);
 +
 +      c->elements[0] = vtn_null_constant(b, glsl_get_array_element(type));
 +      for (unsigned i = 1; i < c->num_elements; i++)
 +         c->elements[i] = c->elements[0];
 +      break;
 +
 +   case GLSL_TYPE_STRUCT:
 +      c->num_elements = glsl_get_length(type);
 +      c->elements = ralloc_array(b, nir_constant *, c->num_elements);
 +
 +      for (unsigned i = 0; i < c->num_elements; i++) {
 +         c->elements[i] = vtn_null_constant(b, glsl_get_struct_field(type, i));
 +      }
 +      break;
 +
 +   default:
 +      unreachable("Invalid type for null constant");
 +   }
 +
 +   return c;
 +}
 +
 +static void
 +spec_constant_deocoration_cb(struct vtn_builder *b, struct vtn_value *v,
 +                             int member, const struct vtn_decoration *dec,
 +                             void *data)
 +{
 +   assert(member == -1);
 +   if (dec->decoration != SpvDecorationSpecId)
 +      return;
 +
 +   uint32_t *const_value = data;
 +
 +   for (unsigned i = 0; i < b->num_specializations; i++) {
 +      if (b->specializations[i].id == dec->literals[0]) {
 +         *const_value = b->specializations[i].data;
 +         return;
 +      }
 +   }
 +}
 +
 +static uint32_t
 +get_specialization(struct vtn_builder *b, struct vtn_value *val,
 +                   uint32_t const_value)
 +{
 +   vtn_foreach_decoration(b, val, spec_constant_deocoration_cb, &const_value);
 +   return const_value;
 +}
 +
 +static void
 +vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
 +                    const uint32_t *w, unsigned count)
 +{
 +   struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_constant);
 +   val->const_type = vtn_value(b, w[1], vtn_value_type_type)->type->type;
 +   val->constant = rzalloc(b, nir_constant);
 +   switch (opcode) {
 +   case SpvOpConstantTrue:
 +      assert(val->const_type == glsl_bool_type());
 +      val->constant->value.u[0] = NIR_TRUE;
 +      break;
 +   case SpvOpConstantFalse:
 +      assert(val->const_type == glsl_bool_type());
 +      val->constant->value.u[0] = NIR_FALSE;
 +      break;
 +
 +   case SpvOpSpecConstantTrue:
 +   case SpvOpSpecConstantFalse: {
 +      assert(val->const_type == glsl_bool_type());
 +      uint32_t int_val =
 +         get_specialization(b, val, (opcode == SpvOpSpecConstantTrue));
 +      val->constant->value.u[0] = int_val ? NIR_TRUE : NIR_FALSE;
 +      break;
 +   }
 +
 +   case SpvOpConstant:
 +      assert(glsl_type_is_scalar(val->const_type));
 +      val->constant->value.u[0] = w[3];
 +      break;
 +   case SpvOpSpecConstant:
 +      assert(glsl_type_is_scalar(val->const_type));
 +      val->constant->value.u[0] = get_specialization(b, val, w[3]);
 +      break;
 +   case SpvOpSpecConstantComposite:
 +   case SpvOpConstantComposite: {
 +      unsigned elem_count = count - 3;
 +      nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
 +      for (unsigned i = 0; i < elem_count; i++)
 +         elems[i] = vtn_value(b, w[i + 3], vtn_value_type_constant)->constant;
 +
 +      switch (glsl_get_base_type(val->const_type)) {
 +      case GLSL_TYPE_UINT:
 +      case GLSL_TYPE_INT:
 +      case GLSL_TYPE_FLOAT:
 +      case GLSL_TYPE_BOOL:
 +         if (glsl_type_is_matrix(val->const_type)) {
 +            unsigned rows = glsl_get_vector_elements(val->const_type);
 +            assert(glsl_get_matrix_columns(val->const_type) == elem_count);
 +            for (unsigned i = 0; i < elem_count; i++)
 +               for (unsigned j = 0; j < rows; j++)
 +                  val->constant->value.u[rows * i + j] = elems[i]->value.u[j];
 +         } else {
 +            assert(glsl_type_is_vector(val->const_type));
 +            assert(glsl_get_vector_elements(val->const_type) == elem_count);
 +            for (unsigned i = 0; i < elem_count; i++)
 +               val->constant->value.u[i] = elems[i]->value.u[0];
 +         }
 +         ralloc_free(elems);
 +         break;
 +
 +      case GLSL_TYPE_STRUCT:
 +      case GLSL_TYPE_ARRAY:
 +         ralloc_steal(val->constant, elems);
 +         val->constant->num_elements = elem_count;
 +         val->constant->elements = elems;
 +         break;
 +
 +      default:
 +         unreachable("Unsupported type for constants");
 +      }
 +      break;
 +   }
 +
 +   case SpvOpSpecConstantOp: {
 +      SpvOp opcode = get_specialization(b, val, w[3]);
 +      switch (opcode) {
 +      case SpvOpVectorShuffle: {
 +         struct vtn_value *v0 = vtn_value(b, w[4], vtn_value_type_constant);
 +         struct vtn_value *v1 = vtn_value(b, w[5], vtn_value_type_constant);
 +         unsigned len0 = glsl_get_vector_elements(v0->const_type);
 +         unsigned len1 = glsl_get_vector_elements(v1->const_type);
 +
 +         uint32_t u[8];
 +         for (unsigned i = 0; i < len0; i++)
 +            u[i] = v0->constant->value.u[i];
 +         for (unsigned i = 0; i < len1; i++)
 +            u[len0 + i] = v1->constant->value.u[i];
 +
 +         for (unsigned i = 0; i < count - 6; i++) {
 +            uint32_t comp = w[i + 6];
 +            if (comp == (uint32_t)-1) {
 +               val->constant->value.u[i] = 0xdeadbeef;
 +            } else {
 +               val->constant->value.u[i] = u[comp];
 +            }
 +         }
 +         return;
 +      }
 +
 +      case SpvOpCompositeExtract:
 +      case SpvOpCompositeInsert: {
 +         struct vtn_value *comp;
 +         unsigned deref_start;
 +         struct nir_constant **c;
 +         if (opcode == SpvOpCompositeExtract) {
 +            comp = vtn_value(b, w[4], vtn_value_type_constant);
 +            deref_start = 5;
 +            c = &comp->constant;
 +         } else {
 +            comp = vtn_value(b, w[5], vtn_value_type_constant);
 +            deref_start = 6;
 +            val->constant = nir_constant_clone(comp->constant,
 +                                               (nir_variable *)b);
 +            c = &val->constant;
 +         }
 +
 +         int elem = -1;
 +         const struct glsl_type *type = comp->const_type;
 +         for (unsigned i = deref_start; i < count; i++) {
 +            switch (glsl_get_base_type(type)) {
 +            case GLSL_TYPE_UINT:
 +            case GLSL_TYPE_INT:
 +            case GLSL_TYPE_FLOAT:
 +            case GLSL_TYPE_BOOL:
 +               /* If we hit this granularity, we're picking off an element */
 +               if (elem < 0)
 +                  elem = 0;
 +
 +               if (glsl_type_is_matrix(type)) {
 +                  elem += w[i] * glsl_get_vector_elements(type);
 +                  type = glsl_get_column_type(type);
 +               } else {
 +                  assert(glsl_type_is_vector(type));
 +                  elem += w[i];
 +                  type = glsl_scalar_type(glsl_get_base_type(type));
 +               }
 +               continue;
 +
 +            case GLSL_TYPE_ARRAY:
 +               c = &(*c)->elements[w[i]];
 +               type = glsl_get_array_element(type);
 +               continue;
 +
 +            case GLSL_TYPE_STRUCT:
 +               c = &(*c)->elements[w[i]];
 +               type = glsl_get_struct_field(type, w[i]);
 +               continue;
 +
 +            default:
 +               unreachable("Invalid constant type");
 +            }
 +         }
 +
 +         if (opcode == SpvOpCompositeExtract) {
 +            if (elem == -1) {
 +               val->constant = *c;
 +            } else {
 +               unsigned num_components = glsl_get_vector_elements(type);
 +               for (unsigned i = 0; i < num_components; i++)
 +                  val->constant->value.u[i] = (*c)->value.u[elem + i];
 +            }
 +         } else {
 +            struct vtn_value *insert =
 +               vtn_value(b, w[4], vtn_value_type_constant);
 +            assert(insert->const_type == type);
 +            if (elem == -1) {
 +               *c = insert->constant;
 +            } else {
 +               unsigned num_components = glsl_get_vector_elements(type);
 +               for (unsigned i = 0; i < num_components; i++)
 +                  (*c)->value.u[elem + i] = insert->constant->value.u[i];
 +            }
 +         }
 +         return;
 +      }
 +
 +      default: {
 +         bool swap;
 +         nir_op op = vtn_nir_alu_op_for_spirv_opcode(opcode, &swap);
 +
 +         unsigned num_components = glsl_get_vector_elements(val->const_type);
 +         unsigned bit_size =
 +            glsl_get_bit_size(glsl_get_base_type(val->const_type));
 +
 +         nir_const_value src[3];
 +         assert(count <= 7);
 +         for (unsigned i = 0; i < count - 4; i++) {
 +            nir_constant *c =
 +               vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
 +
 +            unsigned j = swap ? 1 - i : i;
 +            assert(bit_size == 32);
 +            for (unsigned k = 0; k < num_components; k++)
 +               src[j].u32[k] = c->value.u[k];
 +         }
 +
 +         nir_const_value res = nir_eval_const_opcode(op, num_components,
 +                                                     bit_size, src);
 +
 +         for (unsigned k = 0; k < num_components; k++)
 +            val->constant->value.u[k] = res.u32[k];
 +
 +         return;
 +      } /* default */
 +      }
 +   }
 +
 +   case SpvOpConstantNull:
 +      val->constant = vtn_null_constant(b, val->const_type);
 +      break;
 +
 +   case SpvOpConstantSampler:
 +      assert(!"OpConstantSampler requires Kernel Capability");
 +      break;
 +
 +   default:
 +      unreachable("Unhandled opcode");
 +   }
 +}
 +
 +static void
 +vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
 +                         const uint32_t *w, unsigned count)
 +{
 +   struct nir_function *callee =
 +      vtn_value(b, w[3], vtn_value_type_function)->func->impl->function;
 +
 +   nir_call_instr *call = nir_call_instr_create(b->nb.shader, callee);
 +   for (unsigned i = 0; i < call->num_params; i++) {
 +      unsigned arg_id = w[4 + i];
 +      struct vtn_value *arg = vtn_untyped_value(b, arg_id);
 +      if (arg->value_type == vtn_value_type_access_chain) {
 +         nir_deref_var *d = vtn_access_chain_to_deref(b, arg->access_chain);
 +         call->params[i] = nir_deref_as_var(nir_copy_deref(call, &d->deref));
 +      } else {
 +         struct vtn_ssa_value *arg_ssa = vtn_ssa_value(b, arg_id);
 +
 +         /* Make a temporary to store the argument in */
 +         nir_variable *tmp =
 +            nir_local_variable_create(b->impl, arg_ssa->type, "arg_tmp");
 +         call->params[i] = nir_deref_var_create(call, tmp);
 +
 +         vtn_local_store(b, arg_ssa, call->params[i]);
 +      }
 +   }
 +
 +   nir_variable *out_tmp = NULL;
 +   if (!glsl_type_is_void(callee->return_type)) {
 +      out_tmp = nir_local_variable_create(b->impl, callee->return_type,
 +                                          "out_tmp");
 +      call->return_deref = nir_deref_var_create(call, out_tmp);
 +   }
 +
 +   nir_builder_instr_insert(&b->nb, &call->instr);
 +
 +   if (glsl_type_is_void(callee->return_type)) {
 +      vtn_push_value(b, w[2], vtn_value_type_undef);
 +   } else {
 +      struct vtn_value *retval = vtn_push_value(b, w[2], vtn_value_type_ssa);
 +      retval->ssa = vtn_local_load(b, call->return_deref);
 +   }
 +}
 +
 +struct vtn_ssa_value *
 +vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
 +{
 +   struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
 +   val->type = type;
 +
 +   if (!glsl_type_is_vector_or_scalar(type)) {
 +      unsigned elems = glsl_get_length(type);
 +      val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
 +      for (unsigned i = 0; i < elems; i++) {
 +         const struct glsl_type *child_type;
 +
 +         switch (glsl_get_base_type(type)) {
 +         case GLSL_TYPE_INT:
 +         case GLSL_TYPE_UINT:
 +         case GLSL_TYPE_BOOL:
 +         case GLSL_TYPE_FLOAT:
 +         case GLSL_TYPE_DOUBLE:
 +            child_type = glsl_get_column_type(type);
 +            break;
 +         case GLSL_TYPE_ARRAY:
 +            child_type = glsl_get_array_element(type);
 +            break;
 +         case GLSL_TYPE_STRUCT:
 +            child_type = glsl_get_struct_field(type, i);
 +            break;
 +         default:
 +            unreachable("unkown base type");
 +         }
 +
 +         val->elems[i] = vtn_create_ssa_value(b, child_type);
 +      }
 +   }
 +
 +   return val;
 +}
 +
 +static nir_tex_src
 +vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
 +{
 +   nir_tex_src src;
 +   src.src = nir_src_for_ssa(vtn_ssa_value(b, index)->def);
 +   src.src_type = type;
 +   return src;
 +}
 +
 +static void
 +vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
 +                   const uint32_t *w, unsigned count)
 +{
 +   if (opcode == SpvOpSampledImage) {
 +      struct vtn_value *val =
 +         vtn_push_value(b, w[2], vtn_value_type_sampled_image);
 +      val->sampled_image = ralloc(b, struct vtn_sampled_image);
 +      val->sampled_image->image =
 +         vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
 +      val->sampled_image->sampler =
 +         vtn_value(b, w[4], vtn_value_type_access_chain)->access_chain;
 +      return;
 +   } else if (opcode == SpvOpImage) {
 +      struct vtn_value *val =
 +         vtn_push_value(b, w[2], vtn_value_type_access_chain);
 +      struct vtn_value *src_val = vtn_untyped_value(b, w[3]);
 +      if (src_val->value_type == vtn_value_type_sampled_image) {
 +         val->access_chain = src_val->sampled_image->image;
 +      } else {
 +         assert(src_val->value_type == vtn_value_type_access_chain);
 +         val->access_chain = src_val->access_chain;
 +      }
 +      return;
 +   }
 +
 +   struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
 +   struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
 +
 +   struct vtn_sampled_image sampled;
 +   struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
 +   if (sampled_val->value_type == vtn_value_type_sampled_image) {
 +      sampled = *sampled_val->sampled_image;
 +   } else {
 +      assert(sampled_val->value_type == vtn_value_type_access_chain);
 +      sampled.image = NULL;
 +      sampled.sampler = sampled_val->access_chain;
 +   }
 +
 +   const struct glsl_type *image_type;
 +   if (sampled.image) {
 +      image_type = sampled.image->var->var->interface_type;
 +   } else {
 +      image_type = sampled.sampler->var->var->interface_type;
 +   }
 +
 +   nir_tex_src srcs[8]; /* 8 should be enough */
 +   nir_tex_src *p = srcs;
 +
 +   unsigned idx = 4;
 +
 +   bool has_coord = false;
 +   switch (opcode) {
 +   case SpvOpImageSampleImplicitLod:
 +   case SpvOpImageSampleExplicitLod:
 +   case SpvOpImageSampleDrefImplicitLod:
 +   case SpvOpImageSampleDrefExplicitLod:
 +   case SpvOpImageSampleProjImplicitLod:
 +   case SpvOpImageSampleProjExplicitLod:
 +   case SpvOpImageSampleProjDrefImplicitLod:
 +   case SpvOpImageSampleProjDrefExplicitLod:
 +   case SpvOpImageFetch:
 +   case SpvOpImageGather:
 +   case SpvOpImageDrefGather:
 +   case SpvOpImageQueryLod: {
 +      /* All these types have the coordinate as their first real argument */
 +      struct vtn_ssa_value *coord = vtn_ssa_value(b, w[idx++]);
 +      has_coord = true;
 +      p->src = nir_src_for_ssa(coord->def);
 +      p->src_type = nir_tex_src_coord;
 +      p++;
 +      break;
 +   }
 +
 +   default:
 +      break;
 +   }
 +
 +   /* These all have an explicit depth value as their next source */
 +   switch (opcode) {
 +   case SpvOpImageSampleDrefImplicitLod:
 +   case SpvOpImageSampleDrefExplicitLod:
 +   case SpvOpImageSampleProjDrefImplicitLod:
 +   case SpvOpImageSampleProjDrefExplicitLod:
 +      (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparitor);
 +      break;
 +   default:
 +      break;
 +   }
 +
 +   /* For OpImageQuerySizeLod, we always have an LOD */
 +   if (opcode == SpvOpImageQuerySizeLod)
 +      (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
 +
 +   /* Figure out the base texture operation */
 +   nir_texop texop;
 +   switch (opcode) {
 +   case SpvOpImageSampleImplicitLod:
 +   case SpvOpImageSampleDrefImplicitLod:
 +   case SpvOpImageSampleProjImplicitLod:
 +   case SpvOpImageSampleProjDrefImplicitLod:
 +      texop = nir_texop_tex;
 +      break;
 +
 +   case SpvOpImageSampleExplicitLod:
 +   case SpvOpImageSampleDrefExplicitLod:
 +   case SpvOpImageSampleProjExplicitLod:
 +   case SpvOpImageSampleProjDrefExplicitLod:
 +      texop = nir_texop_txl;
 +      break;
 +
 +   case SpvOpImageFetch:
 +      if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
 +         texop = nir_texop_txf_ms;
 +      } else {
 +         texop = nir_texop_txf;
 +      }
 +      break;
 +
 +   case SpvOpImageGather:
 +   case SpvOpImageDrefGather:
 +      texop = nir_texop_tg4;
 +      break;
 +
 +   case SpvOpImageQuerySizeLod:
 +   case SpvOpImageQuerySize:
 +      texop = nir_texop_txs;
 +      break;
 +
 +   case SpvOpImageQueryLod:
 +      texop = nir_texop_lod;
 +      break;
 +
 +   case SpvOpImageQueryLevels:
 +      texop = nir_texop_query_levels;
 +      break;
 +
 +   case SpvOpImageQuerySamples:
 +   default:
 +      unreachable("Unhandled opcode");
 +   }
 +
 +   /* Now we need to handle some number of optional arguments */
 +   if (idx < count) {
 +      uint32_t operands = w[idx++];
 +
 +      if (operands & SpvImageOperandsBiasMask) {
 +         assert(texop == nir_texop_tex);
 +         texop = nir_texop_txb;
 +         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
 +      }
 +
 +      if (operands & SpvImageOperandsLodMask) {
 +         assert(texop == nir_texop_txl || texop == nir_texop_txf ||
 +                texop == nir_texop_txf_ms || texop == nir_texop_txs);
 +         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
 +      }
 +
 +      if (operands & SpvImageOperandsGradMask) {
 +         assert(texop == nir_texop_tex);
 +         texop = nir_texop_txd;
 +         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
 +         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
 +      }
 +
 +      if (operands & SpvImageOperandsOffsetMask ||
 +          operands & SpvImageOperandsConstOffsetMask)
 +         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
 +
 +      if (operands & SpvImageOperandsConstOffsetsMask)
 +         assert(!"Constant offsets to texture gather not yet implemented");
 +
 +      if (operands & SpvImageOperandsSampleMask) {
 +         assert(texop == nir_texop_txf_ms);
 +         texop = nir_texop_txf_ms;
 +         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
 +      }
 +   }
 +   /* We should have now consumed exactly all of the arguments */
 +   assert(idx == count);
 +
 +   nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
 +   instr->op = texop;
 +
 +   memcpy(instr->src, srcs, instr->num_srcs * sizeof(*instr->src));
 +
 +   instr->sampler_dim = glsl_get_sampler_dim(image_type);
 +   instr->is_array = glsl_sampler_type_is_array(image_type);
 +   instr->is_shadow = glsl_sampler_type_is_shadow(image_type);
 +   instr->is_new_style_shadow = instr->is_shadow;
 +
 +   if (has_coord) {
 +      switch (instr->sampler_dim) {
 +      case GLSL_SAMPLER_DIM_1D:
 +      case GLSL_SAMPLER_DIM_BUF:
 +         instr->coord_components = 1;
 +         break;
 +      case GLSL_SAMPLER_DIM_2D:
 +      case GLSL_SAMPLER_DIM_RECT:
 +      case GLSL_SAMPLER_DIM_MS:
 +         instr->coord_components = 2;
 +         break;
 +      case GLSL_SAMPLER_DIM_3D:
 +      case GLSL_SAMPLER_DIM_CUBE:
 +         instr->coord_components = 3;
 +         break;
 +      default:
 +         assert("Invalid sampler type");
 +      }
 +
 +      if (instr->is_array)
 +         instr->coord_components++;
 +   } else {
 +      instr->coord_components = 0;
 +   }
 +
 +   switch (glsl_get_sampler_result_type(image_type)) {
 +   case GLSL_TYPE_FLOAT:   instr->dest_type = nir_type_float;     break;
 +   case GLSL_TYPE_INT:     instr->dest_type = nir_type_int;       break;
 +   case GLSL_TYPE_UINT:    instr->dest_type = nir_type_uint;  break;
 +   case GLSL_TYPE_BOOL:    instr->dest_type = nir_type_bool;      break;
 +   default:
 +      unreachable("Invalid base type for sampler result");
 +   }
 +
 +   nir_deref_var *sampler = vtn_access_chain_to_deref(b, sampled.sampler);
 +   if (sampled.image) {
 +      nir_deref_var *image = vtn_access_chain_to_deref(b, sampled.image);
 +      instr->texture = nir_deref_as_var(nir_copy_deref(instr, &image->deref));
 +   } else {
 +      instr->texture = nir_deref_as_var(nir_copy_deref(instr, &sampler->deref));
 +   }
 +
 +   switch (instr->op) {
 +   case nir_texop_tex:
 +   case nir_texop_txb:
 +   case nir_texop_txl:
 +   case nir_texop_txd:
 +      /* These operations require a sampler */
 +      instr->sampler = nir_deref_as_var(nir_copy_deref(instr, &sampler->deref));
 +      break;
 +   case nir_texop_txf:
 +   case nir_texop_txf_ms:
 +   case nir_texop_txs:
 +   case nir_texop_lod:
 +   case nir_texop_tg4:
 +   case nir_texop_query_levels:
 +   case nir_texop_texture_samples:
 +   case nir_texop_samples_identical:
 +      /* These don't */
 +      instr->sampler = NULL;
 +      break;
 +   }
 +
 +   nir_ssa_dest_init(&instr->instr, &instr->dest,
 +                     nir_tex_instr_dest_size(instr), 32, NULL);
 +
 +   assert(glsl_get_vector_elements(ret_type->type) ==
 +          nir_tex_instr_dest_size(instr));
 +
 +   val->ssa = vtn_create_ssa_value(b, ret_type->type);
 +   val->ssa->def = &instr->dest.ssa;
 +
 +   nir_builder_instr_insert(&b->nb, &instr->instr);
 +}
 +
 +static nir_ssa_def *
 +get_image_coord(struct vtn_builder *b, uint32_t value)
 +{
 +   struct vtn_ssa_value *coord = vtn_ssa_value(b, value);
 +
 +   /* The image_load_store intrinsics assume a 4-dim coordinate */
 +   unsigned dim = glsl_get_vector_elements(coord->type);
 +   unsigned swizzle[4];
 +   for (unsigned i = 0; i < 4; i++)
 +      swizzle[i] = MIN2(i, dim - 1);
 +
 +   return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
 +}
 +
 +static void
 +vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
 +                 const uint32_t *w, unsigned count)
 +{
 +   /* Just get this one out of the way */
 +   if (opcode == SpvOpImageTexelPointer) {
 +      struct vtn_value *val =
 +         vtn_push_value(b, w[2], vtn_value_type_image_pointer);
 +      val->image = ralloc(b, struct vtn_image_pointer);
 +
 +      val->image->image =
 +         vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
 +      val->image->coord = get_image_coord(b, w[4]);
 +      val->image->sample = vtn_ssa_value(b, w[5])->def;
 +      return;
 +   }
 +
 +   struct vtn_image_pointer image;
 +
 +   switch (opcode) {
 +   case SpvOpAtomicExchange:
 +   case SpvOpAtomicCompareExchange:
 +   case SpvOpAtomicCompareExchangeWeak:
 +   case SpvOpAtomicIIncrement:
 +   case SpvOpAtomicIDecrement:
 +   case SpvOpAtomicIAdd:
 +   case SpvOpAtomicISub:
 +   case SpvOpAtomicSMin:
 +   case SpvOpAtomicUMin:
 +   case SpvOpAtomicSMax:
 +   case SpvOpAtomicUMax:
 +   case SpvOpAtomicAnd:
 +   case SpvOpAtomicOr:
 +   case SpvOpAtomicXor:
 +      image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image;
 +      break;
 +
 +   case SpvOpImageQuerySize:
 +      image.image =
 +         vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
 +      image.coord = NULL;
 +      image.sample = NULL;
 +      break;
 +
 +   case SpvOpImageRead:
 +      image.image =
 +         vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
 +      image.coord = get_image_coord(b, w[4]);
 +
 +      if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) {
 +         assert(w[5] == SpvImageOperandsSampleMask);
 +         image.sample = vtn_ssa_value(b, w[6])->def;
 +      } else {
 +         image.sample = nir_ssa_undef(&b->nb, 1, 32);
 +      }
 +      break;
 +
 +   case SpvOpImageWrite:
 +      image.image =
 +         vtn_value(b, w[1], vtn_value_type_access_chain)->access_chain;
 +      image.coord = get_image_coord(b, w[2]);
 +
 +      /* texel = w[3] */
 +
 +      if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) {
 +         assert(w[4] == SpvImageOperandsSampleMask);
 +         image.sample = vtn_ssa_value(b, w[5])->def;
 +      } else {
 +         image.sample = nir_ssa_undef(&b->nb, 1, 32);
 +      }
 +      break;
 +
 +   default:
 +      unreachable("Invalid image opcode");
 +   }
 +
 +   nir_intrinsic_op op;
 +   switch (opcode) {
 +#define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
 +   OP(ImageQuerySize,         size)
 +   OP(ImageRead,              load)
 +   OP(ImageWrite,             store)
 +   OP(AtomicExchange,         atomic_exchange)
 +   OP(AtomicCompareExchange,  atomic_comp_swap)
 +   OP(AtomicIIncrement,       atomic_add)
 +   OP(AtomicIDecrement,       atomic_add)
 +   OP(AtomicIAdd,             atomic_add)
 +   OP(AtomicISub,             atomic_add)
 +   OP(AtomicSMin,             atomic_min)
 +   OP(AtomicUMin,             atomic_min)
 +   OP(AtomicSMax,             atomic_max)
 +   OP(AtomicUMax,             atomic_max)
 +   OP(AtomicAnd,              atomic_and)
 +   OP(AtomicOr,               atomic_or)
 +   OP(AtomicXor,              atomic_xor)
 +#undef OP
 +   default:
 +      unreachable("Invalid image opcode");
 +   }
 +
 +   nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
 +
 +   nir_deref_var *image_deref = vtn_access_chain_to_deref(b, image.image);
 +   intrin->variables[0] =
 +      nir_deref_as_var(nir_copy_deref(&intrin->instr, &image_deref->deref));
 +
 +   /* ImageQuerySize doesn't take any extra parameters */
 +   if (opcode != SpvOpImageQuerySize) {
 +      /* The image coordinate is always 4 components but we may not have that
 +       * many.  Swizzle to compensate.
 +       */
 +      unsigned swiz[4];
 +      for (unsigned i = 0; i < 4; i++)
 +         swiz[i] = i < image.coord->num_components ? i : 0;
 +      intrin->src[0] = nir_src_for_ssa(nir_swizzle(&b->nb, image.coord,
 +                                                   swiz, 4, false));
 +      intrin->src[1] = nir_src_for_ssa(image.sample);
 +   }
 +
 +   switch (opcode) {
 +   case SpvOpImageQuerySize:
 +   case SpvOpImageRead:
 +      break;
 +   case SpvOpImageWrite:
 +      intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
 +      break;
 +   case SpvOpAtomicIIncrement:
 +      intrin->src[2] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
 +      break;
 +   case SpvOpAtomicIDecrement:
 +      intrin->src[2] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
 +      break;
 +
 +   case SpvOpAtomicExchange:
 +   case SpvOpAtomicIAdd:
 +   case SpvOpAtomicSMin:
 +   case SpvOpAtomicUMin:
 +   case SpvOpAtomicSMax:
 +   case SpvOpAtomicUMax:
 +   case SpvOpAtomicAnd:
 +   case SpvOpAtomicOr:
 +   case SpvOpAtomicXor:
 +      intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
 +      break;
 +
 +   case SpvOpAtomicCompareExchange:
 +      intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
 +      intrin->src[3] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
 +      break;
 +
 +   case SpvOpAtomicISub:
 +      intrin->src[2] = nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
 +      break;
 +
 +   default:
 +      unreachable("Invalid image opcode");
 +   }
 +
 +   if (opcode != SpvOpImageWrite) {
 +      struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
 +      struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
 +      nir_ssa_dest_init(&intrin->instr, &intrin->dest, 4, 32, NULL);
 +
 +      nir_builder_instr_insert(&b->nb, &intrin->instr);
 +
 +      /* The image intrinsics always return 4 channels but we may not want
 +       * that many.  Emit a mov to trim it down.
 +       */
 +      unsigned swiz[4] = {0, 1, 2, 3};
 +      val->ssa = vtn_create_ssa_value(b, type->type);
 +      val->ssa->def = nir_swizzle(&b->nb, &intrin->dest.ssa, swiz,
 +                                  glsl_get_vector_elements(type->type), false);
 +   } else {
 +      nir_builder_instr_insert(&b->nb, &intrin->instr);
 +   }
 +}
 +
 +static nir_intrinsic_op
 +get_ssbo_nir_atomic_op(SpvOp opcode)
 +{
 +   switch (opcode) {
 +#define OP(S, N) case SpvOp##S: return nir_intrinsic_ssbo_##N;
 +   OP(AtomicExchange,         atomic_exchange)
 +   OP(AtomicCompareExchange,  atomic_comp_swap)
 +   OP(AtomicIIncrement,       atomic_add)
 +   OP(AtomicIDecrement,       atomic_add)
 +   OP(AtomicIAdd,             atomic_add)
 +   OP(AtomicISub,             atomic_add)
 +   OP(AtomicSMin,             atomic_imin)
 +   OP(AtomicUMin,             atomic_umin)
 +   OP(AtomicSMax,             atomic_imax)
 +   OP(AtomicUMax,             atomic_umax)
 +   OP(AtomicAnd,              atomic_and)
 +   OP(AtomicOr,               atomic_or)
 +   OP(AtomicXor,              atomic_xor)
 +#undef OP
 +   default:
 +      unreachable("Invalid SSBO atomic");
 +   }
 +}
 +
 +static nir_intrinsic_op
 +get_shared_nir_atomic_op(SpvOp opcode)
 +{
 +   switch (opcode) {
 +#define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
 +   OP(AtomicExchange,         atomic_exchange)
 +   OP(AtomicCompareExchange,  atomic_comp_swap)
 +   OP(AtomicIIncrement,       atomic_add)
 +   OP(AtomicIDecrement,       atomic_add)
 +   OP(AtomicIAdd,             atomic_add)
 +   OP(AtomicISub,             atomic_add)
 +   OP(AtomicSMin,             atomic_imin)
 +   OP(AtomicUMin,             atomic_umin)
 +   OP(AtomicSMax,             atomic_imax)
 +   OP(AtomicUMax,             atomic_umax)
 +   OP(AtomicAnd,              atomic_and)
 +   OP(AtomicOr,               atomic_or)
 +   OP(AtomicXor,              atomic_xor)
 +#undef OP
 +   default:
 +      unreachable("Invalid shared atomic");
 +   }
 +}
 +
 +static void
 +fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
 +                           const uint32_t *w, nir_src *src)
 +{
 +   switch (opcode) {
 +   case SpvOpAtomicIIncrement:
 +      src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
 +      break;
 +
 +   case SpvOpAtomicIDecrement:
 +      src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
 +      break;
 +
 +   case SpvOpAtomicISub:
 +      src[0] =
 +         nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def));
 +      break;
 +
 +   case SpvOpAtomicCompareExchange:
 +      src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def);
 +      src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[8])->def);
 +      break;
 +      /* Fall through */
 +
 +   case SpvOpAtomicExchange:
 +   case SpvOpAtomicIAdd:
 +   case SpvOpAtomicSMin:
 +   case SpvOpAtomicUMin:
 +   case SpvOpAtomicSMax:
 +   case SpvOpAtomicUMax:
 +   case SpvOpAtomicAnd:
 +   case SpvOpAtomicOr:
 +   case SpvOpAtomicXor:
 +      src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def);
 +      break;
 +
 +   default:
 +      unreachable("Invalid SPIR-V atomic");
 +   }
 +}
 +
 +static void
 +vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
 +                                 const uint32_t *w, unsigned count)
 +{
 +   struct vtn_access_chain *chain =
 +      vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
 +   nir_intrinsic_instr *atomic;
 +
 +   /*
 +   SpvScope scope = w[4];
 +   SpvMemorySemanticsMask semantics = w[5];
 +   */
 +
 +   if (chain->var->mode == vtn_variable_mode_workgroup) {
 +      nir_deref *deref = &vtn_access_chain_to_deref(b, chain)->deref;
 +      nir_intrinsic_op op = get_shared_nir_atomic_op(opcode);
 +      atomic = nir_intrinsic_instr_create(b->nb.shader, op);
 +      atomic->variables[0] = nir_deref_as_var(nir_copy_deref(atomic, deref));
 +      fill_common_atomic_sources(b, opcode, w, &atomic->src[0]);
 +   } else {
 +      assert(chain->var->mode == vtn_variable_mode_ssbo);
 +      struct vtn_type *type;
 +      nir_ssa_def *offset, *index;
 +      offset = vtn_access_chain_to_offset(b, chain, &index, &type, NULL, false);
 +
 +      nir_intrinsic_op op = get_ssbo_nir_atomic_op(opcode);
 +
 +      atomic = nir_intrinsic_instr_create(b->nb.shader, op);
 +      atomic->src[0] = nir_src_for_ssa(index);
 +      atomic->src[1] = nir_src_for_ssa(offset);
 +      fill_common_atomic_sources(b, opcode, w, &atomic->src[2]);
 +   }
 +
 +   nir_ssa_dest_init(&atomic->instr, &atomic->dest, 1, 32, NULL);
 +
 +   struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
 +   struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
 +   val->ssa = rzalloc(b, struct vtn_ssa_value);
 +   val->ssa->def = &atomic->dest.ssa;
 +   val->ssa->type = type->type;
 +
 +   nir_builder_instr_insert(&b->nb, &atomic->instr);
 +}
 +
 +static nir_alu_instr *
 +create_vec(nir_shader *shader, unsigned num_components, unsigned bit_size)
 +{
 +   nir_op op;
 +   switch (num_components) {
 +   case 1: op = nir_op_fmov; break;
 +   case 2: op = nir_op_vec2; break;
 +   case 3: op = nir_op_vec3; break;
 +   case 4: op = nir_op_vec4; break;
 +   default: unreachable("bad vector size");
 +   }
 +
 +   nir_alu_instr *vec = nir_alu_instr_create(shader, op);
 +   nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components,
 +                     bit_size, NULL);
 +   vec->dest.write_mask = (1 << num_components) - 1;
 +
 +   return vec;
 +}
 +
 +struct vtn_ssa_value *
 +vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
 +{
 +   if (src->transposed)
 +      return src->transposed;
 +
 +   struct vtn_ssa_value *dest =
 +      vtn_create_ssa_value(b, glsl_transposed_type(src->type));
 +
 +   for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
 +      nir_alu_instr *vec = create_vec(b->shader,
 +                                      glsl_get_matrix_columns(src->type),
 +                                      glsl_get_bit_size(glsl_get_base_type(src->type)));
 +      if (glsl_type_is_vector_or_scalar(src->type)) {
 +          vec->src[0].src = nir_src_for_ssa(src->def);
 +          vec->src[0].swizzle[0] = i;
 +      } else {
 +         for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
 +            vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
 +            vec->src[j].swizzle[0] = i;
 +         }
 +      }
 +      nir_builder_instr_insert(&b->nb, &vec->instr);
 +      dest->elems[i]->def = &vec->dest.dest.ssa;
 +   }
 +
 +   dest->transposed = src;
 +
 +   return dest;
 +}
 +
 +nir_ssa_def *
 +vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
 +{
 +   unsigned swiz[4] = { index };
 +   return nir_swizzle(&b->nb, src, swiz, 1, true);
 +}
 +
 +nir_ssa_def *
 +vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
 +                  unsigned index)
 +{
 +   nir_alu_instr *vec = create_vec(b->shader, src->num_components,
 +                                   src->bit_size);
 +
 +   for (unsigned i = 0; i < src->num_components; i++) {
 +      if (i == index) {
 +         vec->src[i].src = nir_src_for_ssa(insert);
 +      } else {
 +         vec->src[i].src = nir_src_for_ssa(src);
 +         vec->src[i].swizzle[0] = i;
 +      }
 +   }
 +
 +   nir_builder_instr_insert(&b->nb, &vec->instr);
 +
 +   return &vec->dest.dest.ssa;
 +}
 +
 +nir_ssa_def *
 +vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
 +                           nir_ssa_def *index)
 +{
 +   nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
 +   for (unsigned i = 1; i < src->num_components; i++)
 +      dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
 +                       vtn_vector_extract(b, src, i), dest);
 +
 +   return dest;
 +}
 +
 +nir_ssa_def *
 +vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
 +                          nir_ssa_def *insert, nir_ssa_def *index)
 +{
 +   nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
 +   for (unsigned i = 1; i < src->num_components; i++)
 +      dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
 +                       vtn_vector_insert(b, src, insert, i), dest);
 +
 +   return dest;
 +}
 +
 +static nir_ssa_def *
 +vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
 +                   nir_ssa_def *src0, nir_ssa_def *src1,
 +                   const uint32_t *indices)
 +{
 +   nir_alu_instr *vec = create_vec(b->shader, num_components, src0->bit_size);
 +
 +   for (unsigned i = 0; i < num_components; i++) {
 +      uint32_t index = indices[i];
 +      if (index == 0xffffffff) {
 +         vec->src[i].src =
 +            nir_src_for_ssa(nir_ssa_undef(&b->nb, 1, src0->bit_size));
 +      } else if (index < src0->num_components) {
 +         vec->src[i].src = nir_src_for_ssa(src0);
 +         vec->src[i].swizzle[0] = index;
 +      } else {
 +         vec->src[i].src = nir_src_for_ssa(src1);
 +         vec->src[i].swizzle[0] = index - src0->num_components;
 +      }
 +   }
 +
 +   nir_builder_instr_insert(&b->nb, &vec->instr);
 +
 +   return &vec->dest.dest.ssa;
 +}
 +
 +/*
 + * Concatentates a number of vectors/scalars together to produce a vector
 + */
 +static nir_ssa_def *
 +vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
 +                     unsigned num_srcs, nir_ssa_def **srcs)
 +{
 +   nir_alu_instr *vec = create_vec(b->shader, num_components,
 +                                   srcs[0]->bit_size);
 +
 +   unsigned dest_idx = 0;
 +   for (unsigned i = 0; i < num_srcs; i++) {
 +      nir_ssa_def *src = srcs[i];
 +      for (unsigned j = 0; j < src->num_components; j++) {
 +         vec->src[dest_idx].src = nir_src_for_ssa(src);
 +         vec->src[dest_idx].swizzle[0] = j;
 +         dest_idx++;
 +      }
 +   }
 +
 +   nir_builder_instr_insert(&b->nb, &vec->instr);
 +
 +   return &vec->dest.dest.ssa;
 +}
 +
 +static struct vtn_ssa_value *
 +vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
 +{
 +   struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
 +   dest->type = src->type;
 +
 +   if (glsl_type_is_vector_or_scalar(src->type)) {
 +      dest->def = src->def;
 +   } else {
 +      unsigned elems = glsl_get_length(src->type);
 +
 +      dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
 +      for (unsigned i = 0; i < elems; i++)
 +         dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
 +   }
 +
 +   return dest;
 +}
 +
 +static struct vtn_ssa_value *
 +vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
 +                     struct vtn_ssa_value *insert, const uint32_t *indices,
 +                     unsigned num_indices)
 +{
 +   struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
 +
 +   struct vtn_ssa_value *cur = dest;
 +   unsigned i;
 +   for (i = 0; i < num_indices - 1; i++) {
 +      cur = cur->elems[indices[i]];
 +   }
 +
 +   if (glsl_type_is_vector_or_scalar(cur->type)) {
 +      /* According to the SPIR-V spec, OpCompositeInsert may work down to
 +       * the component granularity. In that case, the last index will be
 +       * the index to insert the scalar into the vector.
 +       */
 +
 +      cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
 +   } else {
 +      cur->elems[indices[i]] = insert;
 +   }
 +
 +   return dest;
 +}
 +
 +static struct vtn_ssa_value *
 +vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
 +                      const uint32_t *indices, unsigned num_indices)
 +{
 +   struct vtn_ssa_value *cur = src;
 +   for (unsigned i = 0; i < num_indices; i++) {
 +      if (glsl_type_is_vector_or_scalar(cur->type)) {
 +         assert(i == num_indices - 1);
 +         /* According to the SPIR-V spec, OpCompositeExtract may work down to
 +          * the component granularity. The last index will be the index of the
 +          * vector to extract.
 +          */
 +
 +         struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
 +         ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
 +         ret->def = vtn_vector_extract(b, cur->def, indices[i]);
 +         return ret;
 +      } else {
 +         cur = cur->elems[indices[i]];
 +      }
 +   }
 +
 +   return cur;
 +}
 +
 +static void
 +vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
 +                     const uint32_t *w, unsigned count)
 +{
 +   struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
 +   const struct glsl_type *type =
 +      vtn_value(b, w[1], vtn_value_type_type)->type->type;
 +   val->ssa = vtn_create_ssa_value(b, type);
 +
 +   switch (opcode) {
 +   case SpvOpVectorExtractDynamic:
 +      val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
 +                                                 vtn_ssa_value(b, w[4])->def);
 +      break;
 +
 +   case SpvOpVectorInsertDynamic:
 +      val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
 +                                                vtn_ssa_value(b, w[4])->def,
 +                                                vtn_ssa_value(b, w[5])->def);
 +      break;
 +
 +   case SpvOpVectorShuffle:
 +      val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type),
 +                                         vtn_ssa_value(b, w[3])->def,
 +                                         vtn_ssa_value(b, w[4])->def,
 +                                         w + 5);
 +      break;
 +
 +   case SpvOpCompositeConstruct: {
 +      unsigned elems = count - 3;
 +      if (glsl_type_is_vector_or_scalar(type)) {
 +         nir_ssa_def *srcs[4];
 +         for (unsigned i = 0; i < elems; i++)
 +            srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
 +         val->ssa->def =
 +            vtn_vector_construct(b, glsl_get_vector_elements(type),
 +                                 elems, srcs);
 +      } else {
 +         val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
 +         for (unsigned i = 0; i < elems; i++)
 +            val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
 +      }
 +      break;
 +   }
 +   case SpvOpCompositeExtract:
 +      val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
 +                                       w + 4, count - 4);
 +      break;
 +
 +   case SpvOpCompositeInsert:
 +      val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
 +                                      vtn_ssa_value(b, w[3]),
 +                                      w + 5, count - 5);
 +      break;
 +
 +   case SpvOpCopyObject:
 +      val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
 +      break;
 +
 +   default:
 +      unreachable("unknown composite operation");
 +   }
 +}
 +
 +static void
 +vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
 +                   const uint32_t *w, unsigned count)
 +{
 +   nir_intrinsic_op intrinsic_op;
 +   switch (opcode) {
 +   case SpvOpEmitVertex:
 +   case SpvOpEmitStreamVertex:
 +      intrinsic_op = nir_intrinsic_emit_vertex;
 +      break;
 +   case SpvOpEndPrimitive:
 +   case SpvOpEndStreamPrimitive:
 +      intrinsic_op = nir_intrinsic_end_primitive;
 +      break;
 +   case SpvOpMemoryBarrier:
 +      intrinsic_op = nir_intrinsic_memory_barrier;
 +      break;
 +   case SpvOpControlBarrier:
 +      intrinsic_op = nir_intrinsic_barrier;
 +      break;
 +   default:
 +      unreachable("unknown barrier instruction");
 +   }
 +
 +   nir_intrinsic_instr *intrin =
 +      nir_intrinsic_instr_create(b->shader, intrinsic_op);
 +
 +   if (opcode == SpvOpEmitStreamVertex || opcode == SpvOpEndStreamPrimitive)
 +      nir_intrinsic_set_stream_id(intrin, w[1]);
 +
 +   nir_builder_instr_insert(&b->nb, &intrin->instr);
 +}
 +
 +static unsigned
 +gl_primitive_from_spv_execution_mode(SpvExecutionMode mode)
 +{
 +   switch (mode) {
 +   case SpvExecutionModeInputPoints:
 +   case SpvExecutionModeOutputPoints:
 +      return 0; /* GL_POINTS */
 +   case SpvExecutionModeInputLines:
 +      return 1; /* GL_LINES */
 +   case SpvExecutionModeInputLinesAdjacency:
 +      return 0x000A; /* GL_LINE_STRIP_ADJACENCY_ARB */
 +   case SpvExecutionModeTriangles:
 +      return 4; /* GL_TRIANGLES */
 +   case SpvExecutionModeInputTrianglesAdjacency:
 +      return 0x000C; /* GL_TRIANGLES_ADJACENCY_ARB */
 +   case SpvExecutionModeQuads:
 +      return 7; /* GL_QUADS */
 +   case SpvExecutionModeIsolines:
 +      return 0x8E7A; /* GL_ISOLINES */
 +   case SpvExecutionModeOutputLineStrip:
 +      return 3; /* GL_LINE_STRIP */
 +   case SpvExecutionModeOutputTriangleStrip:
 +      return 5; /* GL_TRIANGLE_STRIP */
 +   default:
 +      assert(!"Invalid primitive type");
 +      return 4;
 +   }
 +}
 +
 +static unsigned
 +vertices_in_from_spv_execution_mode(SpvExecutionMode mode)
 +{
 +   switch (mode) {
 +   case SpvExecutionModeInputPoints:
 +      return 1;
 +   case SpvExecutionModeInputLines:
 +      return 2;
 +   case SpvExecutionModeInputLinesAdjacency:
 +      return 4;
 +   case SpvExecutionModeTriangles:
 +      return 3;
 +   case SpvExecutionModeInputTrianglesAdjacency:
 +      return 6;
 +   default:
 +      assert(!"Invalid GS input mode");
 +      return 0;
 +   }
 +}
 +
 +static gl_shader_stage
 +stage_for_execution_model(SpvExecutionModel model)
 +{
 +   switch (model) {
 +   case SpvExecutionModelVertex:
 +      return MESA_SHADER_VERTEX;
 +   case SpvExecutionModelTessellationControl:
 +      return MESA_SHADER_TESS_CTRL;
 +   case SpvExecutionModelTessellationEvaluation:
 +      return MESA_SHADER_TESS_EVAL;
 +   case SpvExecutionModelGeometry:
 +      return MESA_SHADER_GEOMETRY;
 +   case SpvExecutionModelFragment:
 +      return MESA_SHADER_FRAGMENT;
 +   case SpvExecutionModelGLCompute:
 +      return MESA_SHADER_COMPUTE;
 +   default:
 +      unreachable("Unsupported execution model");
 +   }
 +}
 +
 +static bool
 +vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
 +                                const uint32_t *w, unsigned count)
 +{
 +   switch (opcode) {
 +   case SpvOpSource:
 +   case SpvOpSourceExtension:
 +   case SpvOpSourceContinued:
 +   case SpvOpExtension:
 +      /* Unhandled, but these are for debug so that's ok. */
 +      break;
 +
 +   case SpvOpCapability: {
 +      SpvCapability cap = w[1];
 +      switch (cap) {
 +      case SpvCapabilityMatrix:
 +      case SpvCapabilityShader:
 +      case SpvCapabilityGeometry:
 +      case SpvCapabilityTessellationPointSize:
 +      case SpvCapabilityGeometryPointSize:
 +      case SpvCapabilityUniformBufferArrayDynamicIndexing:
 +      case SpvCapabilitySampledImageArrayDynamicIndexing:
 +      case SpvCapabilityStorageBufferArrayDynamicIndexing:
 +      case SpvCapabilityStorageImageArrayDynamicIndexing:
 +      case SpvCapabilityImageRect:
 +      case SpvCapabilitySampledRect:
 +      case SpvCapabilitySampled1D:
 +      case SpvCapabilityImage1D:
 +      case SpvCapabilitySampledCubeArray:
 +      case SpvCapabilitySampledBuffer:
 +      case SpvCapabilityImageBuffer:
 +      case SpvCapabilityImageQuery:
 +         break;
 +      case SpvCapabilityClipDistance:
 +      case SpvCapabilityCullDistance:
 +      case SpvCapabilityGeometryStreams:
 +         fprintf(stderr, "WARNING: Unsupported SPIR-V Capability\n");
 +         break;
 +      default:
 +         assert(!"Unsupported capability");
 +      }
 +      break;
 +   }
 +
 +   case SpvOpExtInstImport:
 +      vtn_handle_extension(b, opcode, w, count);
 +      break;
 +
 +   case SpvOpMemoryModel:
 +      assert(w[1] == SpvAddressingModelLogical);
 +      assert(w[2] == SpvMemoryModelGLSL450);
 +      break;
 +
 +   case SpvOpEntryPoint: {
 +      struct vtn_value *entry_point = &b->values[w[2]];
 +      /* Let this be a name label regardless */
 +      unsigned name_words;
 +      entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
 +
 +      if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
 +          stage_for_execution_model(w[1]) != b->entry_point_stage)
 +         break;
 +
 +      assert(b->entry_point == NULL);
 +      b->entry_point = entry_point;
 +      break;
 +   }
 +
 +   case SpvOpString:
 +      vtn_push_value(b, w[1], vtn_value_type_string)->str =
 +         vtn_string_literal(b, &w[2], count - 2, NULL);
 +      break;
 +
 +   case SpvOpName:
 +      b->values[w[1]].name = vtn_string_literal(b, &w[2], count - 2, NULL);
 +      break;
 +
 +   case SpvOpMemberName:
 +      /* TODO */
 +      break;
 +
 +   case SpvOpExecutionMode:
 +   case SpvOpDecorationGroup:
 +   case SpvOpDecorate:
 +   case SpvOpMemberDecorate:
 +   case SpvOpGroupDecorate:
 +   case SpvOpGroupMemberDecorate:
 +      vtn_handle_decoration(b, opcode, w, count);
 +      break;
 +
 +   default:
 +      return false; /* End of preamble */
 +   }
 +
 +   return true;
 +}
 +
 +static void
 +vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
 +                          const struct vtn_decoration *mode, void *data)
 +{
 +   assert(b->entry_point == entry_point);
 +
 +   switch(mode->exec_mode) {
 +   case SpvExecutionModeOriginUpperLeft:
 +   case SpvExecutionModeOriginLowerLeft:
 +      b->origin_upper_left =
 +         (mode->exec_mode == SpvExecutionModeOriginUpperLeft);
 +      break;
 +
 +   case SpvExecutionModeEarlyFragmentTests:
 +      assert(b->shader->stage == MESA_SHADER_FRAGMENT);
 +      b->shader->info.fs.early_fragment_tests = true;
 +      break;
 +
 +   case SpvExecutionModeInvocations:
 +      assert(b->shader->stage == MESA_SHADER_GEOMETRY);
 +      b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
 +      break;
 +
 +   case SpvExecutionModeDepthReplacing:
 +      assert(b->shader->stage == MESA_SHADER_FRAGMENT);
 +      b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_ANY;
 +      break;
 +   case SpvExecutionModeDepthGreater:
 +      assert(b->shader->stage == MESA_SHADER_FRAGMENT);
 +      b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_GREATER;
 +      break;
 +   case SpvExecutionModeDepthLess:
 +      assert(b->shader->stage == MESA_SHADER_FRAGMENT);
 +      b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_LESS;
 +      break;
 +   case SpvExecutionModeDepthUnchanged:
 +      assert(b->shader->stage == MESA_SHADER_FRAGMENT);
 +      b->shader->info.fs.depth_layout = FRAG_DEPTH_LAYOUT_UNCHANGED;
 +      break;
 +
 +   case SpvExecutionModeLocalSize:
 +      assert(b->shader->stage == MESA_SHADER_COMPUTE);
 +      b->shader->info.cs.local_size[0] = mode->literals[0];
 +      b->shader->info.cs.local_size[1] = mode->literals[1];
 +      b->shader->info.cs.local_size[2] = mode->literals[2];
 +      break;
 +   case SpvExecutionModeLocalSizeHint:
 +      break; /* Nothing do do with this */
 +
 +   case SpvExecutionModeOutputVertices:
 +      assert(b->shader->stage == MESA_SHADER_GEOMETRY);
 +      b->shader->info.gs.vertices_out = mode->literals[0];
 +      break;
 +
 +   case SpvExecutionModeInputPoints:
 +   case SpvExecutionModeInputLines:
 +   case SpvExecutionModeInputLinesAdjacency:
 +   case SpvExecutionModeTriangles:
 +   case SpvExecutionModeInputTrianglesAdjacency:
 +   case SpvExecutionModeQuads:
 +   case SpvExecutionModeIsolines:
 +      if (b->shader->stage == MESA_SHADER_GEOMETRY) {
 +         b->shader->info.gs.vertices_in =
 +            vertices_in_from_spv_execution_mode(mode->exec_mode);
 +      } else {
 +         assert(!"Tesselation shaders not yet supported");
 +      }
 +      break;
 +
 +   case SpvExecutionModeOutputPoints:
 +   case SpvExecutionModeOutputLineStrip:
 +   case SpvExecutionModeOutputTriangleStrip:
 +      assert(b->shader->stage == MESA_SHADER_GEOMETRY);
 +      b->shader->info.gs.output_primitive =
 +         gl_primitive_from_spv_execution_mode(mode->exec_mode);
 +      break;
 +
 +   case SpvExecutionModeSpacingEqual:
 +   case SpvExecutionModeSpacingFractionalEven:
 +   case SpvExecutionModeSpacingFractionalOdd:
 +   case SpvExecutionModeVertexOrderCw:
 +   case SpvExecutionModeVertexOrderCcw:
 +   case SpvExecutionModePointMode:
 +      assert(!"TODO: Add tessellation metadata");
 +      break;
 +
 +   case SpvExecutionModePixelCenterInteger:
 +   case SpvExecutionModeXfb:
 +      assert(!"Unhandled execution mode");
 +      break;
 +
 +   case SpvExecutionModeVecTypeHint:
 +   case SpvExecutionModeContractionOff:
 +      break; /* OpenCL */
 +   }
 +}
 +
 +static bool
 +vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
 +                                        const uint32_t *w, unsigned count)
 +{
 +   switch (opcode) {
 +   case SpvOpSource:
 +   case SpvOpSourceContinued:
 +   case SpvOpSourceExtension:
 +   case SpvOpExtension:
 +   case SpvOpCapability:
 +   case SpvOpExtInstImport:
 +   case SpvOpMemoryModel:
 +   case SpvOpEntryPoint:
 +   case SpvOpExecutionMode:
 +   case SpvOpString:
 +   case SpvOpName:
 +   case SpvOpMemberName:
 +   case SpvOpDecorationGroup:
 +   case SpvOpDecorate:
 +   case SpvOpMemberDecorate:
 +   case SpvOpGroupDecorate:
 +   case SpvOpGroupMemberDecorate:
 +      assert(!"Invalid opcode types and variables section");
 +      break;
 +
 +   case SpvOpTypeVoid:
 +   case SpvOpTypeBool:
 +   case SpvOpTypeInt:
 +   case SpvOpTypeFloat:
 +   case SpvOpTypeVector:
 +   case SpvOpTypeMatrix:
 +   case SpvOpTypeImage:
 +   case SpvOpTypeSampler:
 +   case SpvOpTypeSampledImage:
 +   case SpvOpTypeArray:
 +   case SpvOpTypeRuntimeArray:
 +   case SpvOpTypeStruct:
 +   case SpvOpTypeOpaque:
 +   case SpvOpTypePointer:
 +   case SpvOpTypeFunction:
 +   case SpvOpTypeEvent:
 +   case SpvOpTypeDeviceEvent:
 +   case SpvOpTypeReserveId:
 +   case SpvOpTypeQueue:
 +   case SpvOpTypePipe:
 +      vtn_handle_type(b, opcode, w, count);
 +      break;
 +
 +   case SpvOpConstantTrue:
 +   case SpvOpConstantFalse:
 +   case SpvOpConstant:
 +   case SpvOpConstantComposite:
 +   case SpvOpConstantSampler:
 +   case SpvOpConstantNull:
 +   case SpvOpSpecConstantTrue:
 +   case SpvOpSpecConstantFalse:
 +   case SpvOpSpecConstant:
 +   case SpvOpSpecConstantComposite:
 +   case SpvOpSpecConstantOp:
 +      vtn_handle_constant(b, opcode, w, count);
 +      break;
 +
 +   case SpvOpVariable:
 +      vtn_handle_variables(b, opcode, w, count);
 +      break;
 +
 +   default:
 +      return false; /* End of preamble */
 +   }
 +
 +   return true;
 +}
 +
 +static bool
 +vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
 +                            const uint32_t *w, unsigned count)
 +{
 +   switch (opcode) {
 +   case SpvOpLabel:
 +      break;
 +
 +   case SpvOpLoopMerge:
 +   case SpvOpSelectionMerge:
 +      /* This is handled by cfg pre-pass and walk_blocks */
 +      break;
 +
 +   case SpvOpUndef: {
 +      struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef);
 +      val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
 +      break;
 +   }
 +
 +   case SpvOpExtInst:
 +      vtn_handle_extension(b, opcode, w, count);
 +      break;
 +
 +   case SpvOpVariable:
 +   case SpvOpLoad:
 +   case SpvOpStore:
 +   case SpvOpCopyMemory:
 +   case SpvOpCopyMemorySized:
 +   case SpvOpAccessChain:
 +   case SpvOpInBoundsAccessChain:
 +   case SpvOpArrayLength:
 +      vtn_handle_variables(b, opcode, w, count);
 +      break;
 +
 +   case SpvOpFunctionCall:
 +      vtn_handle_function_call(b, opcode, w, count);
 +      break;
 +
 +   case SpvOpSampledImage:
 +   case SpvOpImage:
 +   case SpvOpImageSampleImplicitLod:
 +   case SpvOpImageSampleExplicitLod:
 +   case SpvOpImageSampleDrefImplicitLod:
 +   case SpvOpImageSampleDrefExplicitLod:
 +   case SpvOpImageSampleProjImplicitLod:
 +   case SpvOpImageSampleProjExplicitLod:
 +   case SpvOpImageSampleProjDrefImplicitLod:
 +   case SpvOpImageSampleProjDrefExplicitLod:
 +   case SpvOpImageFetch:
 +   case SpvOpImageGather:
 +   case SpvOpImageDrefGather:
 +   case SpvOpImageQuerySizeLod:
 +   case SpvOpImageQueryLod:
 +   case SpvOpImageQueryLevels:
 +   case SpvOpImageQuerySamples:
 +      vtn_handle_texture(b, opcode, w, count);
 +      break;
 +
 +   case SpvOpImageRead:
 +   case SpvOpImageWrite:
 +   case SpvOpImageTexelPointer:
 +      vtn_handle_image(b, opcode, w, count);
 +      break;
 +
 +   case SpvOpImageQuerySize: {
 +      struct vtn_access_chain *image =
 +         vtn_value(b, w[3], vtn_value_type_access_chain)->access_chain;
 +      if (glsl_type_is_image(image->var->var->interface_type)) {
 +         vtn_handle_image(b, opcode, w, count);
 +      } else {
 +         vtn_handle_texture(b, opcode, w, count);
 +      }
 +      break;
 +   }
 +
 +   case SpvOpAtomicExchange:
 +   case SpvOpAtomicCompareExchange:
 +   case SpvOpAtomicCompareExchangeWeak:
 +   case SpvOpAtomicIIncrement:
 +   case SpvOpAtomicIDecrement:
 +   case SpvOpAtomicIAdd:
 +   case SpvOpAtomicISub:
 +   case SpvOpAtomicSMin:
 +   case SpvOpAtomicUMin:
 +   case SpvOpAtomicSMax:
 +   case SpvOpAtomicUMax:
 +   case SpvOpAtomicAnd:
 +   case SpvOpAtomicOr:
 +   case SpvOpAtomicXor: {
 +      struct vtn_value *pointer = vtn_untyped_value(b, w[3]);
 +      if (pointer->value_type == vtn_value_type_image_pointer) {
 +         vtn_handle_image(b, opcode, w, count);
 +      } else {
 +         assert(pointer->value_type == vtn_value_type_access_chain);
 +         vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
 +      }
 +      break;
 +   }
 +
 +   case SpvOpSNegate:
 +   case SpvOpFNegate:
 +   case SpvOpNot:
 +   case SpvOpAny:
 +   case SpvOpAll:
 +   case SpvOpConvertFToU:
 +   case SpvOpConvertFToS:
 +   case SpvOpConvertSToF:
 +   case SpvOpConvertUToF:
 +   case SpvOpUConvert:
 +   case SpvOpSConvert:
 +   case SpvOpFConvert:
 +   case SpvOpQuantizeToF16:
 +   case SpvOpConvertPtrToU:
 +   case SpvOpConvertUToPtr:
 +   case SpvOpPtrCastToGeneric:
 +   case SpvOpGenericCastToPtr:
 +   case SpvOpBitcast:
 +   case SpvOpIsNan:
 +   case SpvOpIsInf:
 +   case SpvOpIsFinite:
 +   case SpvOpIsNormal:
 +   case SpvOpSignBitSet:
 +   case SpvOpLessOrGreater:
 +   case SpvOpOrdered:
 +   case SpvOpUnordered:
 +   case SpvOpIAdd:
 +   case SpvOpFAdd:
 +   case SpvOpISub:
 +   case SpvOpFSub:
 +   case SpvOpIMul:
 +   case SpvOpFMul:
 +   case SpvOpUDiv:
 +   case SpvOpSDiv:
 +   case SpvOpFDiv:
 +   case SpvOpUMod:
 +   case SpvOpSRem:
 +   case SpvOpSMod:
 +   case SpvOpFRem:
 +   case SpvOpFMod:
 +   case SpvOpVectorTimesScalar:
 +   case SpvOpDot:
 +   case SpvOpIAddCarry:
 +   case SpvOpISubBorrow:
 +   case SpvOpUMulExtended:
 +   case SpvOpSMulExtended:
 +   case SpvOpShiftRightLogical:
 +   case SpvOpShiftRightArithmetic:
 +   case SpvOpShiftLeftLogical:
 +   case SpvOpLogicalEqual:
 +   case SpvOpLogicalNotEqual:
 +   case SpvOpLogicalOr:
 +   case SpvOpLogicalAnd:
 +   case SpvOpLogicalNot:
 +   case SpvOpBitwiseOr:
 +   case SpvOpBitwiseXor:
 +   case SpvOpBitwiseAnd:
 +   case SpvOpSelect:
 +   case SpvOpIEqual:
 +   case SpvOpFOrdEqual:
 +   case SpvOpFUnordEqual:
 +   case SpvOpINotEqual:
 +   case SpvOpFOrdNotEqual:
 +   case SpvOpFUnordNotEqual:
 +   case SpvOpULessThan:
 +   case SpvOpSLessThan:
 +   case SpvOpFOrdLessThan:
 +   case SpvOpFUnordLessThan:
 +   case SpvOpUGreaterThan:
 +   case SpvOpSGreaterThan:
 +   case SpvOpFOrdGreaterThan:
 +   case SpvOpFUnordGreaterThan:
 +   case SpvOpULessThanEqual:
 +   case SpvOpSLessThanEqual:
 +   case SpvOpFOrdLessThanEqual:
 +   case SpvOpFUnordLessThanEqual:
 +   case SpvOpUGreaterThanEqual:
 +   case SpvOpSGreaterThanEqual:
 +   case SpvOpFOrdGreaterThanEqual:
 +   case SpvOpFUnordGreaterThanEqual:
 +   case SpvOpDPdx:
 +   case SpvOpDPdy:
 +   case SpvOpFwidth:
 +   case SpvOpDPdxFine:
 +   case SpvOpDPdyFine:
 +   case SpvOpFwidthFine:
 +   case SpvOpDPdxCoarse:
 +   case SpvOpDPdyCoarse:
 +   case SpvOpFwidthCoarse:
 +   case SpvOpBitFieldInsert:
 +   case SpvOpBitFieldSExtract:
 +   case SpvOpBitFieldUExtract:
 +   case SpvOpBitReverse:
 +   case SpvOpBitCount:
 +   case SpvOpTranspose:
 +   case SpvOpOuterProduct:
 +   case SpvOpMatrixTimesScalar:
 +   case SpvOpVectorTimesMatrix:
 +   case SpvOpMatrixTimesVector:
 +   case SpvOpMatrixTimesMatrix:
 +      vtn_handle_alu(b, opcode, w, count);
 +      break;
 +
 +   case SpvOpVectorExtractDynamic:
 +   case SpvOpVectorInsertDynamic:
 +   case SpvOpVectorShuffle:
 +   case SpvOpCompositeConstruct:
 +   case SpvOpCompositeExtract:
 +   case SpvOpCompositeInsert:
 +   case SpvOpCopyObject:
 +      vtn_handle_composite(b, opcode, w, count);
 +      break;
 +
 +   case SpvOpEmitVertex:
 +   case SpvOpEndPrimitive:
 +   case SpvOpEmitStreamVertex:
 +   case SpvOpEndStreamPrimitive:
 +   case SpvOpControlBarrier:
 +   case SpvOpMemoryBarrier:
 +      vtn_handle_barrier(b, opcode, w, count);
 +      break;
 +
 +   default:
 +      unreachable("Unhandled opcode");
 +   }
 +
 +   return true;
 +}
 +
 +nir_function *
 +spirv_to_nir(const uint32_t *words, size_t word_count,
 +             struct nir_spirv_specialization *spec, unsigned num_spec,
 +             gl_shader_stage stage, const char *entry_point_name,
 +             const nir_shader_compiler_options *options)
 +{
 +   const uint32_t *word_end = words + word_count;
 +
 +   /* Handle the SPIR-V header (first 4 dwords)  */
 +   assert(word_count > 5);
 +
 +   assert(words[0] == SpvMagicNumber);
 +   assert(words[1] >= 0x10000);
 +   /* words[2] == generator magic */
 +   unsigned value_id_bound = words[3];
 +   assert(words[4] == 0);
 +
 +   words+= 5;
 +
 +   /* Initialize the stn_builder object */
 +   struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
 +   b->value_id_bound = value_id_bound;
 +   b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
 +   exec_list_make_empty(&b->functions);
 +   b->entry_point_stage = stage;
 +   b->entry_point_name = entry_point_name;
 +
 +   /* Handle all the preamble instructions */
 +   words = vtn_foreach_instruction(b, words, word_end,
 +                                   vtn_handle_preamble_instruction);
 +
 +   if (b->entry_point == NULL) {
 +      assert(!"Entry point not found");
 +      ralloc_free(b);
 +      return NULL;
 +   }
 +
 +   b->shader = nir_shader_create(NULL, stage, options);
 +
 +   /* Set shader info defaults */
 +   b->shader->info.gs.invocations = 1;
 +
 +   /* Parse execution modes */
 +   vtn_foreach_execution_mode(b, b->entry_point,
 +                              vtn_handle_execution_mode, NULL);
 +
 +   b->specializations = spec;
 +   b->num_specializations = num_spec;
 +
 +   /* Handle all variable, type, and constant instructions */
 +   words = vtn_foreach_instruction(b, words, word_end,
 +                                   vtn_handle_variable_or_type_instruction);
 +
 +   vtn_build_cfg(b, words, word_end);
 +
 +   foreach_list_typed(struct vtn_function, func, node, &b->functions) {
 +      b->impl = func->impl;
 +      b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
 +                                               _mesa_key_pointer_equal);
 +
 +      vtn_function_emit(b, func, vtn_handle_body_instruction);
 +   }
 +
 +   assert(b->entry_point->value_type == vtn_value_type_function);
 +   nir_function *entry_point = b->entry_point->func->impl->function;
 +   assert(entry_point);
 +
 +   ralloc_free(b);
 +
 +   return entry_point;
 +}
index 56d79a02d79c56c88652e12a9e02a47a61402758,0000000000000000000000000000000000000000..52748a0619a17351b9778d03884bd1ad4273b228
mode 100644,000000..100644
--- /dev/null
@@@ -1,1370 -1,0 +1,1370 @@@
-    nir = brw_preprocess_nir(nir, compiler->scalar_stage[stage]);
 +/*
 + * 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.
 + */
 +
 +#include <assert.h>
 +#include <stdbool.h>
 +#include <string.h>
 +#include <unistd.h>
 +#include <fcntl.h>
 +
 +#include "util/mesa-sha1.h"
 +#include "anv_private.h"
 +#include "brw_nir.h"
 +#include "anv_nir.h"
 +#include "nir/spirv/nir_spirv.h"
 +
 +/* Needed for SWIZZLE macros */
 +#include "program/prog_instruction.h"
 +
 +// Shader functions
 +
 +VkResult anv_CreateShaderModule(
 +    VkDevice                                    _device,
 +    const VkShaderModuleCreateInfo*             pCreateInfo,
 +    const VkAllocationCallbacks*                pAllocator,
 +    VkShaderModule*                             pShaderModule)
 +{
 +   ANV_FROM_HANDLE(anv_device, device, _device);
 +   struct anv_shader_module *module;
 +
 +   assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO);
 +   assert(pCreateInfo->flags == 0);
 +
 +   module = anv_alloc2(&device->alloc, pAllocator,
 +                       sizeof(*module) + pCreateInfo->codeSize, 8,
 +                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 +   if (module == NULL)
 +      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 +
 +   module->nir = NULL;
 +   module->size = pCreateInfo->codeSize;
 +   memcpy(module->data, pCreateInfo->pCode, module->size);
 +
 +   _mesa_sha1_compute(module->data, module->size, module->sha1);
 +
 +   *pShaderModule = anv_shader_module_to_handle(module);
 +
 +   return VK_SUCCESS;
 +}
 +
 +void anv_DestroyShaderModule(
 +    VkDevice                                    _device,
 +    VkShaderModule                              _module,
 +    const VkAllocationCallbacks*                pAllocator)
 +{
 +   ANV_FROM_HANDLE(anv_device, device, _device);
 +   ANV_FROM_HANDLE(anv_shader_module, module, _module);
 +
 +   anv_free2(&device->alloc, pAllocator, module);
 +}
 +
 +#define SPIR_V_MAGIC_NUMBER 0x07230203
 +
 +/* Eventually, this will become part of anv_CreateShader.  Unfortunately,
 + * we can't do that yet because we don't have the ability to copy nir.
 + */
 +static nir_shader *
 +anv_shader_compile_to_nir(struct anv_device *device,
 +                          struct anv_shader_module *module,
 +                          const char *entrypoint_name,
 +                          gl_shader_stage stage,
 +                          const VkSpecializationInfo *spec_info)
 +{
 +   if (strcmp(entrypoint_name, "main") != 0) {
 +      anv_finishme("Multiple shaders per module not really supported");
 +   }
 +
 +   const struct brw_compiler *compiler =
 +      device->instance->physicalDevice.compiler;
 +   const nir_shader_compiler_options *nir_options =
 +      compiler->glsl_compiler_options[stage].NirOptions;
 +
 +   nir_shader *nir;
 +   nir_function *entry_point;
 +   if (module->nir) {
 +      /* Some things such as our meta clear/blit code will give us a NIR
 +       * shader directly.  In that case, we just ignore the SPIR-V entirely
 +       * and just use the NIR shader */
 +      nir = module->nir;
 +      nir->options = nir_options;
 +      nir_validate_shader(nir);
 +
 +      assert(exec_list_length(&nir->functions) == 1);
 +      struct exec_node *node = exec_list_get_head(&nir->functions);
 +      entry_point = exec_node_data(nir_function, node, node);
 +   } else {
 +      uint32_t *spirv = (uint32_t *) module->data;
 +      assert(spirv[0] == SPIR_V_MAGIC_NUMBER);
 +      assert(module->size % 4 == 0);
 +
 +      uint32_t num_spec_entries = 0;
 +      struct nir_spirv_specialization *spec_entries = NULL;
 +      if (spec_info && spec_info->mapEntryCount > 0) {
 +         num_spec_entries = spec_info->mapEntryCount;
 +         spec_entries = malloc(num_spec_entries * sizeof(*spec_entries));
 +         for (uint32_t i = 0; i < num_spec_entries; i++) {
 +            const uint32_t *data =
 +               spec_info->pData + spec_info->pMapEntries[i].offset;
 +            assert((const void *)(data + 1) <=
 +                   spec_info->pData + spec_info->dataSize);
 +
 +            spec_entries[i].id = spec_info->pMapEntries[i].constantID;
 +            spec_entries[i].data = *data;
 +         }
 +      }
 +
 +      entry_point = spirv_to_nir(spirv, module->size / 4,
 +                                 spec_entries, num_spec_entries,
 +                                 stage, entrypoint_name, nir_options);
 +      nir = entry_point->shader;
 +      assert(nir->stage == stage);
 +      nir_validate_shader(nir);
 +
 +      free(spec_entries);
 +
 +      nir_lower_returns(nir);
 +      nir_validate_shader(nir);
 +
 +      nir_inline_functions(nir);
 +      nir_validate_shader(nir);
 +
 +      /* Pick off the single entrypoint that we want */
 +      foreach_list_typed_safe(nir_function, func, node, &nir->functions) {
 +         if (func != entry_point)
 +            exec_node_remove(&func->node);
 +      }
 +      assert(exec_list_length(&nir->functions) == 1);
 +      entry_point->name = ralloc_strdup(entry_point, "main");
 +
 +      nir_remove_dead_variables(nir, nir_var_shader_in);
 +      nir_remove_dead_variables(nir, nir_var_shader_out);
 +      nir_remove_dead_variables(nir, nir_var_system_value);
 +      nir_validate_shader(nir);
 +
 +      nir_lower_outputs_to_temporaries(entry_point->shader, entry_point);
 +
 +      nir_lower_system_values(nir);
 +      nir_validate_shader(nir);
 +   }
 +
 +   /* Vulkan uses the separate-shader linking model */
 +   nir->info.separate_shader = true;
 +
++   nir = brw_preprocess_nir(compiler, nir);
 +
 +   nir_shader_gather_info(nir, entry_point->impl);
 +
 +   uint32_t indirect_mask = 0;
 +   if (compiler->glsl_compiler_options[stage].EmitNoIndirectInput)
 +      indirect_mask |= (1 << nir_var_shader_in);
 +   if (compiler->glsl_compiler_options[stage].EmitNoIndirectTemp)
 +      indirect_mask |= 1 << nir_var_local;
 +
 +   nir_lower_indirect_derefs(nir, indirect_mask);
 +
 +   return nir;
 +}
 +
 +void anv_DestroyPipeline(
 +    VkDevice                                    _device,
 +    VkPipeline                                  _pipeline,
 +    const VkAllocationCallbacks*                pAllocator)
 +{
 +   ANV_FROM_HANDLE(anv_device, device, _device);
 +   ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline);
 +
 +   anv_reloc_list_finish(&pipeline->batch_relocs,
 +                         pAllocator ? pAllocator : &device->alloc);
 +   if (pipeline->blend_state.map)
 +      anv_state_pool_free(&device->dynamic_state_pool, pipeline->blend_state);
 +   anv_free2(&device->alloc, pAllocator, pipeline);
 +}
 +
 +static const uint32_t vk_to_gen_primitive_type[] = {
 +   [VK_PRIMITIVE_TOPOLOGY_POINT_LIST]                    = _3DPRIM_POINTLIST,
 +   [VK_PRIMITIVE_TOPOLOGY_LINE_LIST]                     = _3DPRIM_LINELIST,
 +   [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP]                    = _3DPRIM_LINESTRIP,
 +   [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST]                 = _3DPRIM_TRILIST,
 +   [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP]                = _3DPRIM_TRISTRIP,
 +   [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN]                  = _3DPRIM_TRIFAN,
 +   [VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY]      = _3DPRIM_LINELIST_ADJ,
 +   [VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY]     = _3DPRIM_LINESTRIP_ADJ,
 +   [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY]  = _3DPRIM_TRILIST_ADJ,
 +   [VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY] = _3DPRIM_TRISTRIP_ADJ,
 +/*   [VK_PRIMITIVE_TOPOLOGY_PATCH_LIST]                = _3DPRIM_PATCHLIST_1 */
 +};
 +
 +static void
 +populate_sampler_prog_key(const struct brw_device_info *devinfo,
 +                          struct brw_sampler_prog_key_data *key)
 +{
 +   /* XXX: Handle texture swizzle on HSW- */
 +   for (int i = 0; i < MAX_SAMPLERS; i++) {
 +      /* Assume color sampler, no swizzling. (Works for BDW+) */
 +      key->swizzles[i] = SWIZZLE_XYZW;
 +   }
 +}
 +
 +static void
 +populate_vs_prog_key(const struct brw_device_info *devinfo,
 +                     struct brw_vs_prog_key *key)
 +{
 +   memset(key, 0, sizeof(*key));
 +
 +   populate_sampler_prog_key(devinfo, &key->tex);
 +
 +   /* XXX: Handle vertex input work-arounds */
 +
 +   /* XXX: Handle sampler_prog_key */
 +}
 +
 +static void
 +populate_gs_prog_key(const struct brw_device_info *devinfo,
 +                     struct brw_gs_prog_key *key)
 +{
 +   memset(key, 0, sizeof(*key));
 +
 +   populate_sampler_prog_key(devinfo, &key->tex);
 +}
 +
 +static void
 +populate_wm_prog_key(const struct brw_device_info *devinfo,
 +                     const VkGraphicsPipelineCreateInfo *info,
 +                     const struct anv_graphics_pipeline_create_info *extra,
 +                     struct brw_wm_prog_key *key)
 +{
 +   ANV_FROM_HANDLE(anv_render_pass, render_pass, info->renderPass);
 +
 +   memset(key, 0, sizeof(*key));
 +
 +   populate_sampler_prog_key(devinfo, &key->tex);
 +
 +   /* TODO: Fill out key->input_slots_valid */
 +
 +   /* Vulkan doesn't specify a default */
 +   key->high_quality_derivatives = false;
 +
 +   /* XXX Vulkan doesn't appear to specify */
 +   key->clamp_fragment_color = false;
 +
 +   /* Vulkan always specifies upper-left coordinates */
 +   key->drawable_height = 0;
 +   key->render_to_fbo = false;
 +
 +   if (extra && extra->color_attachment_count >= 0) {
 +      key->nr_color_regions = extra->color_attachment_count;
 +   } else {
 +      key->nr_color_regions =
 +         render_pass->subpasses[info->subpass].color_count;
 +   }
 +
 +   key->replicate_alpha = key->nr_color_regions > 1 &&
 +                          info->pMultisampleState &&
 +                          info->pMultisampleState->alphaToCoverageEnable;
 +
 +   if (info->pMultisampleState && info->pMultisampleState->rasterizationSamples > 1) {
 +      /* We should probably pull this out of the shader, but it's fairly
 +       * harmless to compute it and then let dead-code take care of it.
 +       */
 +      key->persample_shading = info->pMultisampleState->sampleShadingEnable;
 +      if (key->persample_shading)
 +         key->persample_2x = info->pMultisampleState->rasterizationSamples == 2;
 +
 +      key->compute_pos_offset = info->pMultisampleState->sampleShadingEnable;
 +      key->compute_sample_id = info->pMultisampleState->sampleShadingEnable;
 +   }
 +}
 +
 +static void
 +populate_cs_prog_key(const struct brw_device_info *devinfo,
 +                     struct brw_cs_prog_key *key)
 +{
 +   memset(key, 0, sizeof(*key));
 +
 +   populate_sampler_prog_key(devinfo, &key->tex);
 +}
 +
 +static nir_shader *
 +anv_pipeline_compile(struct anv_pipeline *pipeline,
 +                     struct anv_shader_module *module,
 +                     const char *entrypoint,
 +                     gl_shader_stage stage,
 +                     const VkSpecializationInfo *spec_info,
 +                     struct brw_stage_prog_data *prog_data,
 +                     struct anv_pipeline_bind_map *map)
 +{
 +   const struct brw_compiler *compiler =
 +      pipeline->device->instance->physicalDevice.compiler;
 +
 +   nir_shader *nir = anv_shader_compile_to_nir(pipeline->device,
 +                                               module, entrypoint, stage,
 +                                               spec_info);
 +   if (nir == NULL)
 +      return NULL;
 +
 +   anv_nir_lower_push_constants(nir, compiler->scalar_stage[stage]);
 +
 +   /* Figure out the number of parameters */
 +   prog_data->nr_params = 0;
 +
 +   if (nir->num_uniforms > 0) {
 +      /* If the shader uses any push constants at all, we'll just give
 +       * them the maximum possible number
 +       */
 +      prog_data->nr_params += MAX_PUSH_CONSTANTS_SIZE / sizeof(float);
 +   }
 +
 +   if (pipeline->layout && pipeline->layout->stage[stage].has_dynamic_offsets)
 +      prog_data->nr_params += MAX_DYNAMIC_BUFFERS * 2;
 +
 +   if (nir->info.num_images > 0)
 +      prog_data->nr_params += nir->info.num_images * BRW_IMAGE_PARAM_SIZE;
 +
 +   if (prog_data->nr_params > 0) {
 +      /* XXX: I think we're leaking this */
 +      prog_data->param = (const union gl_constant_value **)
 +         malloc(prog_data->nr_params * sizeof(union gl_constant_value *));
 +
 +      /* We now set the param values to be offsets into a
 +       * anv_push_constant_data structure.  Since the compiler doesn't
 +       * actually dereference any of the gl_constant_value pointers in the
 +       * params array, it doesn't really matter what we put here.
 +       */
 +      struct anv_push_constants *null_data = NULL;
 +      if (nir->num_uniforms > 0) {
 +         /* Fill out the push constants section of the param array */
 +         for (unsigned i = 0; i < MAX_PUSH_CONSTANTS_SIZE / sizeof(float); i++)
 +            prog_data->param[i] = (const union gl_constant_value *)
 +               &null_data->client_data[i * sizeof(float)];
 +      }
 +   }
 +
 +   /* Set up dynamic offsets */
 +   anv_nir_apply_dynamic_offsets(pipeline, nir, prog_data);
 +
 +   /* Apply the actual pipeline layout to UBOs, SSBOs, and textures */
 +   if (pipeline->layout)
 +      anv_nir_apply_pipeline_layout(pipeline, nir, prog_data, map);
 +
 +   /* nir_lower_io will only handle the push constants; we need to set this
 +    * to the full number of possible uniforms.
 +    */
 +   nir->num_uniforms = prog_data->nr_params * 4;
 +
 +   return nir;
 +}
 +
 +static void
 +anv_fill_binding_table(struct brw_stage_prog_data *prog_data, unsigned bias)
 +{
 +   prog_data->binding_table.size_bytes = 0;
 +   prog_data->binding_table.texture_start = bias;
 +   prog_data->binding_table.ubo_start = bias;
 +   prog_data->binding_table.ssbo_start = bias;
 +   prog_data->binding_table.image_start = bias;
 +}
 +
 +static void
 +anv_pipeline_add_compiled_stage(struct anv_pipeline *pipeline,
 +                                gl_shader_stage stage,
 +                                const struct brw_stage_prog_data *prog_data,
 +                                struct anv_pipeline_bind_map *map)
 +{
 +   struct brw_device_info *devinfo = &pipeline->device->info;
 +   uint32_t max_threads[] = {
 +      [MESA_SHADER_VERTEX]                  = devinfo->max_vs_threads,
 +      [MESA_SHADER_TESS_CTRL]               = devinfo->max_hs_threads,
 +      [MESA_SHADER_TESS_EVAL]               = devinfo->max_ds_threads,
 +      [MESA_SHADER_GEOMETRY]                = devinfo->max_gs_threads,
 +      [MESA_SHADER_FRAGMENT]                = devinfo->max_wm_threads,
 +      [MESA_SHADER_COMPUTE]                 = devinfo->max_cs_threads,
 +   };
 +
 +   pipeline->prog_data[stage] = prog_data;
 +   pipeline->active_stages |= mesa_to_vk_shader_stage(stage);
 +   pipeline->scratch_start[stage] = pipeline->total_scratch;
 +   pipeline->total_scratch =
 +      align_u32(pipeline->total_scratch, 1024) +
 +      prog_data->total_scratch * max_threads[stage];
 +   pipeline->bindings[stage] = *map;
 +}
 +
 +static VkResult
 +anv_pipeline_compile_vs(struct anv_pipeline *pipeline,
 +                        struct anv_pipeline_cache *cache,
 +                        const VkGraphicsPipelineCreateInfo *info,
 +                        struct anv_shader_module *module,
 +                        const char *entrypoint,
 +                        const VkSpecializationInfo *spec_info)
 +{
 +   const struct brw_compiler *compiler =
 +      pipeline->device->instance->physicalDevice.compiler;
 +   const struct brw_stage_prog_data *stage_prog_data;
 +   struct anv_pipeline_bind_map map;
 +   struct brw_vs_prog_key key;
 +   uint32_t kernel = NO_KERNEL;
 +   unsigned char sha1[20];
 +
 +   populate_vs_prog_key(&pipeline->device->info, &key);
 +
 +   if (module->size > 0) {
 +      anv_hash_shader(sha1, &key, sizeof(key), module, entrypoint, spec_info);
 +      kernel = anv_pipeline_cache_search(cache, sha1, &stage_prog_data, &map);
 +   }
 +
 +   if (kernel == NO_KERNEL) {
 +      struct brw_vs_prog_data prog_data = { 0, };
 +      struct anv_pipeline_binding surface_to_descriptor[256];
 +      struct anv_pipeline_binding sampler_to_descriptor[256];
 +
 +      map = (struct anv_pipeline_bind_map) {
 +         .surface_to_descriptor = surface_to_descriptor,
 +         .sampler_to_descriptor = sampler_to_descriptor
 +      };
 +
 +      nir_shader *nir = anv_pipeline_compile(pipeline, module, entrypoint,
 +                                             MESA_SHADER_VERTEX, spec_info,
 +                                             &prog_data.base.base, &map);
 +      if (nir == NULL)
 +         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 +
 +      anv_fill_binding_table(&prog_data.base.base, 0);
 +
 +      void *mem_ctx = ralloc_context(NULL);
 +
 +      if (module->nir == NULL)
 +         ralloc_steal(mem_ctx, nir);
 +
 +      prog_data.inputs_read = nir->info.inputs_read;
 +
 +      brw_compute_vue_map(&pipeline->device->info,
 +                          &prog_data.base.vue_map,
 +                          nir->info.outputs_written,
 +                          nir->info.separate_shader);
 +
 +      unsigned code_size;
 +      const unsigned *shader_code =
 +         brw_compile_vs(compiler, NULL, mem_ctx, &key, &prog_data, nir,
 +                        NULL, false, -1, &code_size, NULL);
 +      if (shader_code == NULL) {
 +         ralloc_free(mem_ctx);
 +         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 +      }
 +
 +      stage_prog_data = &prog_data.base.base;
 +      kernel = anv_pipeline_cache_upload_kernel(cache,
 +                                                module->size > 0 ? sha1 : NULL,
 +                                                shader_code, code_size,
 +                                                &stage_prog_data, sizeof(prog_data),
 +                                                &map);
 +      ralloc_free(mem_ctx);
 +   }
 +
 +   const struct brw_vs_prog_data *vs_prog_data =
 +      (const struct brw_vs_prog_data *) stage_prog_data;
 +
 +   if (vs_prog_data->base.dispatch_mode == DISPATCH_MODE_SIMD8) {
 +      pipeline->vs_simd8 = kernel;
 +      pipeline->vs_vec4 = NO_KERNEL;
 +   } else {
 +      pipeline->vs_simd8 = NO_KERNEL;
 +      pipeline->vs_vec4 = kernel;
 +   }
 +
 +   anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_VERTEX,
 +                                   stage_prog_data, &map);
 +
 +   return VK_SUCCESS;
 +}
 +
 +static VkResult
 +anv_pipeline_compile_gs(struct anv_pipeline *pipeline,
 +                        struct anv_pipeline_cache *cache,
 +                        const VkGraphicsPipelineCreateInfo *info,
 +                        struct anv_shader_module *module,
 +                        const char *entrypoint,
 +                        const VkSpecializationInfo *spec_info)
 +{
 +   const struct brw_compiler *compiler =
 +      pipeline->device->instance->physicalDevice.compiler;
 +   const struct brw_stage_prog_data *stage_prog_data;
 +   struct anv_pipeline_bind_map map;
 +   struct brw_gs_prog_key key;
 +   uint32_t kernel = NO_KERNEL;
 +   unsigned char sha1[20];
 +
 +   populate_gs_prog_key(&pipeline->device->info, &key);
 +
 +   if (module->size > 0) {
 +      anv_hash_shader(sha1, &key, sizeof(key), module, entrypoint, spec_info);
 +      kernel = anv_pipeline_cache_search(cache, sha1, &stage_prog_data, &map);
 +   }
 +
 +   if (kernel == NO_KERNEL) {
 +      struct brw_gs_prog_data prog_data = { 0, };
 +      struct anv_pipeline_binding surface_to_descriptor[256];
 +      struct anv_pipeline_binding sampler_to_descriptor[256];
 +
 +      map = (struct anv_pipeline_bind_map) {
 +         .surface_to_descriptor = surface_to_descriptor,
 +         .sampler_to_descriptor = sampler_to_descriptor
 +      };
 +
 +      nir_shader *nir = anv_pipeline_compile(pipeline, module, entrypoint,
 +                                             MESA_SHADER_GEOMETRY, spec_info,
 +                                             &prog_data.base.base, &map);
 +      if (nir == NULL)
 +         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 +
 +      anv_fill_binding_table(&prog_data.base.base, 0);
 +
 +      void *mem_ctx = ralloc_context(NULL);
 +
 +      if (module->nir == NULL)
 +         ralloc_steal(mem_ctx, nir);
 +
 +      brw_compute_vue_map(&pipeline->device->info,
 +                          &prog_data.base.vue_map,
 +                          nir->info.outputs_written,
 +                          nir->info.separate_shader);
 +
 +      unsigned code_size;
 +      const unsigned *shader_code =
 +         brw_compile_gs(compiler, NULL, mem_ctx, &key, &prog_data, nir,
 +                        NULL, -1, &code_size, NULL);
 +      if (shader_code == NULL) {
 +         ralloc_free(mem_ctx);
 +         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 +      }
 +
 +      /* TODO: SIMD8 GS */
 +      stage_prog_data = &prog_data.base.base;
 +      kernel = anv_pipeline_cache_upload_kernel(cache,
 +                                                module->size > 0 ? sha1 : NULL,
 +                                                shader_code, code_size,
 +                                                &stage_prog_data, sizeof(prog_data),
 +                                                &map);
 +
 +      ralloc_free(mem_ctx);
 +   }
 +
 +   pipeline->gs_kernel = kernel;
 +
 +   anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_GEOMETRY,
 +                                   stage_prog_data, &map);
 +
 +   return VK_SUCCESS;
 +}
 +
 +static VkResult
 +anv_pipeline_compile_fs(struct anv_pipeline *pipeline,
 +                        struct anv_pipeline_cache *cache,
 +                        const VkGraphicsPipelineCreateInfo *info,
 +                        const struct anv_graphics_pipeline_create_info *extra,
 +                        struct anv_shader_module *module,
 +                        const char *entrypoint,
 +                        const VkSpecializationInfo *spec_info)
 +{
 +   const struct brw_compiler *compiler =
 +      pipeline->device->instance->physicalDevice.compiler;
 +   const struct brw_stage_prog_data *stage_prog_data;
 +   struct anv_pipeline_bind_map map;
 +   struct brw_wm_prog_key key;
 +   uint32_t kernel = NO_KERNEL;
 +   unsigned char sha1[20];
 +
 +   populate_wm_prog_key(&pipeline->device->info, info, extra, &key);
 +
 +   if (module->size > 0) {
 +      anv_hash_shader(sha1, &key, sizeof(key), module, entrypoint, spec_info);
 +      kernel = anv_pipeline_cache_search(cache, sha1, &stage_prog_data, &map);
 +   }
 +
 +   if (kernel == NO_KERNEL) {
 +      struct brw_wm_prog_data prog_data = { 0, };
 +      struct anv_pipeline_binding surface_to_descriptor[256];
 +      struct anv_pipeline_binding sampler_to_descriptor[256];
 +
 +      map = (struct anv_pipeline_bind_map) {
 +         .surface_to_descriptor = surface_to_descriptor + 8,
 +         .sampler_to_descriptor = sampler_to_descriptor
 +      };
 +
 +      nir_shader *nir = anv_pipeline_compile(pipeline, module, entrypoint,
 +                                             MESA_SHADER_FRAGMENT, spec_info,
 +                                             &prog_data.base, &map);
 +      if (nir == NULL)
 +         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 +
 +      unsigned num_rts = 0;
 +      struct anv_pipeline_binding rt_bindings[8];
 +      nir_function_impl *impl = nir_shader_get_entrypoint(nir)->impl;
 +      nir_foreach_variable_safe(var, &nir->outputs) {
 +         if (var->data.location < FRAG_RESULT_DATA0)
 +            continue;
 +
 +         unsigned rt = var->data.location - FRAG_RESULT_DATA0;
 +         if (rt >= key.nr_color_regions) {
 +            /* Out-of-bounds, throw it away */
 +            var->data.mode = nir_var_local;
 +            exec_node_remove(&var->node);
 +            exec_list_push_tail(&impl->locals, &var->node);
 +            continue;
 +         }
 +
 +         /* Give it a new, compacted, location */
 +         var->data.location = FRAG_RESULT_DATA0 + num_rts;
 +
 +         unsigned array_len =
 +            glsl_type_is_array(var->type) ? glsl_get_length(var->type) : 1;
 +         assert(num_rts + array_len <= 8);
 +
 +         for (unsigned i = 0; i < array_len; i++) {
 +            rt_bindings[num_rts] = (struct anv_pipeline_binding) {
 +               .set = ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS,
 +               .offset = rt + i,
 +            };
 +         }
 +
 +         num_rts += array_len;
 +      }
 +
 +      if (pipeline->use_repclear) {
 +         assert(num_rts == 1);
 +         key.nr_color_regions = 1;
 +      }
 +
 +      if (num_rts == 0) {
 +         /* If we have no render targets, we need a null render target */
 +         rt_bindings[0] = (struct anv_pipeline_binding) {
 +            .set = ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS,
 +            .offset = UINT16_MAX,
 +         };
 +         num_rts = 1;
 +      }
 +
 +      assert(num_rts <= 8);
 +      map.surface_to_descriptor -= num_rts;
 +      map.surface_count += num_rts;
 +      assert(map.surface_count <= 256);
 +      memcpy(map.surface_to_descriptor, rt_bindings,
 +             num_rts * sizeof(*rt_bindings));
 +
 +      anv_fill_binding_table(&prog_data.base, num_rts);
 +
 +      void *mem_ctx = ralloc_context(NULL);
 +
 +      if (module->nir == NULL)
 +         ralloc_steal(mem_ctx, nir);
 +
 +      unsigned code_size;
 +      const unsigned *shader_code =
 +         brw_compile_fs(compiler, NULL, mem_ctx, &key, &prog_data, nir,
 +                        NULL, -1, -1, pipeline->use_repclear, &code_size, NULL);
 +      if (shader_code == NULL) {
 +         ralloc_free(mem_ctx);
 +         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 +      }
 +
 +      stage_prog_data = &prog_data.base;
 +      kernel = anv_pipeline_cache_upload_kernel(cache,
 +                                                module->size > 0 ? sha1 : NULL,
 +                                                shader_code, code_size,
 +                                                &stage_prog_data, sizeof(prog_data),
 +                                                &map);
 +
 +      ralloc_free(mem_ctx);
 +   }
 +
 +   const struct brw_wm_prog_data *wm_prog_data =
 +      (const struct brw_wm_prog_data *) stage_prog_data;
 +
 +   if (wm_prog_data->no_8)
 +      pipeline->ps_simd8 = NO_KERNEL;
 +   else
 +      pipeline->ps_simd8 = kernel;
 +
 +   if (wm_prog_data->no_8 || wm_prog_data->prog_offset_16) {
 +      pipeline->ps_simd16 = kernel + wm_prog_data->prog_offset_16;
 +   } else {
 +      pipeline->ps_simd16 = NO_KERNEL;
 +   }
 +
 +   pipeline->ps_ksp2 = 0;
 +   pipeline->ps_grf_start2 = 0;
 +   if (pipeline->ps_simd8 != NO_KERNEL) {
 +      pipeline->ps_ksp0 = pipeline->ps_simd8;
 +      pipeline->ps_grf_start0 = wm_prog_data->base.dispatch_grf_start_reg;
 +      if (pipeline->ps_simd16 != NO_KERNEL) {
 +         pipeline->ps_ksp2 = pipeline->ps_simd16;
 +         pipeline->ps_grf_start2 = wm_prog_data->dispatch_grf_start_reg_16;
 +      }
 +   } else if (pipeline->ps_simd16 != NO_KERNEL) {
 +      pipeline->ps_ksp0 = pipeline->ps_simd16;
 +      pipeline->ps_grf_start0 = wm_prog_data->dispatch_grf_start_reg_16;
 +   }
 +
 +   anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_FRAGMENT,
 +                                   stage_prog_data, &map);
 +
 +   return VK_SUCCESS;
 +}
 +
 +VkResult
 +anv_pipeline_compile_cs(struct anv_pipeline *pipeline,
 +                        struct anv_pipeline_cache *cache,
 +                        const VkComputePipelineCreateInfo *info,
 +                        struct anv_shader_module *module,
 +                        const char *entrypoint,
 +                        const VkSpecializationInfo *spec_info)
 +{
 +   const struct brw_compiler *compiler =
 +      pipeline->device->instance->physicalDevice.compiler;
 +   const struct brw_stage_prog_data *stage_prog_data;
 +   struct anv_pipeline_bind_map map;
 +   struct brw_cs_prog_key key;
 +   uint32_t kernel = NO_KERNEL;
 +   unsigned char sha1[20];
 +
 +   populate_cs_prog_key(&pipeline->device->info, &key);
 +
 +   if (module->size > 0) {
 +      anv_hash_shader(sha1, &key, sizeof(key), module, entrypoint, spec_info);
 +      kernel = anv_pipeline_cache_search(cache, sha1, &stage_prog_data, &map);
 +   }
 +
 +   if (module->size == 0 || kernel == NO_KERNEL) {
 +      struct brw_cs_prog_data prog_data = { 0, };
 +      struct anv_pipeline_binding surface_to_descriptor[256];
 +      struct anv_pipeline_binding sampler_to_descriptor[256];
 +
 +      map = (struct anv_pipeline_bind_map) {
 +         .surface_to_descriptor = surface_to_descriptor,
 +         .sampler_to_descriptor = sampler_to_descriptor
 +      };
 +
 +      nir_shader *nir = anv_pipeline_compile(pipeline, module, entrypoint,
 +                                             MESA_SHADER_COMPUTE, spec_info,
 +                                             &prog_data.base, &map);
 +      if (nir == NULL)
 +         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 +
 +      anv_fill_binding_table(&prog_data.base, 1);
 +
 +      void *mem_ctx = ralloc_context(NULL);
 +
 +      if (module->nir == NULL)
 +         ralloc_steal(mem_ctx, nir);
 +
 +      unsigned code_size;
 +      const unsigned *shader_code =
 +         brw_compile_cs(compiler, NULL, mem_ctx, &key, &prog_data, nir,
 +                        -1, &code_size, NULL);
 +      if (shader_code == NULL) {
 +         ralloc_free(mem_ctx);
 +         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 +      }
 +
 +      stage_prog_data = &prog_data.base;
 +      kernel = anv_pipeline_cache_upload_kernel(cache,
 +                                                module->size > 0 ? sha1 : NULL,
 +                                                shader_code, code_size,
 +                                                &stage_prog_data, sizeof(prog_data),
 +                                                &map);
 +
 +      ralloc_free(mem_ctx);
 +   }
 +
 +   pipeline->cs_simd = kernel;
 +
 +   anv_pipeline_add_compiled_stage(pipeline, MESA_SHADER_COMPUTE,
 +                                   stage_prog_data, &map);
 +
 +   return VK_SUCCESS;
 +}
 +
 +static void
 +gen7_compute_urb_partition(struct anv_pipeline *pipeline)
 +{
 +   const struct brw_device_info *devinfo = &pipeline->device->info;
 +   bool vs_present = pipeline->active_stages & VK_SHADER_STAGE_VERTEX_BIT;
 +   unsigned vs_size = vs_present ?
 +      get_vs_prog_data(pipeline)->base.urb_entry_size : 1;
 +   unsigned vs_entry_size_bytes = vs_size * 64;
 +   bool gs_present = pipeline->active_stages & VK_SHADER_STAGE_GEOMETRY_BIT;
 +   unsigned gs_size = gs_present ?
 +      get_gs_prog_data(pipeline)->base.urb_entry_size : 1;
 +   unsigned gs_entry_size_bytes = gs_size * 64;
 +
 +   /* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
 +    *
 +    *     VS Number of URB Entries must be divisible by 8 if the VS URB Entry
 +    *     Allocation Size is less than 9 512-bit URB entries.
 +    *
 +    * Similar text exists for GS.
 +    */
 +   unsigned vs_granularity = (vs_size < 9) ? 8 : 1;
 +   unsigned gs_granularity = (gs_size < 9) ? 8 : 1;
 +
 +   /* URB allocations must be done in 8k chunks. */
 +   unsigned chunk_size_bytes = 8192;
 +
 +   /* Determine the size of the URB in chunks. */
 +   unsigned urb_chunks = devinfo->urb.size * 1024 / chunk_size_bytes;
 +
 +   /* Reserve space for push constants */
 +   unsigned push_constant_kb;
 +   if (pipeline->device->info.gen >= 8)
 +      push_constant_kb = 32;
 +   else if (pipeline->device->info.is_haswell)
 +      push_constant_kb = pipeline->device->info.gt == 3 ? 32 : 16;
 +   else
 +      push_constant_kb = 16;
 +
 +   unsigned push_constant_bytes = push_constant_kb * 1024;
 +   unsigned push_constant_chunks =
 +      push_constant_bytes / chunk_size_bytes;
 +
 +   /* Initially, assign each stage the minimum amount of URB space it needs,
 +    * and make a note of how much additional space it "wants" (the amount of
 +    * additional space it could actually make use of).
 +    */
 +
 +   /* VS has a lower limit on the number of URB entries */
 +   unsigned vs_chunks =
 +      ALIGN(devinfo->urb.min_vs_entries * vs_entry_size_bytes,
 +            chunk_size_bytes) / chunk_size_bytes;
 +   unsigned vs_wants =
 +      ALIGN(devinfo->urb.max_vs_entries * vs_entry_size_bytes,
 +            chunk_size_bytes) / chunk_size_bytes - vs_chunks;
 +
 +   unsigned gs_chunks = 0;
 +   unsigned gs_wants = 0;
 +   if (gs_present) {
 +      /* There are two constraints on the minimum amount of URB space we can
 +       * allocate:
 +       *
 +       * (1) We need room for at least 2 URB entries, since we always operate
 +       * the GS in DUAL_OBJECT mode.
 +       *
 +       * (2) We can't allocate less than nr_gs_entries_granularity.
 +       */
 +      gs_chunks = ALIGN(MAX2(gs_granularity, 2) * gs_entry_size_bytes,
 +                        chunk_size_bytes) / chunk_size_bytes;
 +      gs_wants =
 +         ALIGN(devinfo->urb.max_gs_entries * gs_entry_size_bytes,
 +               chunk_size_bytes) / chunk_size_bytes - gs_chunks;
 +   }
 +
 +   /* There should always be enough URB space to satisfy the minimum
 +    * requirements of each stage.
 +    */
 +   unsigned total_needs = push_constant_chunks + vs_chunks + gs_chunks;
 +   assert(total_needs <= urb_chunks);
 +
 +   /* Mete out remaining space (if any) in proportion to "wants". */
 +   unsigned total_wants = vs_wants + gs_wants;
 +   unsigned remaining_space = urb_chunks - total_needs;
 +   if (remaining_space > total_wants)
 +      remaining_space = total_wants;
 +   if (remaining_space > 0) {
 +      unsigned vs_additional = (unsigned)
 +         round(vs_wants * (((double) remaining_space) / total_wants));
 +      vs_chunks += vs_additional;
 +      remaining_space -= vs_additional;
 +      gs_chunks += remaining_space;
 +   }
 +
 +   /* Sanity check that we haven't over-allocated. */
 +   assert(push_constant_chunks + vs_chunks + gs_chunks <= urb_chunks);
 +
 +   /* Finally, compute the number of entries that can fit in the space
 +    * allocated to each stage.
 +    */
 +   unsigned nr_vs_entries = vs_chunks * chunk_size_bytes / vs_entry_size_bytes;
 +   unsigned nr_gs_entries = gs_chunks * chunk_size_bytes / gs_entry_size_bytes;
 +
 +   /* Since we rounded up when computing *_wants, this may be slightly more
 +    * than the maximum allowed amount, so correct for that.
 +    */
 +   nr_vs_entries = MIN2(nr_vs_entries, devinfo->urb.max_vs_entries);
 +   nr_gs_entries = MIN2(nr_gs_entries, devinfo->urb.max_gs_entries);
 +
 +   /* Ensure that we program a multiple of the granularity. */
 +   nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, vs_granularity);
 +   nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, gs_granularity);
 +
 +   /* Finally, sanity check to make sure we have at least the minimum number
 +    * of entries needed for each stage.
 +    */
 +   assert(nr_vs_entries >= devinfo->urb.min_vs_entries);
 +   if (gs_present)
 +      assert(nr_gs_entries >= 2);
 +
 +   /* Lay out the URB in the following order:
 +    * - push constants
 +    * - VS
 +    * - GS
 +    */
 +   pipeline->urb.start[MESA_SHADER_VERTEX] = push_constant_chunks;
 +   pipeline->urb.size[MESA_SHADER_VERTEX] = vs_size;
 +   pipeline->urb.entries[MESA_SHADER_VERTEX] = nr_vs_entries;
 +
 +   pipeline->urb.start[MESA_SHADER_GEOMETRY] = push_constant_chunks + vs_chunks;
 +   pipeline->urb.size[MESA_SHADER_GEOMETRY] = gs_size;
 +   pipeline->urb.entries[MESA_SHADER_GEOMETRY] = nr_gs_entries;
 +
 +   pipeline->urb.start[MESA_SHADER_TESS_CTRL] = push_constant_chunks;
 +   pipeline->urb.size[MESA_SHADER_TESS_CTRL] = 1;
 +   pipeline->urb.entries[MESA_SHADER_TESS_CTRL] = 0;
 +
 +   pipeline->urb.start[MESA_SHADER_TESS_EVAL] = push_constant_chunks;
 +   pipeline->urb.size[MESA_SHADER_TESS_EVAL] = 1;
 +   pipeline->urb.entries[MESA_SHADER_TESS_EVAL] = 0;
 +
 +   const unsigned stages =
 +      _mesa_bitcount(pipeline->active_stages & VK_SHADER_STAGE_ALL_GRAPHICS);
 +   unsigned size_per_stage = stages ? (push_constant_kb / stages) : 0;
 +   unsigned used_kb = 0;
 +
 +   /* Broadwell+ and Haswell gt3 require that the push constant sizes be in
 +    * units of 2KB.  Incidentally, these are the same platforms that have
 +    * 32KB worth of push constant space.
 +    */
 +   if (push_constant_kb == 32)
 +      size_per_stage &= ~1u;
 +
 +   for (int i = MESA_SHADER_VERTEX; i < MESA_SHADER_FRAGMENT; i++) {
 +      pipeline->urb.push_size[i] =
 +         (pipeline->active_stages & (1 << i)) ? size_per_stage : 0;
 +      used_kb += pipeline->urb.push_size[i];
 +      assert(used_kb <= push_constant_kb);
 +   }
 +
 +   pipeline->urb.push_size[MESA_SHADER_FRAGMENT] =
 +      push_constant_kb - used_kb;
 +}
 +
 +static void
 +anv_pipeline_init_dynamic_state(struct anv_pipeline *pipeline,
 +                                const VkGraphicsPipelineCreateInfo *pCreateInfo)
 +{
 +   anv_cmd_dirty_mask_t states = ANV_CMD_DIRTY_DYNAMIC_ALL;
 +   ANV_FROM_HANDLE(anv_render_pass, pass, pCreateInfo->renderPass);
 +   struct anv_subpass *subpass = &pass->subpasses[pCreateInfo->subpass];
 +
 +   pipeline->dynamic_state = default_dynamic_state;
 +
 +   if (pCreateInfo->pDynamicState) {
 +      /* Remove all of the states that are marked as dynamic */
 +      uint32_t count = pCreateInfo->pDynamicState->dynamicStateCount;
 +      for (uint32_t s = 0; s < count; s++)
 +         states &= ~(1 << pCreateInfo->pDynamicState->pDynamicStates[s]);
 +   }
 +
 +   struct anv_dynamic_state *dynamic = &pipeline->dynamic_state;
 +
 +   dynamic->viewport.count = pCreateInfo->pViewportState->viewportCount;
 +   if (states & (1 << VK_DYNAMIC_STATE_VIEWPORT)) {
 +      typed_memcpy(dynamic->viewport.viewports,
 +                   pCreateInfo->pViewportState->pViewports,
 +                   pCreateInfo->pViewportState->viewportCount);
 +   }
 +
 +   dynamic->scissor.count = pCreateInfo->pViewportState->scissorCount;
 +   if (states & (1 << VK_DYNAMIC_STATE_SCISSOR)) {
 +      typed_memcpy(dynamic->scissor.scissors,
 +                   pCreateInfo->pViewportState->pScissors,
 +                   pCreateInfo->pViewportState->scissorCount);
 +   }
 +
 +   if (states & (1 << VK_DYNAMIC_STATE_LINE_WIDTH)) {
 +      assert(pCreateInfo->pRasterizationState);
 +      dynamic->line_width = pCreateInfo->pRasterizationState->lineWidth;
 +   }
 +
 +   if (states & (1 << VK_DYNAMIC_STATE_DEPTH_BIAS)) {
 +      assert(pCreateInfo->pRasterizationState);
 +      dynamic->depth_bias.bias =
 +         pCreateInfo->pRasterizationState->depthBiasConstantFactor;
 +      dynamic->depth_bias.clamp =
 +         pCreateInfo->pRasterizationState->depthBiasClamp;
 +      dynamic->depth_bias.slope =
 +         pCreateInfo->pRasterizationState->depthBiasSlopeFactor;
 +   }
 +
 +   if (states & (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS)) {
 +      assert(pCreateInfo->pColorBlendState);
 +      typed_memcpy(dynamic->blend_constants,
 +                   pCreateInfo->pColorBlendState->blendConstants, 4);
 +   }
 +
 +   /* If there is no depthstencil attachment, then don't read
 +    * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
 +    * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
 +    * no need to override the depthstencil defaults in
 +    * anv_pipeline::dynamic_state when there is no depthstencil attachment.
 +    *
 +    * From the Vulkan spec (20 Oct 2015, git-aa308cb):
 +    *
 +    *    pDepthStencilState [...] may only be NULL if renderPass and subpass
 +    *    specify a subpass that has no depth/stencil attachment.
 +    */
 +   if (subpass->depth_stencil_attachment != VK_ATTACHMENT_UNUSED) {
 +      if (states & (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS)) {
 +         assert(pCreateInfo->pDepthStencilState);
 +         dynamic->depth_bounds.min =
 +            pCreateInfo->pDepthStencilState->minDepthBounds;
 +         dynamic->depth_bounds.max =
 +            pCreateInfo->pDepthStencilState->maxDepthBounds;
 +      }
 +
 +      if (states & (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK)) {
 +         assert(pCreateInfo->pDepthStencilState);
 +         dynamic->stencil_compare_mask.front =
 +            pCreateInfo->pDepthStencilState->front.compareMask;
 +         dynamic->stencil_compare_mask.back =
 +            pCreateInfo->pDepthStencilState->back.compareMask;
 +      }
 +
 +      if (states & (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK)) {
 +         assert(pCreateInfo->pDepthStencilState);
 +         dynamic->stencil_write_mask.front =
 +            pCreateInfo->pDepthStencilState->front.writeMask;
 +         dynamic->stencil_write_mask.back =
 +            pCreateInfo->pDepthStencilState->back.writeMask;
 +      }
 +
 +      if (states & (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE)) {
 +         assert(pCreateInfo->pDepthStencilState);
 +         dynamic->stencil_reference.front =
 +            pCreateInfo->pDepthStencilState->front.reference;
 +         dynamic->stencil_reference.back =
 +            pCreateInfo->pDepthStencilState->back.reference;
 +      }
 +   }
 +
 +   pipeline->dynamic_state_mask = states;
 +}
 +
 +static void
 +anv_pipeline_validate_create_info(const VkGraphicsPipelineCreateInfo *info)
 +{
 +   struct anv_render_pass *renderpass = NULL;
 +   struct anv_subpass *subpass = NULL;
 +
 +   /* Assert that all required members of VkGraphicsPipelineCreateInfo are
 +    * present, as explained by the Vulkan (20 Oct 2015, git-aa308cb), Section
 +    * 4.2 Graphics Pipeline.
 +    */
 +   assert(info->sType == VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO);
 +
 +   renderpass = anv_render_pass_from_handle(info->renderPass);
 +   assert(renderpass);
 +
 +   if (renderpass != &anv_meta_dummy_renderpass) {
 +      assert(info->subpass < renderpass->subpass_count);
 +      subpass = &renderpass->subpasses[info->subpass];
 +   }
 +
 +   assert(info->stageCount >= 1);
 +   assert(info->pVertexInputState);
 +   assert(info->pInputAssemblyState);
 +   assert(info->pViewportState);
 +   assert(info->pRasterizationState);
 +
 +   if (subpass && subpass->depth_stencil_attachment != VK_ATTACHMENT_UNUSED)
 +      assert(info->pDepthStencilState);
 +
 +   if (subpass && subpass->color_count > 0)
 +      assert(info->pColorBlendState);
 +
 +   for (uint32_t i = 0; i < info->stageCount; ++i) {
 +      switch (info->pStages[i].stage) {
 +      case VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT:
 +      case VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT:
 +         assert(info->pTessellationState);
 +         break;
 +      default:
 +         break;
 +      }
 +   }
 +}
 +
 +VkResult
 +anv_pipeline_init(struct anv_pipeline *pipeline,
 +                  struct anv_device *device,
 +                  struct anv_pipeline_cache *cache,
 +                  const VkGraphicsPipelineCreateInfo *pCreateInfo,
 +                  const struct anv_graphics_pipeline_create_info *extra,
 +                  const VkAllocationCallbacks *alloc)
 +{
 +   VkResult result;
 +
 +   anv_validate {
 +      anv_pipeline_validate_create_info(pCreateInfo);
 +   }
 +
 +   if (alloc == NULL)
 +      alloc = &device->alloc;
 +
 +   pipeline->device = device;
 +   pipeline->layout = anv_pipeline_layout_from_handle(pCreateInfo->layout);
 +
 +   result = anv_reloc_list_init(&pipeline->batch_relocs, alloc);
 +   if (result != VK_SUCCESS)
 +      return result;
 +
 +   pipeline->batch.alloc = alloc;
 +   pipeline->batch.next = pipeline->batch.start = pipeline->batch_data;
 +   pipeline->batch.end = pipeline->batch.start + sizeof(pipeline->batch_data);
 +   pipeline->batch.relocs = &pipeline->batch_relocs;
 +
 +   anv_pipeline_init_dynamic_state(pipeline, pCreateInfo);
 +
 +   pipeline->use_repclear = extra && extra->use_repclear;
 +
 +   /* When we free the pipeline, we detect stages based on the NULL status
 +    * of various prog_data pointers.  Make them NULL by default.
 +    */
 +   memset(pipeline->prog_data, 0, sizeof(pipeline->prog_data));
 +   memset(pipeline->scratch_start, 0, sizeof(pipeline->scratch_start));
 +   memset(pipeline->bindings, 0, sizeof(pipeline->bindings));
 +
 +   pipeline->vs_simd8 = NO_KERNEL;
 +   pipeline->vs_vec4 = NO_KERNEL;
 +   pipeline->gs_kernel = NO_KERNEL;
 +   pipeline->ps_ksp0 = NO_KERNEL;
 +
 +   pipeline->active_stages = 0;
 +   pipeline->total_scratch = 0;
 +
 +   const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, };
 +   struct anv_shader_module *modules[MESA_SHADER_STAGES] = { 0, };
 +   for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) {
 +      gl_shader_stage stage = ffs(pCreateInfo->pStages[i].stage) - 1;
 +      pStages[stage] = &pCreateInfo->pStages[i];
 +      modules[stage] = anv_shader_module_from_handle(pStages[stage]->module);
 +   }
 +
 +   if (modules[MESA_SHADER_VERTEX]) {
 +      anv_pipeline_compile_vs(pipeline, cache, pCreateInfo,
 +                              modules[MESA_SHADER_VERTEX],
 +                              pStages[MESA_SHADER_VERTEX]->pName,
 +                              pStages[MESA_SHADER_VERTEX]->pSpecializationInfo);
 +   }
 +
 +   if (modules[MESA_SHADER_TESS_CTRL] || modules[MESA_SHADER_TESS_EVAL])
 +      anv_finishme("no tessellation support");
 +
 +   if (modules[MESA_SHADER_GEOMETRY]) {
 +      anv_pipeline_compile_gs(pipeline, cache, pCreateInfo,
 +                              modules[MESA_SHADER_GEOMETRY],
 +                              pStages[MESA_SHADER_GEOMETRY]->pName,
 +                              pStages[MESA_SHADER_GEOMETRY]->pSpecializationInfo);
 +   }
 +
 +   if (modules[MESA_SHADER_FRAGMENT]) {
 +      anv_pipeline_compile_fs(pipeline, cache, pCreateInfo, extra,
 +                              modules[MESA_SHADER_FRAGMENT],
 +                              pStages[MESA_SHADER_FRAGMENT]->pName,
 +                              pStages[MESA_SHADER_FRAGMENT]->pSpecializationInfo);
 +   }
 +
 +   if (!(pipeline->active_stages & VK_SHADER_STAGE_VERTEX_BIT)) {
 +      /* Vertex is only optional if disable_vs is set */
 +      assert(extra->disable_vs);
 +   }
 +
 +   gen7_compute_urb_partition(pipeline);
 +
 +   const VkPipelineVertexInputStateCreateInfo *vi_info =
 +      pCreateInfo->pVertexInputState;
 +
 +   uint64_t inputs_read;
 +   if (extra && extra->disable_vs) {
 +      /* If the VS is disabled, just assume the user knows what they're
 +       * doing and apply the layout blindly.  This can only come from
 +       * meta, so this *should* be safe.
 +       */
 +      inputs_read = ~0ull;
 +   } else {
 +      inputs_read = get_vs_prog_data(pipeline)->inputs_read;
 +   }
 +
 +   pipeline->vb_used = 0;
 +   for (uint32_t i = 0; i < vi_info->vertexAttributeDescriptionCount; i++) {
 +      const VkVertexInputAttributeDescription *desc =
 +         &vi_info->pVertexAttributeDescriptions[i];
 +
 +      if (inputs_read & (1 << (VERT_ATTRIB_GENERIC0 + desc->location)))
 +         pipeline->vb_used |= 1 << desc->binding;
 +   }
 +
 +   for (uint32_t i = 0; i < vi_info->vertexBindingDescriptionCount; i++) {
 +      const VkVertexInputBindingDescription *desc =
 +         &vi_info->pVertexBindingDescriptions[i];
 +
 +      pipeline->binding_stride[desc->binding] = desc->stride;
 +
 +      /* Step rate is programmed per vertex element (attribute), not
 +       * binding. Set up a map of which bindings step per instance, for
 +       * reference by vertex element setup. */
 +      switch (desc->inputRate) {
 +      default:
 +      case VK_VERTEX_INPUT_RATE_VERTEX:
 +         pipeline->instancing_enable[desc->binding] = false;
 +         break;
 +      case VK_VERTEX_INPUT_RATE_INSTANCE:
 +         pipeline->instancing_enable[desc->binding] = true;
 +         break;
 +      }
 +   }
 +
 +   const VkPipelineInputAssemblyStateCreateInfo *ia_info =
 +      pCreateInfo->pInputAssemblyState;
 +   pipeline->primitive_restart = ia_info->primitiveRestartEnable;
 +   pipeline->topology = vk_to_gen_primitive_type[ia_info->topology];
 +
 +   if (extra && extra->use_rectlist)
 +      pipeline->topology = _3DPRIM_RECTLIST;
 +
 +   while (anv_block_pool_size(&device->scratch_block_pool) <
 +          pipeline->total_scratch)
 +      anv_block_pool_alloc(&device->scratch_block_pool);
 +
 +   return VK_SUCCESS;
 +}
 +
 +VkResult
 +anv_graphics_pipeline_create(
 +   VkDevice _device,
 +   VkPipelineCache _cache,
 +   const VkGraphicsPipelineCreateInfo *pCreateInfo,
 +   const struct anv_graphics_pipeline_create_info *extra,
 +   const VkAllocationCallbacks *pAllocator,
 +   VkPipeline *pPipeline)
 +{
 +   ANV_FROM_HANDLE(anv_device, device, _device);
 +   ANV_FROM_HANDLE(anv_pipeline_cache, cache, _cache);
 +
 +   if (cache == NULL)
 +      cache = &device->default_pipeline_cache;
 +
 +   switch (device->info.gen) {
 +   case 7:
 +      if (device->info.is_haswell)
 +         return gen75_graphics_pipeline_create(_device, cache, pCreateInfo, extra, pAllocator, pPipeline);
 +      else
 +         return gen7_graphics_pipeline_create(_device, cache, pCreateInfo, extra, pAllocator, pPipeline);
 +   case 8:
 +      return gen8_graphics_pipeline_create(_device, cache, pCreateInfo, extra, pAllocator, pPipeline);
 +   case 9:
 +      return gen9_graphics_pipeline_create(_device, cache, pCreateInfo, extra, pAllocator, pPipeline);
 +   default:
 +      unreachable("unsupported gen\n");
 +   }
 +}
 +
 +VkResult anv_CreateGraphicsPipelines(
 +    VkDevice                                    _device,
 +    VkPipelineCache                             pipelineCache,
 +    uint32_t                                    count,
 +    const VkGraphicsPipelineCreateInfo*         pCreateInfos,
 +    const VkAllocationCallbacks*                pAllocator,
 +    VkPipeline*                                 pPipelines)
 +{
 +   VkResult result = VK_SUCCESS;
 +
 +   unsigned i = 0;
 +   for (; i < count; i++) {
 +      result = anv_graphics_pipeline_create(_device,
 +                                            pipelineCache,
 +                                            &pCreateInfos[i],
 +                                            NULL, pAllocator, &pPipelines[i]);
 +      if (result != VK_SUCCESS) {
 +         for (unsigned j = 0; j < i; j++) {
 +            anv_DestroyPipeline(_device, pPipelines[j], pAllocator);
 +         }
 +
 +         return result;
 +      }
 +   }
 +
 +   return VK_SUCCESS;
 +}
 +
 +static VkResult anv_compute_pipeline_create(
 +    VkDevice                                    _device,
 +    VkPipelineCache                             _cache,
 +    const VkComputePipelineCreateInfo*          pCreateInfo,
 +    const VkAllocationCallbacks*                pAllocator,
 +    VkPipeline*                                 pPipeline)
 +{
 +   ANV_FROM_HANDLE(anv_device, device, _device);
 +   ANV_FROM_HANDLE(anv_pipeline_cache, cache, _cache);
 +
 +   if (cache == NULL)
 +      cache = &device->default_pipeline_cache;
 +
 +   switch (device->info.gen) {
 +   case 7:
 +      if (device->info.is_haswell)
 +         return gen75_compute_pipeline_create(_device, cache, pCreateInfo, pAllocator, pPipeline);
 +      else
 +         return gen7_compute_pipeline_create(_device, cache, pCreateInfo, pAllocator, pPipeline);
 +   case 8:
 +      return gen8_compute_pipeline_create(_device, cache, pCreateInfo, pAllocator, pPipeline);
 +   case 9:
 +      return gen9_compute_pipeline_create(_device, cache, pCreateInfo, pAllocator, pPipeline);
 +   default:
 +      unreachable("unsupported gen\n");
 +   }
 +}
 +
 +VkResult anv_CreateComputePipelines(
 +    VkDevice                                    _device,
 +    VkPipelineCache                             pipelineCache,
 +    uint32_t                                    count,
 +    const VkComputePipelineCreateInfo*          pCreateInfos,
 +    const VkAllocationCallbacks*                pAllocator,
 +    VkPipeline*                                 pPipelines)
 +{
 +   VkResult result = VK_SUCCESS;
 +
 +   unsigned i = 0;
 +   for (; i < count; i++) {
 +      result = anv_compute_pipeline_create(_device, pipelineCache,
 +                                           &pCreateInfos[i],
 +                                           pAllocator, &pPipelines[i]);
 +      if (result != VK_SUCCESS) {
 +         for (unsigned j = 0; j < i; j++) {
 +            anv_DestroyPipeline(_device, pPipelines[j], pAllocator);
 +         }
 +
 +         return result;
 +      }
 +   }
 +
 +   return VK_SUCCESS;
 +}
Simple merge