spirv: Tell which opcode or value is unhandled when failing
[mesa.git] / src / compiler / spirv / spirv_to_nir.c
index 3de45c47371e24ceaffdf46a0822345f4e6dec9c..b297a4c2e4adff38e27d5572d31b62b8bf0946d5 100644 (file)
 #include "nir/nir_vla.h"
 #include "nir/nir_control_flow.h"
 #include "nir/nir_constant_expressions.h"
+#include "nir/nir_deref.h"
 #include "spirv_info.h"
 
+#include "util/u_math.h"
+
 #include <stdio.h>
 
 void
@@ -129,6 +132,18 @@ _vtn_warn(struct vtn_builder *b, const char *file, unsigned line,
    va_end(args);
 }
 
+void
+_vtn_err(struct vtn_builder *b, const char *file, unsigned line,
+          const char *fmt, ...)
+{
+   va_list args;
+
+   va_start(args, fmt);
+   vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_ERROR, "SPIR-V ERROR:\n",
+               file, line, fmt, args);
+   va_end(args);
+}
+
 void
 _vtn_fail(struct vtn_builder *b, const char *file, unsigned line,
           const char *fmt, ...)
@@ -207,6 +222,8 @@ vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
    case GLSL_TYPE_UINT:
    case GLSL_TYPE_INT16:
    case GLSL_TYPE_UINT16:
+   case GLSL_TYPE_UINT8:
+   case GLSL_TYPE_INT8:
    case GLSL_TYPE_INT64:
    case GLSL_TYPE_UINT64:
    case GLSL_TYPE_BOOL:
@@ -219,7 +236,8 @@ vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
          nir_load_const_instr *load =
             nir_load_const_instr_create(b->shader, num_components, bit_size);
 
-         load->value = constant->values[0];
+         memcpy(load->value, constant->values[0],
+                sizeof(nir_const_value) * load->def.num_components);
 
          nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
          val->def = &load->def;
@@ -235,7 +253,8 @@ vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
             nir_load_const_instr *load =
                nir_load_const_instr_create(b->shader, rows, bit_size);
 
-            load->value = constant->values[i];
+            memcpy(load->value, constant->values[i],
+                   sizeof(nir_const_value) * load->def.num_components);
 
             nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr);
             col_val->def = &load->def;
@@ -368,16 +387,22 @@ static void
 vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
                      const uint32_t *w, unsigned count)
 {
+   const char *ext = (const char *)&w[2];
    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) {
+      if (strcmp(ext, "GLSL.std.450") == 0) {
          val->ext_handler = vtn_handle_glsl450_instruction;
-      } else if ((strcmp((const char *)&w[2], "SPV_AMD_gcn_shader") == 0)
-                && (b->options && b->options->exts.AMD_gcn_shader)) {
+      } else if ((strcmp(ext, "SPV_AMD_gcn_shader") == 0)
+                && (b->options && b->options->caps.gcn_shader)) {
          val->ext_handler = vtn_handle_amd_gcn_shader_instruction;
+      } else if ((strcmp(ext, "SPV_AMD_shader_trinary_minmax") == 0)
+                && (b->options && b->options->caps.trinary_minmax)) {
+         val->ext_handler = vtn_handle_amd_shader_trinary_minmax_instruction;
+      } else if (strcmp(ext, "OpenCL.std") == 0) {
+         val->ext_handler = vtn_handle_opencl_instruction;
       } else {
-         vtn_fail("Unsupported extension");
+         vtn_fail("Unsupported extension: %s", ext);
       }
       break;
    }
@@ -390,7 +415,7 @@ vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
    }
 
    default:
-      vtn_fail("Unhandled opcode");
+      vtn_fail_with_opcode("Unhandled opcode", opcode);
    }
 }
 
@@ -461,7 +486,7 @@ vtn_foreach_execution_mode(struct vtn_builder *b, struct vtn_value *value,
    }
 }
 
-static void
+void
 vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
                       const uint32_t *w, unsigned count)
 {
@@ -476,20 +501,26 @@ vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
 
    case SpvOpDecorate:
    case SpvOpMemberDecorate:
-   case SpvOpExecutionMode: {
+   case SpvOpDecorateStringGOOGLE:
+   case SpvOpMemberDecorateStringGOOGLE:
+   case SpvOpExecutionMode:
+   case SpvOpExecutionModeId: {
       struct vtn_value *val = vtn_untyped_value(b, target);
 
       struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
       switch (opcode) {
       case SpvOpDecorate:
+      case SpvOpDecorateStringGOOGLE:
          dec->scope = VTN_DEC_DECORATION;
          break;
       case SpvOpMemberDecorate:
+      case SpvOpMemberDecorateStringGOOGLE:
          dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(w++);
          vtn_fail_if(dec->scope < VTN_DEC_STRUCT_MEMBER0, /* overflow */
                      "Member argument of OpMemberDecorate too large");
          break;
       case SpvOpExecutionMode:
+      case SpvOpExecutionModeId:
          dec->scope = VTN_DEC_EXECUTION_MODE;
          break;
       default:
@@ -540,6 +571,29 @@ struct member_decoration_ctx {
    struct vtn_type *type;
 };
 
+/**
+ * Returns true if the given type contains a struct decorated Block or
+ * BufferBlock
+ */
+bool
+vtn_type_contains_block(struct vtn_builder *b, struct vtn_type *type)
+{
+   switch (type->base_type) {
+   case vtn_base_type_array:
+      return vtn_type_contains_block(b, type->array_element);
+   case vtn_base_type_struct:
+      if (type->block || type->buffer_block)
+         return true;
+      for (unsigned i = 0; i < type->length; i++) {
+         if (vtn_type_contains_block(b, type->members[i]))
+            return true;
+      }
+      return false;
+   default:
+      return false;
+   }
+}
+
 /** Returns true if two types are "compatible", i.e. you can do an OpLoad,
  * OpStore, or OpCopyMemory between them without breaking anything.
  * Technically, the SPIR-V rules require the exact same type ID but this lets
@@ -649,6 +703,29 @@ mutable_matrix_member(struct vtn_builder *b, struct vtn_type *type, int member)
    return type;
 }
 
+static void
+vtn_handle_access_qualifier(struct vtn_builder *b, struct vtn_type *type,
+                            int member, enum gl_access_qualifier access)
+{
+   type->members[member] = vtn_type_copy(b, type->members[member]);
+   type = type->members[member];
+
+   type->access |= access;
+}
+
+static void
+array_stride_decoration_cb(struct vtn_builder *b,
+                           struct vtn_value *val, int member,
+                           const struct vtn_decoration *dec, void *void_ctx)
+{
+   struct vtn_type *type = val->type;
+
+   if (dec->decoration == SpvDecorationArrayStride) {
+      vtn_fail_if(dec->literals[0] == 0, "ArrayStride must be non-zero");
+      type->stride = dec->literals[0];
+   }
+}
+
 static void
 struct_member_decoration_cb(struct vtn_builder *b,
                             struct vtn_value *val, int member,
@@ -662,13 +739,21 @@ struct_member_decoration_cb(struct vtn_builder *b,
    assert(member < ctx->num_fields);
 
    switch (dec->decoration) {
+   case SpvDecorationRelaxedPrecision:
+   case SpvDecorationUniform:
+      break; /* FIXME: Do nothing with this for now. */
    case SpvDecorationNonWritable:
+      vtn_handle_access_qualifier(b, ctx->type, member, ACCESS_NON_WRITEABLE);
+      break;
    case SpvDecorationNonReadable:
-   case SpvDecorationRelaxedPrecision:
+      vtn_handle_access_qualifier(b, ctx->type, member, ACCESS_NON_READABLE);
+      break;
    case SpvDecorationVolatile:
+      vtn_handle_access_qualifier(b, ctx->type, member, ACCESS_VOLATILE);
+      break;
    case SpvDecorationCoherent:
-   case SpvDecorationUniform:
-      break; /* FIXME: Do nothing with this for now. */
+      vtn_handle_access_qualifier(b, ctx->type, member, ACCESS_COHERENT);
+      break;
    case SpvDecorationNoPerspective:
       ctx->fields[member].interpolation = INTERP_MODE_NOPERSPECTIVE;
       break;
@@ -698,6 +783,7 @@ struct_member_decoration_cb(struct vtn_builder *b,
       break;
    case SpvDecorationOffset:
       ctx->type->offsets[member] = dec->literals[0];
+      ctx->fields[member].offset = dec->literals[0];
       break;
    case SpvDecorationMatrixStride:
       /* Handled as a second pass */
@@ -737,20 +823,48 @@ struct_member_decoration_cb(struct vtn_builder *b,
       break;
 
    case SpvDecorationCPacked:
+      if (b->shader->info.stage != MESA_SHADER_KERNEL)
+         vtn_warn("Decoration only allowed for CL-style kernels: %s",
+                  spirv_decoration_to_string(dec->decoration));
+      else
+         ctx->type->packed = true;
+      break;
+
    case SpvDecorationSaturatedConversion:
    case SpvDecorationFuncParamAttr:
    case SpvDecorationFPRoundingMode:
    case SpvDecorationFPFastMathMode:
    case SpvDecorationAlignment:
-      vtn_warn("Decoration only allowed for CL-style kernels: %s",
-               spirv_decoration_to_string(dec->decoration));
+      if (b->shader->info.stage != MESA_SHADER_KERNEL) {
+         vtn_warn("Decoration only allowed for CL-style kernels: %s",
+                  spirv_decoration_to_string(dec->decoration));
+      }
+      break;
+
+   case SpvDecorationHlslSemanticGOOGLE:
+      /* HLSL semantic decorations can safely be ignored by the driver. */
       break;
 
    default:
-      vtn_fail("Unhandled decoration");
+      vtn_fail_with_decoration("Unhandled decoration", dec->decoration);
    }
 }
 
+/** Chases the array type all the way down to the tail and rewrites the
+ * glsl_types to be based off the tail's glsl_type.
+ */
+static void
+vtn_array_type_rewrite_glsl_type(struct vtn_type *type)
+{
+   if (type->base_type != vtn_base_type_array)
+      return;
+
+   vtn_array_type_rewrite_glsl_type(type->array_element);
+
+   type->type = glsl_array_type(type->array_element->type,
+                                type->length, type->stride);
+}
+
 /* Matrix strides are handled as a separate pass because we need to know
  * whether the matrix is row-major or not first.
  */
@@ -766,6 +880,7 @@ struct_member_matrix_stride_cb(struct vtn_builder *b,
    vtn_fail_if(member < 0,
                "The MatrixStride decoration is only allowed on members "
                "of OpTypeStruct");
+   vtn_fail_if(dec->literals[0] == 0, "MatrixStride must be non-zero");
 
    struct member_decoration_ctx *ctx = void_ctx;
 
@@ -774,10 +889,39 @@ struct_member_matrix_stride_cb(struct vtn_builder *b,
       mat_type->array_element = vtn_type_copy(b, mat_type->array_element);
       mat_type->stride = mat_type->array_element->stride;
       mat_type->array_element->stride = dec->literals[0];
+
+      mat_type->type = glsl_explicit_matrix_type(mat_type->type,
+                                                 dec->literals[0], true);
+      mat_type->array_element->type = glsl_get_column_type(mat_type->type);
    } else {
       vtn_assert(mat_type->array_element->stride > 0);
       mat_type->stride = dec->literals[0];
+
+      mat_type->type = glsl_explicit_matrix_type(mat_type->type,
+                                                 dec->literals[0], false);
    }
+
+   /* Now that we've replaced the glsl_type with a properly strided matrix
+    * type, rewrite the member type so that it's an array of the proper kind
+    * of glsl_type.
+    */
+   vtn_array_type_rewrite_glsl_type(ctx->type->members[member]);
+   ctx->fields[member].type = ctx->type->members[member]->type;
+}
+
+static void
+struct_block_decoration_cb(struct vtn_builder *b,
+                           struct vtn_value *val, int member,
+                           const struct vtn_decoration *dec, void *ctx)
+{
+   if (member != -1)
+      return;
+
+   struct vtn_type *type = val->type;
+   if (dec->decoration == SpvDecorationBlock)
+      type->block = true;
+   else if (dec->decoration == SpvDecorationBufferBlock)
+      type->buffer_block = true;
 }
 
 static void
@@ -796,18 +940,16 @@ type_decoration_cb(struct vtn_builder *b,
 
    switch (dec->decoration) {
    case SpvDecorationArrayStride:
-      vtn_assert(type->base_type == vtn_base_type_matrix ||
-                 type->base_type == vtn_base_type_array ||
+      vtn_assert(type->base_type == vtn_base_type_array ||
                  type->base_type == vtn_base_type_pointer);
-      type->stride = dec->literals[0];
       break;
    case SpvDecorationBlock:
       vtn_assert(type->base_type == vtn_base_type_struct);
-      type->block = true;
+      vtn_assert(type->block);
       break;
    case SpvDecorationBufferBlock:
       vtn_assert(type->base_type == vtn_base_type_struct);
-      type->buffer_block = true;
+      vtn_assert(type->buffer_block);
       break;
    case SpvDecorationGLSLShared:
    case SpvDecorationGLSLPacked:
@@ -828,16 +970,24 @@ type_decoration_cb(struct vtn_builder *b,
    case SpvDecorationNonWritable:
    case SpvDecorationNonReadable:
    case SpvDecorationUniform:
-   case SpvDecorationStream:
    case SpvDecorationLocation:
    case SpvDecorationComponent:
    case SpvDecorationOffset:
    case SpvDecorationXfbBuffer:
    case SpvDecorationXfbStride:
+   case SpvDecorationHlslSemanticGOOGLE:
       vtn_warn("Decoration only allowed for struct members: %s",
                spirv_decoration_to_string(dec->decoration));
       break;
 
+   case SpvDecorationStream:
+      /* We don't need to do anything here, as stream is filled up when
+       * aplying the decoration to a variable, just check that if it is not a
+       * struct member, it should be a struct.
+       */
+      vtn_assert(type->base_type == vtn_base_type_struct);
+      break;
+
    case SpvDecorationRelaxedPrecision:
    case SpvDecorationSpecId:
    case SpvDecorationInvariant:
@@ -855,6 +1005,13 @@ type_decoration_cb(struct vtn_builder *b,
       break;
 
    case SpvDecorationCPacked:
+      if (b->shader->info.stage != MESA_SHADER_KERNEL)
+         vtn_warn("Decoration only allowed for CL-style kernels: %s",
+                  spirv_decoration_to_string(dec->decoration));
+      else
+         type->packed = true;
+      break;
+
    case SpvDecorationSaturatedConversion:
    case SpvDecorationFuncParamAttr:
    case SpvDecorationFPRoundingMode:
@@ -865,7 +1022,7 @@ type_decoration_cb(struct vtn_builder *b,
       break;
 
    default:
-      vtn_fail("Unhandled decoration");
+      vtn_fail_with_decoration("Unhandled decoration", dec->decoration);
    }
 }
 
@@ -914,7 +1071,8 @@ translate_image_format(struct vtn_builder *b, SpvImageFormat format)
    case SpvImageFormatR16ui:        return 0x8234; /* GL_R16UI */
    case SpvImageFormatR8ui:         return 0x8232; /* GL_R8UI */
    default:
-      vtn_fail("Invalid image format");
+      vtn_fail("Invalid image format: %s (%u)",
+               spirv_imageformat_to_string(format), format);
    }
 }
 
@@ -924,15 +1082,16 @@ vtn_type_layout_std430(struct vtn_builder *b, struct vtn_type *type,
 {
    switch (type->base_type) {
    case vtn_base_type_scalar: {
-      uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
+      uint32_t comp_size = glsl_type_is_boolean(type->type)
+         ? 4 : glsl_get_bit_size(type->type) / 8;
       *size_out = comp_size;
       *align_out = comp_size;
       return type;
    }
 
    case vtn_base_type_vector: {
-      uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
-      assert(type->length > 0 && type->length <= 4);
+      uint32_t comp_size = glsl_type_is_boolean(type->type)
+         ? 4 : glsl_get_bit_size(type->type) / 8;
       unsigned align_comps = type->length == 3 ? 4 : type->length;
       *size_out = comp_size * type->length,
       *align_out = comp_size * align_comps;
@@ -980,10 +1139,18 @@ 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);
+   struct vtn_value *val = NULL;
 
-   val->type = rzalloc(b, struct vtn_type);
-   val->type->id = w[1];
+   /* In order to properly handle forward declarations, we have to defer
+    * allocation for pointer types.
+    */
+   if (opcode != SpvOpTypePointer && opcode != SpvOpTypeForwardPointer) {
+      val = vtn_push_value(b, w[1], vtn_value_type_type);
+      vtn_fail_if(val->type != NULL,
+                  "Only pointers can have forward declarations");
+      val->type = rzalloc(b, struct vtn_type);
+      val->type->id = w[1];
+   }
 
    switch (opcode) {
    case SpvOpTypeVoid:
@@ -1009,8 +1176,11 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
       case 16:
          val->type->type = (signedness ? glsl_int16_t_type() : glsl_uint16_t_type());
          break;
+      case 8:
+         val->type->type = (signedness ? glsl_int8_t_type() : glsl_uint8_t_type());
+         break;
       default:
-         vtn_fail("Invalid int bit size");
+         vtn_fail("Invalid int bit size: %u", bit_size);
       }
       val->type->length = 1;
       break;
@@ -1030,7 +1200,7 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
          val->type->type = glsl_double_type();
          break;
       default:
-         vtn_fail("Invalid float bit size");
+         vtn_fail("Invalid float bit size: %u", bit_size);
       }
       val->type->length = 1;
       break;
@@ -1042,13 +1212,14 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
 
       vtn_fail_if(base->base_type != vtn_base_type_scalar,
                   "Base type for OpTypeVector must be a scalar");
-      vtn_fail_if(elems < 2 || elems > 4,
+      vtn_fail_if((elems < 2 || elems > 4) && (elems != 8) && (elems != 16),
                   "Invalid component count for OpTypeVector");
 
       val->type->base_type = vtn_base_type_vector;
       val->type->type = glsl_vector_type(glsl_get_base_type(base->type), elems);
       val->type->length = elems;
-      val->type->stride = glsl_get_bit_size(base->type) / 8;
+      val->type->stride = glsl_type_is_boolean(val->type->type)
+         ? 4 : glsl_get_bit_size(base->type) / 8;
       val->type->array_element = base;
       break;
    }
@@ -1086,13 +1257,17 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
          val->type->length = 0;
       } else {
          val->type->length =
-            vtn_value(b, w[3], vtn_value_type_constant)->constant->values[0].u32[0];
+            vtn_value(b, w[3], vtn_value_type_constant)->constant->values[0][0].u32;
       }
 
       val->type->base_type = vtn_base_type_array;
-      val->type->type = glsl_array_type(array_element->type, val->type->length);
       val->type->array_element = array_element;
-      val->type->stride = 0;
+      if (b->shader->info.stage == MESA_SHADER_KERNEL)
+         val->type->stride = glsl_get_cl_size(array_element->type);
+
+      vtn_foreach_decoration(b, val, array_stride_decoration_cb, NULL);
+      val->type->type = glsl_array_type(array_element->type, val->type->length,
+                                        val->type->stride);
       break;
    }
 
@@ -1102,6 +1277,7 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
       val->type->length = num_fields;
       val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
       val->type->offsets = ralloc_array(b, unsigned, num_fields);
+      val->type->packed = false;
 
       NIR_VLA(struct glsl_struct_field, fields, count);
       for (unsigned i = 0; i < num_fields; i++) {
@@ -1111,9 +1287,19 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
             .type = val->type->members[i]->type,
             .name = ralloc_asprintf(b, "field%d", i),
             .location = -1,
+            .offset = -1,
          };
       }
 
+      if (b->shader->info.stage == MESA_SHADER_KERNEL) {
+         unsigned offset = 0;
+         for (unsigned i = 0; i < num_fields; i++) {
+            offset = align(offset, glsl_get_cl_alignment(fields[i].type));
+            fields[i].offset = offset;
+            offset += glsl_get_cl_size(fields[i].type);
+         }
+      }
+
       struct member_decoration_ctx ctx = {
          .num_fields = num_fields,
          .fields = fields,
@@ -1123,9 +1309,21 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
       vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
       vtn_foreach_decoration(b, val, struct_member_matrix_stride_cb, &ctx);
 
-      const char *name = val->name ? val->name : "struct";
+      vtn_foreach_decoration(b, val, struct_block_decoration_cb, NULL);
+
+      const char *name = val->name;
 
-      val->type->type = glsl_struct_type(fields, num_fields, name);
+      if (val->type->block || val->type->buffer_block) {
+         /* Packing will be ignored since types coming from SPIR-V are
+          * explicitly laid out.
+          */
+         val->type->type = glsl_interface_type(fields, num_fields,
+                                               /* packing */ 0, false,
+                                               name ? name : "block");
+      } else {
+         val->type->type = glsl_struct_type(fields, num_fields,
+                                            name ? name : "struct", false);
+      }
       break;
    }
 
@@ -1145,34 +1343,93 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
       break;
    }
 
-   case SpvOpTypePointer: {
+   case SpvOpTypePointer:
+   case SpvOpTypeForwardPointer: {
+      /* We can't blindly push the value because it might be a forward
+       * declaration.
+       */
+      val = vtn_untyped_value(b, w[1]);
+
       SpvStorageClass storage_class = w[2];
-      struct vtn_type *deref_type =
-         vtn_value(b, w[3], vtn_value_type_type)->type;
 
-      val->type->base_type = vtn_base_type_pointer;
-      val->type->storage_class = storage_class;
-      val->type->deref = deref_type;
+      if (val->value_type == vtn_value_type_invalid) {
+         val->value_type = vtn_value_type_type;
+         val->type = rzalloc(b, struct vtn_type);
+         val->type->id = w[1];
+         val->type->base_type = vtn_base_type_pointer;
+         val->type->storage_class = storage_class;
 
-      if (storage_class == SpvStorageClassUniform ||
-          storage_class == SpvStorageClassStorageBuffer) {
          /* These can actually be stored to nir_variables and used as SSA
           * values so they need a real glsl_type.
           */
-         val->type->type = glsl_vector_type(GLSL_TYPE_UINT, 2);
+         switch (storage_class) {
+         case SpvStorageClassUniform:
+            val->type->type = b->options->ubo_ptr_type;
+            break;
+         case SpvStorageClassStorageBuffer:
+            val->type->type = b->options->ssbo_ptr_type;
+            break;
+         case SpvStorageClassPhysicalStorageBufferEXT:
+            val->type->type = b->options->phys_ssbo_ptr_type;
+            break;
+         case SpvStorageClassPushConstant:
+            val->type->type = b->options->push_const_ptr_type;
+            break;
+         case SpvStorageClassWorkgroup:
+            val->type->type = b->options->shared_ptr_type;
+            break;
+         case SpvStorageClassCrossWorkgroup:
+            val->type->type = b->options->global_ptr_type;
+            break;
+         case SpvStorageClassFunction:
+            if (b->physical_ptrs)
+               val->type->type = b->options->temp_ptr_type;
+            break;
+         default:
+            /* In this case, no variable pointers are allowed so all deref
+             * chains are complete back to the variable and it doesn't matter
+             * what type gets used so we leave it NULL.
+             */
+            break;
+         }
+      } else {
+         vtn_fail_if(val->type->storage_class != storage_class,
+                     "The storage classes of an OpTypePointer and any "
+                     "OpTypeForwardPointers that provide forward "
+                     "declarations of it must match.");
       }
 
-      if (storage_class == SpvStorageClassWorkgroup &&
-          b->options->lower_workgroup_access_to_offsets) {
-         uint32_t size, align;
-         val->type->deref = vtn_type_layout_std430(b, val->type->deref,
-                                                   &size, &align);
-         val->type->length = size;
-         val->type->align = align;
-         /* These can actually be stored to nir_variables and used as SSA
-          * values so they need a real glsl_type.
-          */
-         val->type->type = glsl_uint_type();
+      if (opcode == SpvOpTypePointer) {
+         vtn_fail_if(val->type->deref != NULL,
+                     "While OpTypeForwardPointer can be used to provide a "
+                     "forward declaration of a pointer, OpTypePointer can "
+                     "only be used once for a given id.");
+
+         val->type->deref = vtn_value(b, w[3], vtn_value_type_type)->type;
+
+         vtn_foreach_decoration(b, val, array_stride_decoration_cb, NULL);
+
+         if (b->physical_ptrs) {
+            switch (storage_class) {
+            case SpvStorageClassFunction:
+            case SpvStorageClassWorkgroup:
+            case SpvStorageClassCrossWorkgroup:
+               val->type->stride = align(glsl_get_cl_size(val->type->deref->type),
+                                         glsl_get_cl_alignment(val->type->deref->type));
+               break;
+            default:
+               break;
+            }
+         }
+
+         if (storage_class == SpvStorageClassWorkgroup &&
+             b->options->lower_workgroup_access_to_offsets) {
+            uint32_t size, align;
+            val->type->deref = vtn_type_layout_std430(b, val->type->deref,
+                                                      &size, &align);
+            val->type->length = size;
+            val->type->align = align;
+         }
       }
       break;
    }
@@ -1197,10 +1454,13 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
       case SpvDimBuffer:   dim = GLSL_SAMPLER_DIM_BUF;   break;
       case SpvDimSubpassData: dim = GLSL_SAMPLER_DIM_SUBPASS; break;
       default:
-         vtn_fail("Invalid SPIR-V image dimensionality");
+         vtn_fail("Invalid SPIR-V image dimensionality: %s (%u)",
+                  spirv_dim_to_string((SpvDim)w[3]), w[3]);
       }
 
-      bool is_shadow = w[4];
+      /* w[4]: as per Vulkan spec "Validation Rules within a Module",
+       *       The “Depth” operand of OpTypeImage is ignored.
+       */
       bool is_array = w[5];
       bool multisampled = w[6];
       unsigned sampled = w[7];
@@ -1226,10 +1486,9 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
          glsl_get_base_type(sampled_type->type);
       if (sampled == 1) {
          val->type->sampled = true;
-         val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
+         val->type->type = glsl_sampler_type(dim, false, is_array,
                                              sampled_base_type);
       } else if (sampled == 2) {
-         vtn_assert(!is_shadow);
          val->type->sampled = false;
          val->type->type = glsl_image_type(dim, is_array, sampled_base_type);
       } else {
@@ -1261,10 +1520,21 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
    case SpvOpTypeQueue:
    case SpvOpTypePipe:
    default:
-      vtn_fail("Unhandled opcode");
+      vtn_fail_with_opcode("Unhandled opcode", opcode);
    }
 
    vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
+
+   if (val->type->base_type == vtn_base_type_struct &&
+       (val->type->block || val->type->buffer_block)) {
+      for (unsigned i = 0; i < val->type->length; i++) {
+         vtn_fail_if(vtn_type_contains_block(b, val->type->members[i]),
+                     "Block and BufferBlock decorations cannot decorate a "
+                     "structure type that is nested at any level inside "
+                     "another structure type decorated with Block or "
+                     "BufferBlock.");
+      }
+   }
 }
 
 static nir_constant *
@@ -1283,6 +1553,8 @@ vtn_null_constant(struct vtn_builder *b, const struct glsl_type *type)
    case GLSL_TYPE_UINT:
    case GLSL_TYPE_INT16:
    case GLSL_TYPE_UINT16:
+   case GLSL_TYPE_UINT8:
+   case GLSL_TYPE_INT8:
    case GLSL_TYPE_INT64:
    case GLSL_TYPE_UINT64:
    case GLSL_TYPE_BOOL:
@@ -1375,10 +1647,7 @@ handle_workgroup_size_decoration_cb(struct vtn_builder *b,
       return;
 
    vtn_assert(val->type->type == glsl_vector_type(GLSL_TYPE_UINT, 3));
-
-   b->shader->info.cs.local_size[0] = val->constant->values[0].u32[0];
-   b->shader->info.cs.local_size[1] = val->constant->values[0].u32[1];
-   b->shader->info.cs.local_size[2] = val->constant->values[0].u32[2];
+   b->workgroup_size_builtin = val;
 }
 
 static void
@@ -1403,7 +1672,7 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
           opcode == SpvOpSpecConstantFalse)
          int_val = get_specialization(b, val, int_val);
 
-      val->constant->values[0].u32[0] = int_val ? NIR_TRUE : NIR_FALSE;
+      val->constant->values[0][0].b = int_val != 0;
       break;
    }
 
@@ -1414,16 +1683,19 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
       int bit_size = glsl_get_bit_size(val->type->type);
       switch (bit_size) {
       case 64:
-         val->constant->values->u64[0] = vtn_u64_literal(&w[3]);
+         val->constant->values[0][0].u64 = vtn_u64_literal(&w[3]);
          break;
       case 32:
-         val->constant->values->u32[0] = w[3];
+         val->constant->values[0][0].u32 = w[3];
          break;
       case 16:
-         val->constant->values->u16[0] = w[3];
+         val->constant->values[0][0].u16 = w[3];
+         break;
+      case 8:
+         val->constant->values[0][0].u8 = w[3];
          break;
       default:
-         vtn_fail("Unsupported SpvOpConstant bit size");
+         vtn_fail("Unsupported SpvOpConstant bit size: %u", bit_size);
       }
       break;
    }
@@ -1435,14 +1707,17 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
       int bit_size = glsl_get_bit_size(val->type->type);
       switch (bit_size) {
       case 64:
-         val->constant->values[0].u64[0] =
+         val->constant->values[0][0].u64 =
             get_specialization64(b, val, vtn_u64_literal(&w[3]));
          break;
       case 32:
-         val->constant->values[0].u32[0] = get_specialization(b, val, w[3]);
+         val->constant->values[0][0].u32 = get_specialization(b, val, w[3]);
          break;
       case 16:
-         val->constant->values[0].u16[0] = get_specialization(b, val, w[3]);
+         val->constant->values[0][0].u16 = get_specialization(b, val, w[3]);
+         break;
+      case 8:
+         val->constant->values[0][0].u8 = get_specialization(b, val, w[3]);
          break;
       default:
          vtn_fail("Unsupported SpvOpSpecConstant bit size");
@@ -1458,35 +1733,36 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
                   spirv_op_to_string(opcode), elem_count, val->type->length);
 
       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;
+      for (unsigned i = 0; i < elem_count; i++) {
+         struct vtn_value *val = vtn_untyped_value(b, w[i + 3]);
+
+         if (val->value_type == vtn_value_type_constant) {
+            elems[i] = val->constant;
+         } else {
+            vtn_fail_if(val->value_type != vtn_value_type_undef,
+                        "only constants or undefs allowed for "
+                        "SpvOpConstantComposite");
+            /* to make it easier, just insert a NULL constant for now */
+            elems[i] = vtn_null_constant(b, val->type->type);
+         }
+      }
 
       switch (val->type->base_type) {
       case vtn_base_type_vector: {
          assert(glsl_type_is_vector(val->type->type));
-         int bit_size = glsl_get_bit_size(val->type->type);
-         for (unsigned i = 0; i < elem_count; i++) {
-            switch (bit_size) {
-            case 64:
-               val->constant->values[0].u64[i] = elems[i]->values[0].u64[0];
-               break;
-            case 32:
-               val->constant->values[0].u32[i] = elems[i]->values[0].u32[0];
-               break;
-            case 16:
-               val->constant->values[0].u16[i] = elems[i]->values[0].u16[0];
-               break;
-            default:
-               vtn_fail("Invalid SpvOpConstantComposite bit size");
-            }
-         }
+         for (unsigned i = 0; i < elem_count; i++)
+            val->constant->values[0][i] = elems[i]->values[0][0];
          break;
       }
 
       case vtn_base_type_matrix:
          assert(glsl_type_is_matrix(val->type->type));
-         for (unsigned i = 0; i < elem_count; i++)
-            val->constant->values[i] = elems[i]->values[0];
+         for (unsigned i = 0; i < elem_count; i++) {
+            unsigned components =
+               glsl_get_components(glsl_get_column_type(val->type->type));
+            memcpy(val->constant->values[i], elems[i]->values,
+                   sizeof(nir_const_value) * components);
+         }
          break;
 
       case vtn_base_type_struct:
@@ -1531,11 +1807,11 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
             uint64_t u64[8];
             if (v0->value_type == vtn_value_type_constant) {
                for (unsigned i = 0; i < len0; i++)
-                  u64[i] = v0->constant->values[0].u64[i];
+                  u64[i] = v0->constant->values[0][i].u64;
             }
             if (v1->value_type == vtn_value_type_constant) {
                for (unsigned i = 0; i < len1; i++)
-                  u64[len0 + i] = v1->constant->values[0].u64[i];
+                  u64[len0 + i] = v1->constant->values[0][i].u64;
             }
 
             for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
@@ -1544,20 +1820,20 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
                 * to detect if it is wrongly used.
                 */
                if (comp == (uint32_t)-1)
-                  val->constant->values[0].u64[j] = 0xdeadbeefdeadbeef;
+                  val->constant->values[0][j].u64 = 0xdeadbeefdeadbeef;
                else
-                  val->constant->values[0].u64[j] = u64[comp];
+                  val->constant->values[0][j].u64 = u64[comp];
             }
          } else {
             /* This is for both 32-bit and 16-bit values */
             uint32_t u32[8];
             if (v0->value_type == vtn_value_type_constant) {
                for (unsigned i = 0; i < len0; i++)
-                  u32[i] = v0->constant->values[0].u32[i];
+                  u32[i] = v0->constant->values[0][i].u32;
             }
             if (v1->value_type == vtn_value_type_constant) {
                for (unsigned i = 0; i < len1; i++)
-                  u32[len0 + i] = v1->constant->values[0].u32[i];
+                  u32[len0 + i] = v1->constant->values[0][i].u32;
             }
 
             for (unsigned i = 0, j = 0; i < count - 6; i++, j++) {
@@ -1566,9 +1842,9 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
                 * to detect if it is wrongly used.
                 */
                if (comp == (uint32_t)-1)
-                  val->constant->values[0].u32[j] = 0xdeadbeef;
+                  val->constant->values[0][j].u32 = 0xdeadbeef;
                else
-                  val->constant->values[0].u32[j] = u32[comp];
+                  val->constant->values[0][j].u32 = u32[comp];
             }
          }
          break;
@@ -1634,21 +1910,8 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
                val->constant = *c;
             } else {
                unsigned num_components = type->length;
-               unsigned bit_size = glsl_get_bit_size(type->type);
                for (unsigned i = 0; i < num_components; i++)
-                  switch(bit_size) {
-                  case 64:
-                     val->constant->values[0].u64[i] = (*c)->values[col].u64[elem + i];
-                     break;
-                  case 32:
-                     val->constant->values[0].u32[i] = (*c)->values[col].u32[elem + i];
-                     break;
-                  case 16:
-                     val->constant->values[0].u16[i] = (*c)->values[col].u16[elem + i];
-                     break;
-                  default:
-                     vtn_fail("Invalid SpvOpCompositeExtract bit size");
-                  }
+                  val->constant->values[0][i] = (*c)->values[col][elem + i];
             }
          } else {
             struct vtn_value *insert =
@@ -1658,21 +1921,8 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
                *c = insert->constant;
             } else {
                unsigned num_components = type->length;
-               unsigned bit_size = glsl_get_bit_size(type->type);
                for (unsigned i = 0; i < num_components; i++)
-                  switch (bit_size) {
-                  case 64:
-                     (*c)->values[col].u64[elem + i] = insert->constant->values[0].u64[i];
-                     break;
-                  case 32:
-                     (*c)->values[col].u32[elem + i] = insert->constant->values[0].u32[i];
-                     break;
-                  case 16:
-                     (*c)->values[col].u16[elem + i] = insert->constant->values[0].u16[i];
-                     break;
-                  default:
-                     vtn_fail("Invalid SpvOpCompositeInsert bit size");
-                  }
+                  (*c)->values[col][elem + i] = insert->constant->values[0][i];
             }
          }
          break;
@@ -1703,20 +1953,48 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
          };
 
          nir_op op = vtn_nir_alu_op_for_spirv_opcode(b, opcode, &swap,
-                                                     src_alu_type,
-                                                     dst_alu_type);
-         nir_const_value src[4];
+                                                     nir_alu_type_get_type_size(src_alu_type),
+                                                     nir_alu_type_get_type_size(dst_alu_type));
+         nir_const_value src[3][NIR_MAX_VEC_COMPONENTS];
 
          for (unsigned i = 0; i < count - 4; i++) {
-            nir_constant *c =
-               vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
+            struct vtn_value *src_val =
+               vtn_value(b, w[4 + i], vtn_value_type_constant);
+
+            /* If this is an unsized source, pull the bit size from the
+             * source; otherwise, we'll use the bit size from the destination.
+             */
+            if (!nir_alu_type_get_type_size(nir_op_infos[op].input_types[i]))
+               bit_size = glsl_get_bit_size(src_val->type->type);
 
             unsigned j = swap ? 1 - i : i;
-            src[j] = c->values[0];
+            memcpy(src[j], src_val->constant->values[0], sizeof(src[j]));
+         }
+
+         /* fix up fixed size sources */
+         switch (op) {
+         case nir_op_ishl:
+         case nir_op_ishr:
+         case nir_op_ushr: {
+            if (bit_size == 32)
+               break;
+            for (unsigned i = 0; i < num_components; ++i) {
+               switch (bit_size) {
+               case 64: src[1][i].u32 = src[1][i].u64; break;
+               case 16: src[1][i].u32 = src[1][i].u16; break;
+               case  8: src[1][i].u32 = src[1][i].u8;  break;
+               }
+            }
+            break;
+         }
+         default:
+            break;
          }
 
-         val->constant->values[0] =
-            nir_eval_const_opcode(op, num_components, bit_size, src);
+         nir_const_value *srcs[3] = {
+            src[0], src[1], src[2],
+         };
+         nir_eval_const_opcode(op, val->constant->values[0], num_components, bit_size, srcs);
          break;
       } /* default */
       }
@@ -1732,61 +2010,13 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
       break;
 
    default:
-      vtn_fail("Unhandled opcode");
+      vtn_fail_with_opcode("Unhandled opcode", opcode);
    }
 
    /* Now that we have the value, update the workgroup size if needed */
    vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
 }
 
-static void
-vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
-                         const uint32_t *w, unsigned count)
-{
-   struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
-   struct vtn_function *vtn_callee =
-      vtn_value(b, w[3], vtn_value_type_function)->func;
-   struct nir_function *callee = vtn_callee->impl->function;
-
-   vtn_callee->referenced = true;
-
-   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_pointer &&
-          arg->pointer->ptr_type->type == NULL) {
-         nir_deref_var *d = vtn_pointer_to_deref(b, arg->pointer);
-         call->params[i] = nir_deref_var_clone(d, call);
-      } 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->nb.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;
-   vtn_assert(res_type->type == callee->return_type);
-   if (!glsl_type_is_void(callee->return_type)) {
-      out_tmp = nir_local_variable_create(b->nb.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 {
-      vtn_push_ssa(b, w[2], res_type, vtn_local_load(b, call->return_deref));
-   }
-}
-
 struct vtn_ssa_value *
 vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
 {
@@ -1804,6 +2034,8 @@ vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
          case GLSL_TYPE_UINT:
          case GLSL_TYPE_INT16:
          case GLSL_TYPE_UINT16:
+         case GLSL_TYPE_UINT8:
+         case GLSL_TYPE_INT8:
          case GLSL_TYPE_INT64:
          case GLSL_TYPE_UINT64:
          case GLSL_TYPE_BOOL:
@@ -1816,6 +2048,7 @@ vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
             child_type = glsl_get_array_element(type);
             break;
          case GLSL_TYPE_STRUCT:
+         case GLSL_TYPE_INTERFACE:
             child_type = glsl_get_struct_field(type, i);
             break;
          default:
@@ -1882,7 +2115,6 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
    const struct glsl_type *image_type = sampled.type->type;
    const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
    const bool is_array = glsl_sampler_type_is_array(image_type);
-   const bool is_shadow = glsl_sampler_type_is_shadow(image_type);
 
    /* Figure out the base texture operation */
    nir_texop texop;
@@ -1932,12 +2164,44 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
       break;
 
    default:
-      vtn_fail("Unhandled opcode");
+      vtn_fail_with_opcode("Unhandled opcode", opcode);
    }
 
-   nir_tex_src srcs[8]; /* 8 should be enough */
+   nir_tex_src srcs[10]; /* 10 should be enough */
    nir_tex_src *p = srcs;
 
+   nir_deref_instr *sampler = vtn_pointer_to_deref(b, sampled.sampler);
+   nir_deref_instr *texture =
+      sampled.image ? vtn_pointer_to_deref(b, sampled.image) : sampler;
+
+   p->src = nir_src_for_ssa(&texture->dest.ssa);
+   p->src_type = nir_tex_src_texture_deref;
+   p++;
+
+   switch (texop) {
+   case nir_texop_tex:
+   case nir_texop_txb:
+   case nir_texop_txl:
+   case nir_texop_txd:
+   case nir_texop_tg4:
+   case nir_texop_lod:
+      /* These operations require a sampler */
+      p->src = nir_src_for_ssa(&sampler->dest.ssa);
+      p->src_type = nir_tex_src_sampler_deref;
+      p++;
+      break;
+   case nir_texop_txf:
+   case nir_texop_txf_ms:
+   case nir_texop_txs:
+   case nir_texop_query_levels:
+   case nir_texop_texture_samples:
+   case nir_texop_samples_identical:
+      /* These don't */
+      break;
+   case nir_texop_txf_ms_mcs:
+      vtn_fail("unexpected nir_texop_txf_ms_mcs");
+   }
+
    unsigned idx = 4;
 
    struct nir_ssa_def *coord;
@@ -2006,6 +2270,7 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
       break;
    }
 
+   bool is_shadow = false;
    unsigned gather_component = 0;
    switch (opcode) {
    case SpvOpImageSampleDrefImplicitLod:
@@ -2014,13 +2279,14 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
    case SpvOpImageSampleProjDrefExplicitLod:
    case SpvOpImageDrefGather:
       /* These all have an explicit depth value as their next source */
+      is_shadow = true;
       (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparator);
       break;
 
    case SpvOpImageGather:
       /* This has a component as its next source */
       gather_component =
-         vtn_value(b, w[idx++], vtn_value_type_constant)->constant->values[0].u32[0];
+         vtn_value(b, w[idx++], vtn_value_type_constant)->constant->values[0][0].u32;
       break;
 
    default:
@@ -2032,7 +2298,7 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
       (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
 
    /* Now we need to handle some number of optional arguments */
-   const struct vtn_ssa_value *gather_offsets = NULL;
+   struct vtn_value *gather_offsets = NULL;
    if (idx < count) {
       uint32_t operands = w[idx++];
 
@@ -2060,8 +2326,8 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
          (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
 
       if (operands & SpvImageOperandsConstOffsetsMask) {
-         gather_offsets = vtn_ssa_value(b, w[idx++]);
-         (*p++) = (nir_tex_src){};
+         vtn_assert(texop == nir_texop_tg4);
+         gather_offsets = vtn_value(b, w[idx++], vtn_value_type_constant);
       }
 
       if (operands & SpvImageOperandsSampleMask) {
@@ -2069,6 +2335,13 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
          texop = nir_texop_txf_ms;
          (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
       }
+
+      if (operands & SpvImageOperandsMinLodMask) {
+         vtn_assert(texop == nir_texop_tex ||
+                    texop == nir_texop_txb ||
+                    texop == nir_texop_txd);
+         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_min_lod);
+      }
    }
    /* We should have now consumed exactly all of the arguments */
    vtn_assert(idx == count);
@@ -2086,6 +2359,12 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
       is_shadow && glsl_get_components(ret_type->type) == 1;
    instr->component = gather_component;
 
+   if (sampled.image && (sampled.image->access & ACCESS_NON_UNIFORM))
+      instr->texture_non_uniform = true;
+
+   if (sampled.sampler && (sampled.sampler->access & ACCESS_NON_UNIFORM))
+      instr->sampler_non_uniform = true;
+
    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;
@@ -2095,105 +2374,46 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
       vtn_fail("Invalid base type for sampler result");
    }
 
-   nir_deref_var *sampler = vtn_pointer_to_deref(b, sampled.sampler);
-   nir_deref_var *texture;
-   if (sampled.image) {
-      nir_deref_var *image = vtn_pointer_to_deref(b, sampled.image);
-      texture = image;
-   } else {
-      texture = sampler;
-   }
-
-   instr->texture = nir_deref_var_clone(texture, instr);
-
-   switch (instr->op) {
-   case nir_texop_tex:
-   case nir_texop_txb:
-   case nir_texop_txl:
-   case nir_texop_txd:
-   case nir_texop_tg4:
-      /* These operations require a sampler */
-      instr->sampler = nir_deref_var_clone(sampler, instr);
-      break;
-   case nir_texop_txf:
-   case nir_texop_txf_ms:
-   case nir_texop_txs:
-   case nir_texop_lod:
-   case nir_texop_query_levels:
-   case nir_texop_texture_samples:
-   case nir_texop_samples_identical:
-      /* These don't */
-      instr->sampler = NULL;
-      break;
-   case nir_texop_txf_ms_mcs:
-      vtn_fail("unexpected nir_texop_txf_ms_mcs");
-   }
-
    nir_ssa_dest_init(&instr->instr, &instr->dest,
                      nir_tex_instr_dest_size(instr), 32, NULL);
 
    vtn_assert(glsl_get_vector_elements(ret_type->type) ==
               nir_tex_instr_dest_size(instr));
 
-   nir_ssa_def *def;
-   nir_instr *instruction;
    if (gather_offsets) {
-      vtn_assert(glsl_get_base_type(gather_offsets->type) == GLSL_TYPE_ARRAY);
-      vtn_assert(glsl_get_length(gather_offsets->type) == 4);
-      nir_tex_instr *instrs[4] = {instr, NULL, NULL, NULL};
-
-      /* Copy the current instruction 4x */
-      for (uint32_t i = 1; i < 4; i++) {
-         instrs[i] = nir_tex_instr_create(b->shader, instr->num_srcs);
-         instrs[i]->op = instr->op;
-         instrs[i]->coord_components = instr->coord_components;
-         instrs[i]->sampler_dim = instr->sampler_dim;
-         instrs[i]->is_array = instr->is_array;
-         instrs[i]->is_shadow = instr->is_shadow;
-         instrs[i]->is_new_style_shadow = instr->is_new_style_shadow;
-         instrs[i]->component = instr->component;
-         instrs[i]->dest_type = instr->dest_type;
-         instrs[i]->texture = nir_deref_var_clone(texture, instrs[i]);
-         instrs[i]->sampler = NULL;
-
-         memcpy(instrs[i]->src, srcs, instr->num_srcs * sizeof(*instr->src));
-
-         nir_ssa_dest_init(&instrs[i]->instr, &instrs[i]->dest,
-                           nir_tex_instr_dest_size(instr), 32, NULL);
-      }
-
-      /* Fill in the last argument with the offset from the passed in offsets
-       * and insert the instruction into the stream.
-       */
+      vtn_fail_if(gather_offsets->type->base_type != vtn_base_type_array ||
+                  gather_offsets->type->length != 4,
+                  "ConstOffsets must be an array of size four of vectors "
+                  "of two integer components");
+
+      struct vtn_type *vec_type = gather_offsets->type->array_element;
+      vtn_fail_if(vec_type->base_type != vtn_base_type_vector ||
+                  vec_type->length != 2 ||
+                  !glsl_type_is_integer(vec_type->type),
+                  "ConstOffsets must be an array of size four of vectors "
+                  "of two integer components");
+
+      unsigned bit_size = glsl_get_bit_size(vec_type->type);
       for (uint32_t i = 0; i < 4; i++) {
-         nir_tex_src src;
-         src.src = nir_src_for_ssa(gather_offsets->elems[i]->def);
-         src.src_type = nir_tex_src_offset;
-         instrs[i]->src[instrs[i]->num_srcs - 1] = src;
-         nir_builder_instr_insert(&b->nb, &instrs[i]->instr);
-      }
-
-      /* Combine the results of the 4 instructions by taking their .w
-       * components
-       */
-      nir_alu_instr *vec4 = nir_alu_instr_create(b->shader, nir_op_vec4);
-      nir_ssa_dest_init(&vec4->instr, &vec4->dest.dest, 4, 32, NULL);
-      vec4->dest.write_mask = 0xf;
-      for (uint32_t i = 0; i < 4; i++) {
-         vec4->src[i].src = nir_src_for_ssa(&instrs[i]->dest.ssa);
-         vec4->src[i].swizzle[0] = 3;
+         const nir_const_value *cvec =
+            gather_offsets->constant->elements[i]->values[0];
+         for (uint32_t j = 0; j < 2; j++) {
+            switch (bit_size) {
+            case 8:  instr->tg4_offsets[i][j] = cvec[j].i8;    break;
+            case 16: instr->tg4_offsets[i][j] = cvec[j].i16;   break;
+            case 32: instr->tg4_offsets[i][j] = cvec[j].i32;   break;
+            case 64: instr->tg4_offsets[i][j] = cvec[j].i64;   break;
+            default:
+               vtn_fail("Unsupported bit size: %u", bit_size);
+            }
+         }
       }
-      def = &vec4->dest.dest.ssa;
-      instruction = &vec4->instr;
-   } else {
-      def = &instr->dest.ssa;
-      instruction = &instr->instr;
    }
 
    val->ssa = vtn_create_ssa_value(b, ret_type->type);
-   val->ssa->def = def;
+   val->ssa->def = &instr->dest.ssa;
 
-   nir_builder_instr_insert(&b->nb, instruction);
+   nir_builder_instr_insert(&b->nb, &instr->instr);
 }
 
 static void
@@ -2232,7 +2452,7 @@ fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
       break;
 
    default:
-      vtn_fail("Invalid SPIR-V atomic");
+      vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode);
    }
 }
 
@@ -2250,6 +2470,18 @@ get_image_coord(struct vtn_builder *b, uint32_t value)
    return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
 }
 
+static nir_ssa_def *
+expand_to_vec4(nir_builder *b, nir_ssa_def *value)
+{
+   if (value->num_components == 4)
+      return value;
+
+   unsigned swiz[4];
+   for (unsigned i = 0; i < 4; i++)
+      swiz[i] = i < value->num_components ? i : 0;
+   return nir_swizzle(b, value, swiz, 4, false);
+}
+
 static void
 vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
                  const uint32_t *w, unsigned count)
@@ -2324,12 +2556,12 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
       break;
 
    default:
-      vtn_fail("Invalid image opcode");
+      vtn_fail_with_opcode("Invalid image opcode", opcode);
    }
 
    nir_intrinsic_op op;
    switch (opcode) {
-#define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_##N; break;
+#define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
    OP(ImageQuerySize,         size)
    OP(ImageRead,              load)
    OP(ImageWrite,             store)
@@ -2350,25 +2582,21 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
    OP(AtomicXor,              atomic_xor)
 #undef OP
    default:
-      vtn_fail("Invalid image opcode");
+      vtn_fail_with_opcode("Invalid image opcode", opcode);
    }
 
    nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
 
-   nir_deref_var *image_deref = vtn_pointer_to_deref(b, image.image);
-   intrin->variables[0] = nir_deref_var_clone(image_deref, intrin);
+   nir_deref_instr *image_deref = vtn_pointer_to_deref(b, image.image);
+   intrin->src[0] = nir_src_for_ssa(&image_deref->dest.ssa);
 
    /* 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);
+      intrin->src[1] = nir_src_for_ssa(expand_to_vec4(&b->nb, image.coord));
+      intrin->src[2] = nir_src_for_ssa(image.sample);
    }
 
    switch (opcode) {
@@ -2377,11 +2605,15 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
    case SpvOpImageRead:
       break;
    case SpvOpAtomicStore:
-      intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
-      break;
-   case SpvOpImageWrite:
-      intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
+   case SpvOpImageWrite: {
+      const uint32_t value_id = opcode == SpvOpAtomicStore ? w[4] : w[3];
+      nir_ssa_def *value = vtn_ssa_value(b, value_id)->def;
+      /* nir_intrinsic_image_deref_store always takes a vec4 value */
+      assert(op == nir_intrinsic_image_deref_store);
+      intrin->num_components = 4;
+      intrin->src[3] = nir_src_for_ssa(expand_to_vec4(&b->nb, value));
       break;
+   }
 
    case SpvOpAtomicCompareExchange:
    case SpvOpAtomicIIncrement:
@@ -2396,31 +2628,33 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
    case SpvOpAtomicAnd:
    case SpvOpAtomicOr:
    case SpvOpAtomicXor:
-      fill_common_atomic_sources(b, opcode, w, &intrin->src[2]);
+      fill_common_atomic_sources(b, opcode, w, &intrin->src[3]);
       break;
 
    default:
-      vtn_fail("Invalid image opcode");
+      vtn_fail_with_opcode("Invalid image opcode", opcode);
    }
 
-   if (opcode != SpvOpImageWrite) {
+   if (opcode != SpvOpImageWrite && opcode != SpvOpAtomicStore) {
       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;
 
-      unsigned dest_components =
-         nir_intrinsic_infos[intrin->intrinsic].dest_components;
-      if (intrin->intrinsic == nir_intrinsic_image_size) {
-         dest_components = intrin->num_components =
-            glsl_get_vector_elements(type->type);
-      }
+      unsigned dest_components = glsl_get_vector_elements(type->type);
+      intrin->num_components = nir_intrinsic_infos[op].dest_components;
+      if (intrin->num_components == 0)
+         intrin->num_components = dest_components;
 
       nir_ssa_dest_init(&intrin->instr, &intrin->dest,
-                        dest_components, 32, NULL);
+                        intrin->num_components, 32, NULL);
 
       nir_builder_instr_insert(&b->nb, &intrin->instr);
 
+      nir_ssa_def *result = &intrin->dest.ssa;
+      if (intrin->num_components != dest_components)
+         result = nir_channels(&b->nb, result, (1 << dest_components) - 1);
+
       val->ssa = vtn_create_ssa_value(b, type->type);
-      val->ssa->def = &intrin->dest.ssa;
+      val->ssa->def = result;
    } else {
       nir_builder_instr_insert(&b->nb, &intrin->instr);
    }
@@ -2448,7 +2682,36 @@ get_ssbo_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
    OP(AtomicXor,              atomic_xor)
 #undef OP
    default:
-      vtn_fail("Invalid SSBO atomic");
+      vtn_fail_with_opcode("Invalid SSBO atomic", opcode);
+   }
+}
+
+static nir_intrinsic_op
+get_uniform_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
+{
+   switch (opcode) {
+#define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
+   OP(AtomicLoad,             read_deref)
+   OP(AtomicExchange,         exchange)
+   OP(AtomicCompareExchange,  comp_swap)
+   OP(AtomicIIncrement,       inc_deref)
+   OP(AtomicIDecrement,       post_dec_deref)
+   OP(AtomicIAdd,             add_deref)
+   OP(AtomicISub,             add_deref)
+   OP(AtomicUMin,             min_deref)
+   OP(AtomicUMax,             max_deref)
+   OP(AtomicAnd,              and_deref)
+   OP(AtomicOr,               or_deref)
+   OP(AtomicXor,              xor_deref)
+#undef OP
+   default:
+      /* We left the following out: AtomicStore, AtomicSMin and
+       * AtomicSmax. Right now there are not nir intrinsics for them. At this
+       * moment Atomic Counter support is needed for ARB_spirv support, so is
+       * only need to support GLSL Atomic Counters that are uints and don't
+       * allow direct storage.
+       */
+      unreachable("Invalid uniform atomic");
    }
 }
 
@@ -2474,17 +2737,17 @@ get_shared_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
    OP(AtomicXor,              atomic_xor)
 #undef OP
    default:
-      vtn_fail("Invalid shared atomic");
+      vtn_fail_with_opcode("Invalid shared atomic", opcode);
    }
 }
 
 static nir_intrinsic_op
-get_var_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
+get_deref_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
 {
    switch (opcode) {
-   case SpvOpAtomicLoad:      return nir_intrinsic_load_var;
-   case SpvOpAtomicStore:     return nir_intrinsic_store_var;
-#define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
+   case SpvOpAtomicLoad:      return nir_intrinsic_load_deref;
+   case SpvOpAtomicStore:     return nir_intrinsic_store_deref;
+#define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
    OP(AtomicExchange,         atomic_exchange)
    OP(AtomicCompareExchange,  atomic_comp_swap)
    OP(AtomicIIncrement,       atomic_add)
@@ -2500,13 +2763,16 @@ get_var_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
    OP(AtomicXor,              atomic_xor)
 #undef OP
    default:
-      vtn_fail("Invalid shared atomic");
+      vtn_fail_with_opcode("Invalid shared atomic", opcode);
    }
 }
 
+/*
+ * Handles shared atomics, ssbo atomics and atomic counters.
+ */
 static void
-vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
-                                 const uint32_t *w, unsigned count)
+vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode,
+                   const uint32_t *w, unsigned count)
 {
    struct vtn_pointer *ptr;
    nir_intrinsic_instr *atomic;
@@ -2535,7 +2801,7 @@ vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
       break;
 
    default:
-      vtn_fail("Invalid SPIR-V atomic");
+      vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode);
    }
 
    /*
@@ -2543,13 +2809,18 @@ vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
    SpvMemorySemanticsMask semantics = w[5];
    */
 
-   if (ptr->mode == vtn_variable_mode_workgroup &&
-       !b->options->lower_workgroup_access_to_offsets) {
-      nir_deref_var *deref = vtn_pointer_to_deref(b, ptr);
-      const struct glsl_type *deref_type = nir_deref_tail(&deref->deref)->type;
-      nir_intrinsic_op op = get_var_nir_atomic_op(b, opcode);
+   /* uniform as "atomic counter uniform" */
+   if (ptr->mode == vtn_variable_mode_uniform) {
+      nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
+      const struct glsl_type *deref_type = deref->type;
+      nir_intrinsic_op op = get_uniform_nir_atomic_op(b, opcode);
       atomic = nir_intrinsic_instr_create(b->nb.shader, op);
-      atomic->variables[0] = nir_deref_var_clone(deref, atomic);
+      atomic->src[0] = nir_src_for_ssa(&deref->dest.ssa);
+
+      /* SSBO needs to initialize index/offset. In this case we don't need to,
+       * as that info is already stored on the ptr->var->var nir_variable (see
+       * vtn_create_variable)
+       */
 
       switch (opcode) {
       case SpvOpAtomicLoad:
@@ -2559,7 +2830,6 @@ vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
       case SpvOpAtomicStore:
          atomic->num_components = glsl_get_vector_elements(deref_type);
          nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
-         atomic->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
          break;
 
       case SpvOpAtomicExchange:
@@ -2576,16 +2846,18 @@ vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
       case SpvOpAtomicAnd:
       case SpvOpAtomicOr:
       case SpvOpAtomicXor:
-         fill_common_atomic_sources(b, opcode, w, &atomic->src[0]);
+         /* Nothing: we don't need to call fill_common_atomic_sources here, as
+          * atomic counter uniforms doesn't have sources
+          */
          break;
 
       default:
-         vtn_fail("Invalid SPIR-V atomic");
+         unreachable("Invalid SPIR-V atomic");
 
       }
-   } else {
+   } else if (vtn_pointer_uses_ssa_offset(b, ptr)) {
       nir_ssa_def *offset, *index;
-      offset = vtn_pointer_to_offset(b, ptr, &index, NULL);
+      offset = vtn_pointer_to_offset(b, ptr, &index);
 
       nir_intrinsic_op op;
       if (ptr->mode == vtn_variable_mode_ssbo) {
@@ -2602,6 +2874,7 @@ vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
       switch (opcode) {
       case SpvOpAtomicLoad:
          atomic->num_components = glsl_get_vector_elements(ptr->type->type);
+         nir_intrinsic_set_align(atomic, 4, 0);
          if (ptr->mode == vtn_variable_mode_ssbo)
             atomic->src[src++] = nir_src_for_ssa(index);
          atomic->src[src++] = nir_src_for_ssa(offset);
@@ -2610,6 +2883,7 @@ vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
       case SpvOpAtomicStore:
          atomic->num_components = glsl_get_vector_elements(ptr->type->type);
          nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
+         nir_intrinsic_set_align(atomic, 4, 0);
          atomic->src[src++] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
          if (ptr->mode == vtn_variable_mode_ssbo)
             atomic->src[src++] = nir_src_for_ssa(index);
@@ -2637,7 +2911,45 @@ vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
          break;
 
       default:
-         vtn_fail("Invalid SPIR-V atomic");
+         vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode);
+      }
+   } else {
+      nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
+      const struct glsl_type *deref_type = deref->type;
+      nir_intrinsic_op op = get_deref_nir_atomic_op(b, opcode);
+      atomic = nir_intrinsic_instr_create(b->nb.shader, op);
+      atomic->src[0] = nir_src_for_ssa(&deref->dest.ssa);
+
+      switch (opcode) {
+      case SpvOpAtomicLoad:
+         atomic->num_components = glsl_get_vector_elements(deref_type);
+         break;
+
+      case SpvOpAtomicStore:
+         atomic->num_components = glsl_get_vector_elements(deref_type);
+         nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
+         atomic->src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
+         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:
+         fill_common_atomic_sources(b, opcode, w, &atomic->src[1]);
+         break;
+
+      default:
+         vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode);
       }
    }
 
@@ -2662,11 +2974,11 @@ create_vec(struct vtn_builder *b, unsigned num_components, unsigned bit_size)
 {
    nir_op op;
    switch (num_components) {
-   case 1: op = nir_op_fmov; break;
+   case 1: op = nir_op_imov; break;
    case 2: op = nir_op_vec2; break;
    case 3: op = nir_op_vec3; break;
    case 4: op = nir_op_vec4; break;
-   default: vtn_fail("bad vector size");
+   default: vtn_fail("bad vector size: %u", num_components);
    }
 
    nir_alu_instr *vec = nir_alu_instr_create(b->shader, op);
@@ -2710,8 +3022,7 @@ vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
 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);
+   return nir_channel(&b->nb, src, index);
 }
 
 nir_ssa_def *
@@ -2735,16 +3046,17 @@ vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
    return &vec->dest.dest.ssa;
 }
 
+static nir_ssa_def *
+nir_ieq_imm(nir_builder *b, nir_ssa_def *x, uint64_t i)
+{
+   return nir_ieq(b, x, nir_imm_intN_t(b, i, x->bit_size));
+}
+
 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;
+   return nir_vector_extract(&b->nb, src, nir_i2i(&b->nb, index, 32));
 }
 
 nir_ssa_def *
@@ -2753,7 +3065,7 @@ vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
 {
    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)),
+      dest = nir_bcsel(&b->nb, nir_ieq_imm(&b->nb, index, i),
                        vtn_vector_insert(b, src, insert, i), dest);
 
    return dest;
@@ -2927,7 +3239,7 @@ vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
       unsigned elems = count - 3;
       assume(elems >= 1);
       if (glsl_type_is_vector_or_scalar(type)) {
-         nir_ssa_def *srcs[4];
+         nir_ssa_def *srcs[NIR_MAX_VEC_COMPONENTS];
          for (unsigned i = 0; i < elems; i++)
             srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
          val->ssa->def =
@@ -2956,7 +3268,7 @@ vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
       break;
 
    default:
-      vtn_fail("unknown composite operation");
+      vtn_fail_with_opcode("unknown composite operation", opcode);
    }
 }
 
@@ -3051,9 +3363,12 @@ vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
 
       switch (opcode) {
       case SpvOpEmitStreamVertex:
-      case SpvOpEndStreamPrimitive:
-         nir_intrinsic_set_stream_id(intrin, w[1]);
+      case SpvOpEndStreamPrimitive: {
+         unsigned stream = vtn_constant_uint(b, w[1]);
+         nir_intrinsic_set_stream_id(intrin, stream);
          break;
+      }
+
       default:
          break;
       }
@@ -3063,23 +3378,19 @@ vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode,
    }
 
    case SpvOpMemoryBarrier: {
-      SpvScope scope = vtn_constant_value(b, w[1])->values[0].u32[0];
-      SpvMemorySemanticsMask semantics =
-         vtn_constant_value(b, w[2])->values[0].u32[0];
+      SpvScope scope = vtn_constant_uint(b, w[1]);
+      SpvMemorySemanticsMask semantics = vtn_constant_uint(b, w[2]);
       vtn_emit_memory_barrier(b, scope, semantics);
       return;
    }
 
    case SpvOpControlBarrier: {
-      SpvScope execution_scope =
-         vtn_constant_value(b, w[1])->values[0].u32[0];
+      SpvScope execution_scope = vtn_constant_uint(b, w[1]);
       if (execution_scope == SpvScopeWorkgroup)
          vtn_emit_barrier(b, nir_intrinsic_barrier);
 
-      SpvScope memory_scope =
-         vtn_constant_value(b, w[2])->values[0].u32[0];
-      SpvMemorySemanticsMask memory_semantics =
-         vtn_constant_value(b, w[3])->values[0].u32[0];
+      SpvScope memory_scope = vtn_constant_uint(b, w[2]);
+      SpvMemorySemanticsMask memory_semantics = vtn_constant_uint(b, w[3]);
       vtn_emit_memory_barrier(b, memory_scope, memory_semantics);
       break;
    }
@@ -3114,7 +3425,8 @@ gl_primitive_from_spv_execution_mode(struct vtn_builder *b,
    case SpvExecutionModeOutputTriangleStrip:
       return 5; /* GL_TRIANGLE_STRIP */
    default:
-      vtn_fail("Invalid primitive type");
+      vtn_fail("Invalid primitive type: %s (%u)",
+               spirv_executionmode_to_string(mode), mode);
    }
 }
 
@@ -3134,7 +3446,8 @@ vertices_in_from_spv_execution_mode(struct vtn_builder *b,
    case SpvExecutionModeInputTrianglesAdjacency:
       return 6;
    default:
-      vtn_fail("Invalid GS input mode");
+      vtn_fail("Invalid GS input mode: %s (%u)",
+               spirv_executionmode_to_string(mode), mode);
    }
 }
 
@@ -3154,17 +3467,38 @@ stage_for_execution_model(struct vtn_builder *b, SpvExecutionModel model)
       return MESA_SHADER_FRAGMENT;
    case SpvExecutionModelGLCompute:
       return MESA_SHADER_COMPUTE;
+   case SpvExecutionModelKernel:
+      return MESA_SHADER_KERNEL;
    default:
-      vtn_fail("Unsupported execution model");
+      vtn_fail("Unsupported execution model: %s (%u)",
+               spirv_executionmodel_to_string(model), model);
    }
 }
 
-#define spv_check_supported(name, cap) do {            \
-      if (!(b->options && b->options->caps.name))      \
-         vtn_warn("Unsupported SPIR-V capability: %s",  \
-                  spirv_capability_to_string(cap));     \
+#define spv_check_supported(name, cap) do {                 \
+      if (!(b->options && b->options->caps.name))           \
+         vtn_warn("Unsupported SPIR-V capability: %s (%u)", \
+                  spirv_capability_to_string(cap), cap);    \
    } while(0)
 
+
+void
+vtn_handle_entry_point(struct vtn_builder *b, const uint32_t *w,
+                       unsigned count)
+{
+   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(b, w[1]) != b->entry_point_stage)
+      return;
+
+   vtn_assert(b->entry_point == NULL);
+   b->entry_point = entry_point;
+}
+
 static bool
 vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
                                 const uint32_t *w, unsigned count)
@@ -3229,32 +3563,60 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
       case SpvCapabilityStorageImageExtendedFormats:
          break;
 
-      case SpvCapabilityGeometryStreams:
       case SpvCapabilityLinkage:
       case SpvCapabilityVector16:
       case SpvCapabilityFloat16Buffer:
-      case SpvCapabilityFloat16:
-      case SpvCapabilityInt64Atomics:
-      case SpvCapabilityAtomicStorage:
-      case SpvCapabilityInt16:
-      case SpvCapabilityStorageImageMultisample:
-      case SpvCapabilityInt8:
       case SpvCapabilitySparseResidency:
-      case SpvCapabilityMinLod:
-      case SpvCapabilityTransformFeedback:
          vtn_warn("Unsupported SPIR-V capability: %s",
                   spirv_capability_to_string(cap));
          break;
 
+      case SpvCapabilityMinLod:
+         spv_check_supported(min_lod, cap);
+         break;
+
+      case SpvCapabilityAtomicStorage:
+         spv_check_supported(atomic_storage, cap);
+         break;
+
       case SpvCapabilityFloat64:
          spv_check_supported(float64, cap);
          break;
       case SpvCapabilityInt64:
          spv_check_supported(int64, cap);
          break;
+      case SpvCapabilityInt16:
+         spv_check_supported(int16, cap);
+         break;
+
+      case SpvCapabilityTransformFeedback:
+         spv_check_supported(transform_feedback, cap);
+         break;
+
+      case SpvCapabilityGeometryStreams:
+         spv_check_supported(geometry_streams, cap);
+         break;
+
+      case SpvCapabilityInt64Atomics:
+         spv_check_supported(int64_atomics, cap);
+         break;
+
+      case SpvCapabilityInt8:
+         spv_check_supported(int8, cap);
+         break;
+
+      case SpvCapabilityStorageImageMultisample:
+         spv_check_supported(storage_image_ms, cap);
+         break;
 
       case SpvCapabilityAddresses:
+         spv_check_supported(address, cap);
+         break;
+
       case SpvCapabilityKernel:
+         spv_check_supported(kernel, cap);
+         break;
+
       case SpvCapabilityImageBasic:
       case SpvCapabilityImageReadWrite:
       case SpvCapabilityImageMipmap:
@@ -3316,14 +3678,17 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
 
       case SpvCapabilityGroupNonUniformQuad:
          spv_check_supported(subgroup_quad, cap);
+         break;
 
       case SpvCapabilityGroupNonUniformArithmetic:
       case SpvCapabilityGroupNonUniformClustered:
          spv_check_supported(subgroup_arithmetic, cap);
+         break;
 
       case SpvCapabilityVariablePointersStorageBuffer:
       case SpvCapabilityVariablePointers:
          spv_check_supported(variable_pointers, cap);
+         b->variable_pointers = true;
          break;
 
       case SpvCapabilityStorageUniformBufferBlock16:
@@ -3337,8 +3702,46 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
          spv_check_supported(shader_viewport_index_layer, cap);
          break;
 
+      case SpvCapabilityStorageBuffer8BitAccess:
+      case SpvCapabilityUniformAndStorageBuffer8BitAccess:
+      case SpvCapabilityStoragePushConstant8:
+         spv_check_supported(storage_8bit, cap);
+         break;
+
+      case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT:
+      case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT:
+      case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT:
+         spv_check_supported(descriptor_array_dynamic_indexing, cap);
+         break;
+
+      case SpvCapabilityRuntimeDescriptorArrayEXT:
+         spv_check_supported(runtime_descriptor_array, cap);
+         break;
+
+      case SpvCapabilityStencilExportEXT:
+         spv_check_supported(stencil_export, cap);
+         break;
+
+      case SpvCapabilitySampleMaskPostDepthCoverage:
+         spv_check_supported(post_depth_coverage, cap);
+         break;
+
+      case SpvCapabilityPhysicalStorageBufferAddressesEXT:
+         spv_check_supported(physical_storage_buffer_address, cap);
+         break;
+
+      case SpvCapabilityComputeDerivativeGroupQuadsNV:
+      case SpvCapabilityComputeDerivativeGroupLinearNV:
+         spv_check_supported(derivative_group, cap);
+         break;
+
+      case SpvCapabilityFloat16:
+         spv_check_supported(float16, cap);
+         break;
+
       default:
-         vtn_fail("Unhandled capability");
+         vtn_fail("Unhandled capability: %s (%u)",
+                  spirv_capability_to_string(cap), cap);
       }
       break;
    }
@@ -3348,25 +3751,50 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
       break;
 
    case SpvOpMemoryModel:
-      vtn_assert(w[1] == SpvAddressingModelLogical);
+      switch (w[1]) {
+      case SpvAddressingModelPhysical32:
+         vtn_fail_if(b->shader->info.stage != MESA_SHADER_KERNEL,
+                     "AddressingModelPhysical32 only supported for kernels");
+         b->shader->info.cs.ptr_size = 32;
+         b->physical_ptrs = true;
+         b->options->shared_ptr_type = glsl_uint_type();
+         b->options->global_ptr_type = glsl_uint_type();
+         b->options->temp_ptr_type = glsl_uint_type();
+         break;
+      case SpvAddressingModelPhysical64:
+         vtn_fail_if(b->shader->info.stage != MESA_SHADER_KERNEL,
+                     "AddressingModelPhysical64 only supported for kernels");
+         b->shader->info.cs.ptr_size = 64;
+         b->physical_ptrs = true;
+         b->options->shared_ptr_type = glsl_uint64_t_type();
+         b->options->global_ptr_type = glsl_uint64_t_type();
+         b->options->temp_ptr_type = glsl_uint64_t_type();
+         break;
+      case SpvAddressingModelLogical:
+         vtn_fail_if(b->shader->info.stage >= MESA_SHADER_STAGES,
+                     "AddressingModelLogical only supported for shaders");
+         b->shader->info.cs.ptr_size = 0;
+         b->physical_ptrs = false;
+         break;
+      case SpvAddressingModelPhysicalStorageBuffer64EXT:
+         vtn_fail_if(!b->options ||
+                     !b->options->caps.physical_storage_buffer_address,
+                     "AddressingModelPhysicalStorageBuffer64EXT not supported");
+         break;
+      default:
+         vtn_fail("Unknown addressing model: %s (%u)",
+                  spirv_addressingmodel_to_string(w[1]), w[1]);
+         break;
+      }
+
       vtn_assert(w[2] == SpvMemoryModelSimple ||
-                 w[2] == SpvMemoryModelGLSL450);
+                 w[2] == SpvMemoryModelGLSL450 ||
+                 w[2] == SpvMemoryModelOpenCL);
       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(b, w[1]) != b->entry_point_stage)
-         break;
-
-      vtn_assert(b->entry_point == NULL);
-      b->entry_point = entry_point;
+   case SpvOpEntryPoint:
+      vtn_handle_entry_point(b, w, count);
       break;
-   }
 
    case SpvOpString:
       vtn_push_value(b, w[1], vtn_value_type_string)->str =
@@ -3382,11 +3810,14 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
       break;
 
    case SpvOpExecutionMode:
+   case SpvOpExecutionModeId:
    case SpvOpDecorationGroup:
    case SpvOpDecorate:
    case SpvOpMemberDecorate:
    case SpvOpGroupDecorate:
    case SpvOpGroupMemberDecorate:
+   case SpvOpDecorateStringGOOGLE:
+   case SpvOpMemberDecorateStringGOOGLE:
       vtn_handle_decoration(b, opcode, w, count);
       break;
 
@@ -3406,7 +3837,8 @@ vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
    switch(mode->exec_mode) {
    case SpvExecutionModeOriginUpperLeft:
    case SpvExecutionModeOriginLowerLeft:
-      b->origin_upper_left =
+      vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
+      b->shader->info.fs.origin_upper_left =
          (mode->exec_mode == SpvExecutionModeOriginUpperLeft);
       break;
 
@@ -3415,6 +3847,11 @@ vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
       b->shader->info.fs.early_fragment_tests = true;
       break;
 
+   case SpvExecutionModePostDepthCoverage:
+      vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
+      b->shader->info.fs.post_depth_coverage = true;
+      break;
+
    case SpvExecutionModeInvocations:
       vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
       b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
@@ -3438,12 +3875,20 @@ vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
       break;
 
    case SpvExecutionModeLocalSize:
-      vtn_assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
+      vtn_assert(gl_shader_stage_is_compute(b->shader->info.stage));
       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 SpvExecutionModeLocalSizeId:
+      b->shader->info.cs.local_size[0] = vtn_constant_uint(b, mode->literals[0]);
+      b->shader->info.cs.local_size[1] = vtn_constant_uint(b, mode->literals[1]);
+      b->shader->info.cs.local_size[2] = vtn_constant_uint(b, mode->literals[2]);
+      break;
+
    case SpvExecutionModeLocalSizeHint:
+   case SpvExecutionModeLocalSizeHintId:
       break; /* Nothing to do with this */
 
    case SpvExecutionModeOutputVertices:
@@ -3471,6 +3916,8 @@ vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
          vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
          b->shader->info.gs.vertices_in =
             vertices_in_from_spv_execution_mode(b, mode->exec_mode);
+         b->shader->info.gs.input_primitive =
+            gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
       }
       break;
 
@@ -3514,19 +3961,43 @@ vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point,
       break;
 
    case SpvExecutionModePixelCenterInteger:
-      b->pixel_center_integer = true;
+      vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
+      b->shader->info.fs.pixel_center_integer = true;
       break;
 
    case SpvExecutionModeXfb:
-      vtn_fail("Unhandled execution mode");
+      b->shader->info.has_transform_feedback_varyings = true;
       break;
 
    case SpvExecutionModeVecTypeHint:
-   case SpvExecutionModeContractionOff:
       break; /* OpenCL */
 
+   case SpvExecutionModeContractionOff:
+      if (b->shader->info.stage != MESA_SHADER_KERNEL)
+         vtn_warn("ExectionMode only allowed for CL-style kernels: %s",
+                  spirv_executionmode_to_string(mode->exec_mode));
+      else
+         b->exact = true;
+      break;
+
+   case SpvExecutionModeStencilRefReplacingEXT:
+      vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
+      break;
+
+   case SpvExecutionModeDerivativeGroupQuadsNV:
+      vtn_assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
+      b->shader->info.cs.derivative_group = DERIVATIVE_GROUP_QUADS;
+      break;
+
+   case SpvExecutionModeDerivativeGroupLinearNV:
+      vtn_assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
+      b->shader->info.cs.derivative_group = DERIVATIVE_GROUP_LINEAR;
+      break;
+
    default:
-      vtn_fail("Unhandled execution mode");
+      vtn_fail("Unhandled execution mode: %s (%u)",
+               spirv_executionmode_to_string(mode->exec_mode),
+               mode->exec_mode);
    }
 }
 
@@ -3554,6 +4025,8 @@ vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
    case SpvOpMemberDecorate:
    case SpvOpGroupDecorate:
    case SpvOpGroupMemberDecorate:
+   case SpvOpDecorateStringGOOGLE:
+   case SpvOpMemberDecorateStringGOOGLE:
       vtn_fail("Invalid opcode types and variables section");
       break;
 
@@ -3571,6 +4044,7 @@ vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
    case SpvOpTypeStruct:
    case SpvOpTypeOpaque:
    case SpvOpTypePointer:
+   case SpvOpTypeForwardPointer:
    case SpvOpTypeFunction:
    case SpvOpTypeEvent:
    case SpvOpTypeDeviceEvent:
@@ -3637,7 +4111,10 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
    case SpvOpAccessChain:
    case SpvOpPtrAccessChain:
    case SpvOpInBoundsAccessChain:
+   case SpvOpInBoundsPtrAccessChain:
    case SpvOpArrayLength:
+   case SpvOpConvertPtrToU:
+   case SpvOpConvertUToPtr:
       vtn_handle_variables(b, opcode, w, count);
       break;
 
@@ -3674,10 +4151,10 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
    case SpvOpImageQuerySize: {
       struct vtn_pointer *image =
          vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
-      if (image->mode == vtn_variable_mode_image) {
+      if (glsl_type_is_image(image->type->type)) {
          vtn_handle_image(b, opcode, w, count);
       } else {
-         vtn_assert(image->mode == vtn_variable_mode_sampler);
+         vtn_assert(glsl_type_is_sampler(image->type->type));
          vtn_handle_texture(b, opcode, w, count);
       }
       break;
@@ -3703,7 +4180,7 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
          vtn_handle_image(b, opcode, w, count);
       } else {
          vtn_assert(pointer->value_type == vtn_value_type_pointer);
-         vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
+         vtn_handle_atomics(b, opcode, w, count);
       }
       break;
    }
@@ -3714,7 +4191,7 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
          vtn_handle_image(b, opcode, w, count);
       } else {
          vtn_assert(pointer->value_type == vtn_value_type_pointer);
-         vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
+         vtn_handle_atomics(b, opcode, w, count);
       }
       break;
    }
@@ -3794,8 +4271,6 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
    case SpvOpSConvert:
    case SpvOpFConvert:
    case SpvOpQuantizeToF16:
-   case SpvOpConvertPtrToU:
-   case SpvOpConvertUToPtr:
    case SpvOpPtrCastToGeneric:
    case SpvOpGenericCastToPtr:
    case SpvOpBitcast:
@@ -3940,21 +4415,23 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
       break;
 
    default:
-      vtn_fail("Unhandled opcode");
+      vtn_fail_with_opcode("Unhandled opcode", 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 struct spirv_to_nir_options *options,
-             const nir_shader_compiler_options *nir_options)
+struct vtn_builder*
+vtn_create_builder(const uint32_t *words, size_t word_count,
+                   gl_shader_stage stage, const char *entry_point_name,
+                   const struct spirv_to_nir_options *options)
 {
-   /* Initialize the stn_builder object */
+   /* Initialize the vtn_builder object */
    struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
+   struct spirv_to_nir_options *dup_options =
+      ralloc(b, struct spirv_to_nir_options);
+   *dup_options = *options;
+
    b->spirv = words;
    b->spirv_word_count = word_count;
    b->file = NULL;
@@ -3963,30 +4440,139 @@ spirv_to_nir(const uint32_t *words, size_t word_count,
    exec_list_make_empty(&b->functions);
    b->entry_point_stage = stage;
    b->entry_point_name = entry_point_name;
-   b->options = options;
+   b->options = dup_options;
 
-   /* See also _vtn_fail() */
-   if (setjmp(b->fail_jump)) {
-      ralloc_free(b);
-      return NULL;
+   /*
+    * Handle the SPIR-V header (first 5 dwords).
+    * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
+    */
+   if (word_count <= 5)
+      goto fail;
+
+   if (words[0] != SpvMagicNumber) {
+      vtn_err("words[0] was 0x%x, want 0x%x", words[0], SpvMagicNumber);
+      goto fail;
+   }
+   if (words[1] < 0x10000) {
+      vtn_err("words[1] was 0x%x, want >= 0x10000", words[1]);
+      goto fail;
    }
 
-   const uint32_t *word_end = words + word_count;
+   uint16_t generator_id = words[2] >> 16;
+   uint16_t generator_version = words[2];
 
-   /* Handle the SPIR-V header (first 4 dwords)  */
-   vtn_assert(word_count > 5);
+   /* The first GLSLang version bump actually 1.5 years after #179 was fixed
+    * but this should at least let us shut the workaround off for modern
+    * versions of GLSLang.
+    */
+   b->wa_glslang_179 = (generator_id == 8 && generator_version == 1);
 
-   vtn_assert(words[0] == SpvMagicNumber);
-   vtn_assert(words[1] >= 0x10000);
    /* words[2] == generator magic */
    unsigned value_id_bound = words[3];
-   vtn_assert(words[4] == 0);
-
-   words+= 5;
+   if (words[4] != 0) {
+      vtn_err("words[4] was %u, want 0", words[4]);
+      goto fail;
+   }
 
    b->value_id_bound = value_id_bound;
    b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
 
+   return b;
+ fail:
+   ralloc_free(b);
+   return NULL;
+}
+
+static nir_function *
+vtn_emit_kernel_entry_point_wrapper(struct vtn_builder *b,
+                                    nir_function *entry_point)
+{
+   vtn_assert(entry_point == b->entry_point->func->impl->function);
+   vtn_fail_if(!entry_point->name, "entry points are required to have a name");
+   const char *func_name =
+      ralloc_asprintf(b->shader, "__wrapped_%s", entry_point->name);
+
+   /* we shouldn't have any inputs yet */
+   vtn_assert(!entry_point->shader->num_inputs);
+   vtn_assert(b->shader->info.stage == MESA_SHADER_KERNEL);
+
+   nir_function *main_entry_point = nir_function_create(b->shader, func_name);
+   main_entry_point->impl = nir_function_impl_create(main_entry_point);
+   nir_builder_init(&b->nb, main_entry_point->impl);
+   b->nb.cursor = nir_after_cf_list(&main_entry_point->impl->body);
+   b->func_param_idx = 0;
+
+   nir_call_instr *call = nir_call_instr_create(b->nb.shader, entry_point);
+
+   for (unsigned i = 0; i < entry_point->num_params; ++i) {
+      struct vtn_type *param_type = b->entry_point->func->type->params[i];
+
+      /* consider all pointers to function memory to be parameters passed
+       * by value
+       */
+      bool is_by_val = param_type->base_type == vtn_base_type_pointer &&
+         param_type->storage_class == SpvStorageClassFunction;
+
+      /* input variable */
+      nir_variable *in_var = rzalloc(b->nb.shader, nir_variable);
+      in_var->data.mode = nir_var_shader_in;
+      in_var->data.read_only = true;
+      in_var->data.location = i;
+
+      if (is_by_val)
+         in_var->type = param_type->deref->type;
+      else
+         in_var->type = param_type->type;
+
+      nir_shader_add_variable(b->nb.shader, in_var);
+      b->nb.shader->num_inputs++;
+
+      /* we have to copy the entire variable into function memory */
+      if (is_by_val) {
+         nir_variable *copy_var =
+            nir_local_variable_create(main_entry_point->impl, in_var->type,
+                                      "copy_in");
+         nir_copy_var(&b->nb, copy_var, in_var);
+         call->params[i] =
+            nir_src_for_ssa(&nir_build_deref_var(&b->nb, copy_var)->dest.ssa);
+      } else {
+         call->params[i] = nir_src_for_ssa(nir_load_var(&b->nb, in_var));
+      }
+   }
+
+   nir_builder_instr_insert(&b->nb, &call->instr);
+
+   return main_entry_point;
+}
+
+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 struct spirv_to_nir_options *options,
+             const nir_shader_compiler_options *nir_options)
+
+{
+   const uint32_t *word_end = words + word_count;
+
+   struct vtn_builder *b = vtn_create_builder(words, word_count,
+                                              stage, entry_point_name,
+                                              options);
+
+   if (b == NULL)
+      return NULL;
+
+   /* See also _vtn_fail() */
+   if (setjmp(b->fail_jump)) {
+      ralloc_free(b);
+      return NULL;
+   }
+
+   /* Skip the SPIR-V header, handled at vtn_create_builder */
+   words+= 5;
+
+   b->shader = nir_shader_create(b, stage, nir_options, NULL);
+
    /* Handle all the preamble instructions */
    words = vtn_foreach_instruction(b, words, word_end,
                                    vtn_handle_preamble_instruction);
@@ -3997,15 +4583,9 @@ spirv_to_nir(const uint32_t *words, size_t word_count,
       return NULL;
    }
 
-   b->shader = nir_shader_create(b, stage, nir_options, NULL);
-
    /* 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;
 
@@ -4013,6 +4593,22 @@ spirv_to_nir(const uint32_t *words, size_t word_count,
    words = vtn_foreach_instruction(b, words, word_end,
                                    vtn_handle_variable_or_type_instruction);
 
+   /* Parse execution modes */
+   vtn_foreach_execution_mode(b, b->entry_point,
+                              vtn_handle_execution_mode, NULL);
+
+   if (b->workgroup_size_builtin) {
+      vtn_assert(b->workgroup_size_builtin->type->type ==
+                 glsl_vector_type(GLSL_TYPE_UINT, 3));
+
+      nir_const_value *const_size =
+         b->workgroup_size_builtin->constant->values[0];
+
+      b->shader->info.cs.local_size[0] = const_size[0].u32;
+      b->shader->info.cs.local_size[1] = const_size[1].u32;
+      b->shader->info.cs.local_size[2] = const_size[2].u32;
+   }
+
    /* Set types on all vtn_values */
    vtn_foreach_instruction(b, words, word_end, vtn_set_instruction_result_type);
 
@@ -4026,8 +4622,7 @@ spirv_to_nir(const uint32_t *words, size_t word_count,
       progress = false;
       foreach_list_typed(struct vtn_function, func, node, &b->functions) {
          if (func->referenced && !func->emitted) {
-            b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
-                                                     _mesa_key_pointer_equal);
+            b->const_table = _mesa_pointer_hash_table_create(b);
 
             vtn_function_emit(b, func, vtn_handle_body_instruction);
             progress = true;
@@ -4039,6 +4634,31 @@ spirv_to_nir(const uint32_t *words, size_t word_count,
    nir_function *entry_point = b->entry_point->func->impl->function;
    vtn_assert(entry_point);
 
+   /* post process entry_points with input params */
+   if (entry_point->num_params && b->shader->info.stage == MESA_SHADER_KERNEL)
+      entry_point = vtn_emit_kernel_entry_point_wrapper(b, entry_point);
+
+   entry_point->is_entrypoint = true;
+
+   /* When multiple shader stages exist in the same SPIR-V module, we
+    * generate input and output variables for every stage, in the same
+    * NIR program.  These dead variables can be invalid NIR.  For example,
+    * TCS outputs must be per-vertex arrays (or decorated 'patch'), while
+    * VS output variables wouldn't be.
+    *
+    * To ensure we have valid NIR, we eliminate any dead inputs and outputs
+    * right away.  In order to do so, we must lower any constant initializers
+    * on outputs so nir_remove_dead_variables sees that they're written to.
+    */
+   nir_lower_constant_initializers(b->shader, nir_var_shader_out);
+   nir_remove_dead_variables(b->shader,
+                             nir_var_shader_in | nir_var_shader_out);
+
+   /* We sometimes generate bogus derefs that, while never used, give the
+    * validator a bit of heartburn.  Run dead code to get rid of them.
+    */
+   nir_opt_dce(b->shader);
+
    /* Unparent the shader from the vtn_builder before we delete the builder */
    ralloc_steal(NULL, b->shader);