* their own.
*/
-#include "ir.h"
+#include "lower_buffer_access.h"
#include "ir_builder.h"
-#include "ir_rvalue_visitor.h"
#include "main/macros.h"
#include "glsl_parser_extras.h"
using namespace ir_builder;
-/**
- * Determine if a thing being dereferenced is row-major
- *
- * There is some trickery here.
- *
- * If the thing being dereferenced is a member of uniform block \b without an
- * instance name, then the name of the \c ir_variable is the field name of an
- * interface type. If this field is row-major, then the thing referenced is
- * row-major.
- *
- * If the thing being dereferenced is a member of uniform block \b with an
- * instance name, then the last dereference in the tree will be an
- * \c ir_dereference_record. If that record field is row-major, then the
- * thing referenced is row-major.
- */
-static bool
-is_dereferenced_thing_row_major(const ir_rvalue *deref)
-{
- bool matrix = false;
- const ir_rvalue *ir = deref;
-
- while (true) {
- matrix = matrix || ir->type->without_array()->is_matrix();
-
- switch (ir->ir_type) {
- case ir_type_dereference_array: {
- const ir_dereference_array *const array_deref =
- (const ir_dereference_array *) ir;
-
- ir = array_deref->array;
- break;
- }
-
- case ir_type_dereference_record: {
- const ir_dereference_record *const record_deref =
- (const ir_dereference_record *) ir;
-
- ir = record_deref->record;
-
- const int idx = ir->type->field_index(record_deref->field);
- assert(idx >= 0);
-
- const enum glsl_matrix_layout matrix_layout =
- glsl_matrix_layout(ir->type->fields.structure[idx].matrix_layout);
-
- switch (matrix_layout) {
- case GLSL_MATRIX_LAYOUT_INHERITED:
- break;
- case GLSL_MATRIX_LAYOUT_COLUMN_MAJOR:
- return false;
- case GLSL_MATRIX_LAYOUT_ROW_MAJOR:
- return matrix || deref->type->without_array()->is_record();
- }
-
- break;
- }
-
- case ir_type_dereference_variable: {
- const ir_dereference_variable *const var_deref =
- (const ir_dereference_variable *) ir;
-
- const enum glsl_matrix_layout matrix_layout =
- glsl_matrix_layout(var_deref->var->data.matrix_layout);
-
- switch (matrix_layout) {
- case GLSL_MATRIX_LAYOUT_INHERITED:
- assert(!matrix);
- return false;
- case GLSL_MATRIX_LAYOUT_COLUMN_MAJOR:
- return false;
- case GLSL_MATRIX_LAYOUT_ROW_MAJOR:
- return matrix || deref->type->without_array()->is_record();
- }
-
- unreachable("invalid matrix layout");
- break;
- }
-
- default:
- return false;
- }
- }
-
- /* The tree must have ended with a dereference that wasn't an
- * ir_dereference_variable. That is invalid, and it should be impossible.
- */
- unreachable("invalid dereference tree");
- return false;
-}
-
namespace {
-class lower_ubo_reference_visitor : public ir_rvalue_enter_visitor {
+class lower_ubo_reference_visitor :
+ public lower_buffer_access::lower_buffer_access {
public:
lower_ubo_reference_visitor(struct gl_shader *shader)
: shader(shader)
void handle_rvalue(ir_rvalue **rvalue);
ir_visitor_status visit_enter(ir_assignment *ir);
- void setup_for_load_or_store(ir_variable *var,
+ void setup_for_load_or_store(void *mem_ctx,
+ ir_variable *var,
ir_rvalue *deref,
ir_rvalue **offset,
unsigned *const_offset,
bool *row_major,
int *matrix_columns,
unsigned packing);
- ir_expression *ubo_load(const struct glsl_type *type,
+ ir_expression *ubo_load(void *mem_ctx, const struct glsl_type *type,
ir_rvalue *offset);
- ir_call *ssbo_load(const struct glsl_type *type,
+ ir_call *ssbo_load(void *mem_ctx, const struct glsl_type *type,
ir_rvalue *offset);
+ bool check_for_buffer_array_copy(ir_assignment *ir);
+ bool check_for_buffer_struct_copy(ir_assignment *ir);
void check_for_ssbo_store(ir_assignment *ir);
- void write_to_memory(ir_dereference *deref,
- ir_variable *var,
- ir_variable *write_var,
- unsigned write_mask);
- ir_call *ssbo_store(ir_rvalue *deref, ir_rvalue *offset,
+ void write_to_memory(void *mem_ctx, ir_dereference *deref, ir_variable *var,
+ ir_variable *write_var, unsigned write_mask);
+ ir_call *ssbo_store(void *mem_ctx, ir_rvalue *deref, ir_rvalue *offset,
unsigned write_mask);
- void emit_access(bool is_write, ir_dereference *deref,
- ir_variable *base_offset, unsigned int deref_offset,
- bool row_major, int matrix_columns,
- unsigned packing, unsigned write_mask);
+ enum {
+ ubo_load_access,
+ ssbo_load_access,
+ ssbo_store_access,
+ ssbo_unsized_array_length_access,
+ ssbo_atomic_access,
+ } buffer_access_type;
+
+ void insert_buffer_access(void *mem_ctx, ir_dereference *deref,
+ const glsl_type *type, ir_rvalue *offset,
+ unsigned mask, int channel);
ir_visitor_status visit_enter(class ir_expression *);
ir_expression *calculate_ssbo_unsized_array_length(ir_expression *expr);
ir_expression *process_ssbo_unsized_array_length(ir_rvalue **,
ir_dereference *,
ir_variable *);
- ir_expression *emit_ssbo_get_buffer_size();
+ ir_expression *emit_ssbo_get_buffer_size(void *mem_ctx);
unsigned calculate_unsized_array_stride(ir_dereference *deref,
unsigned packing);
ir_call *check_for_ssbo_atomic_intrinsic(ir_call *ir);
ir_visitor_status visit_enter(ir_call *ir);
- void *mem_ctx;
struct gl_shader *shader;
struct gl_uniform_buffer_variable *ubo_var;
ir_rvalue *uniform_block;
bool progress;
- bool is_shader_storage;
};
/**
interface_field_name(void *mem_ctx, char *base_name, ir_rvalue *d,
ir_rvalue **nonconst_block_index)
{
- ir_rvalue *previous_index = NULL;
*nonconst_block_index = NULL;
+ char *name_copy = NULL;
+ size_t base_length = 0;
+
+ /* Loop back through the IR until we find the uniform block */
+ ir_rvalue *ir = d;
+ while (ir != NULL) {
+ switch (ir->ir_type) {
+ case ir_type_dereference_variable: {
+ /* Exit loop */
+ ir = NULL;
+ break;
+ }
+
+ case ir_type_dereference_record: {
+ ir_dereference_record *r = (ir_dereference_record *) ir;
+ ir = r->record->as_dereference();
+
+ /* If we got here it means any previous array subscripts belong to
+ * block members and not the block itself so skip over them in the
+ * next pass.
+ */
+ d = ir;
+ break;
+ }
+
+ case ir_type_dereference_array: {
+ ir_dereference_array *a = (ir_dereference_array *) ir;
+ ir = a->array->as_dereference();
+ break;
+ }
+
+ case ir_type_swizzle: {
+ ir_swizzle *s = (ir_swizzle *) ir;
+ ir = s->val->as_dereference();
+ /* Skip swizzle in the next pass */
+ d = ir;
+ break;
+ }
+
+ default:
+ assert(!"Should not get here.");
+ break;
+ }
+ }
while (d != NULL) {
switch (d->ir_type) {
case ir_type_dereference_variable: {
ir_dereference_variable *v = (ir_dereference_variable *) d;
- if (previous_index
- && v->var->is_interface_instance()
- && v->var->type->is_array()) {
-
- ir_constant *const_index = previous_index->as_constant();
- if (!const_index) {
- *nonconst_block_index = previous_index;
- return ralloc_asprintf(mem_ctx, "%s[0]", base_name);
- } else {
- return ralloc_asprintf(mem_ctx,
- "%s[%d]",
- base_name,
- const_index->get_uint_component(0));
- }
+ if (name_copy != NULL &&
+ v->var->is_interface_instance() &&
+ v->var->type->is_array()) {
+ return name_copy;
} else {
+ *nonconst_block_index = NULL;
return base_name;
}
break;
}
- case ir_type_dereference_record: {
- ir_dereference_record *r = (ir_dereference_record *) d;
-
- d = r->record->as_dereference();
- break;
- }
-
case ir_type_dereference_array: {
ir_dereference_array *a = (ir_dereference_array *) d;
+ size_t new_length;
+
+ if (name_copy == NULL) {
+ name_copy = ralloc_strdup(mem_ctx, base_name);
+ base_length = strlen(name_copy);
+ }
+
+ /* For arrays of arrays we start at the innermost array and work our
+ * way out so we need to insert the subscript at the base of the
+ * name string rather than just attaching it to the end.
+ */
+ new_length = base_length;
+ ir_constant *const_index = a->array_index->as_constant();
+ char *end = ralloc_strdup(NULL, &name_copy[new_length]);
+ if (!const_index) {
+ ir_rvalue *array_index = a->array_index;
+ if (array_index->type != glsl_type::uint_type)
+ array_index = i2u(array_index);
+
+ if (a->array->type->is_array() &&
+ a->array->type->fields.array->is_array()) {
+ ir_constant *base_size = new(mem_ctx)
+ ir_constant(a->array->type->fields.array->arrays_of_arrays_size());
+ array_index = mul(array_index, base_size);
+ }
+
+ if (*nonconst_block_index) {
+ *nonconst_block_index = add(*nonconst_block_index, array_index);
+ } else {
+ *nonconst_block_index = array_index;
+ }
+
+ ralloc_asprintf_rewrite_tail(&name_copy, &new_length, "[0]%s",
+ end);
+ } else {
+ ralloc_asprintf_rewrite_tail(&name_copy, &new_length, "[%d]%s",
+ const_index->get_uint_component(0),
+ end);
+ }
+ ralloc_free(end);
d = a->array->as_dereference();
- previous_index = a->array_index;
break;
}
- case ir_type_swizzle: {
- ir_swizzle *s = (ir_swizzle *) d;
- d = s->val->as_dereference();
- break;
- }
default:
assert(!"Should not get here.");
break;
}
void
-lower_ubo_reference_visitor::setup_for_load_or_store(ir_variable *var,
+lower_ubo_reference_visitor::setup_for_load_or_store(void *mem_ctx,
+ ir_variable *var,
ir_rvalue *deref,
ir_rvalue **offset,
unsigned *const_offset,
deref, &nonconst_block_index);
/* Locate the block by interface name */
- this->is_shader_storage = var->is_in_shader_storage_block();
unsigned num_blocks;
struct gl_uniform_block **blocks;
- if (this->is_shader_storage) {
+ if (this->buffer_access_type != ubo_load_access) {
num_blocks = shader->NumShaderStorageBlocks;
blocks = shader->ShaderStorageBlocks;
} else {
ir_constant *index = new(mem_ctx) ir_constant(i);
if (nonconst_block_index) {
- if (nonconst_block_index->type != glsl_type::uint_type)
- nonconst_block_index = i2u(nonconst_block_index);
this->uniform_block = add(nonconst_block_index, index);
} else {
this->uniform_block = index;
assert(this->uniform_block);
- *offset = new(mem_ctx) ir_constant(0u);
- *const_offset = 0;
- *row_major = is_dereferenced_thing_row_major(deref);
- *matrix_columns = 1;
-
- /* Calculate the offset to the start of the region of the UBO
- * dereferenced by *rvalue. This may be a variable offset if an
- * array dereference has a variable index.
- */
- while (deref) {
- switch (deref->ir_type) {
- case ir_type_dereference_variable: {
- *const_offset += ubo_var->Offset;
- deref = NULL;
- break;
- }
-
- case ir_type_dereference_array: {
- ir_dereference_array *deref_array = (ir_dereference_array *) deref;
- unsigned array_stride;
- if (deref_array->array->type->is_matrix() && *row_major) {
- /* When loading a vector out of a row major matrix, the
- * step between the columns (vectors) is the size of a
- * float, while the step between the rows (elements of a
- * vector) is handled below in emit_ubo_loads.
- */
- array_stride = 4;
- if (deref_array->array->type->is_double())
- array_stride *= 2;
- *matrix_columns = deref_array->array->type->matrix_columns;
- } else if (deref_array->type->is_interface()) {
- /* We're processing an array dereference of an interface instance
- * array. The thing being dereferenced *must* be a variable
- * dereference because interfaces cannot be embedded in other
- * types. In terms of calculating the offsets for the lowering
- * pass, we don't care about the array index. All elements of an
- * interface instance array will have the same offsets relative to
- * the base of the block that backs them.
- */
- assert(deref_array->array->as_dereference_variable());
- deref = deref_array->array->as_dereference();
- break;
- } else {
- /* Whether or not the field is row-major (because it might be a
- * bvec2 or something) does not affect the array itself. We need
- * to know whether an array element in its entirety is row-major.
- */
- const bool array_row_major =
- is_dereferenced_thing_row_major(deref_array);
-
- /* The array type will give the correct interface packing
- * information
- */
- if (packing == GLSL_INTERFACE_PACKING_STD430) {
- array_stride = deref_array->type->std430_array_stride(array_row_major);
- } else {
- array_stride = deref_array->type->std140_size(array_row_major);
- array_stride = glsl_align(array_stride, 16);
- }
- }
-
- ir_rvalue *array_index = deref_array->array_index;
- if (array_index->type->base_type == GLSL_TYPE_INT)
- array_index = i2u(array_index);
-
- ir_constant *const_index =
- array_index->constant_expression_value(NULL);
- if (const_index) {
- *const_offset += array_stride * const_index->value.u[0];
- } else {
- *offset = add(*offset,
- mul(array_index,
- new(mem_ctx) ir_constant(array_stride)));
- }
- deref = deref_array->array->as_dereference();
- break;
- }
-
- case ir_type_dereference_record: {
- ir_dereference_record *deref_record = (ir_dereference_record *) deref;
- const glsl_type *struct_type = deref_record->record->type;
- unsigned intra_struct_offset = 0;
-
- for (unsigned int i = 0; i < struct_type->length; i++) {
- const glsl_type *type = struct_type->fields.structure[i].type;
-
- ir_dereference_record *field_deref = new(mem_ctx)
- ir_dereference_record(deref_record->record,
- struct_type->fields.structure[i].name);
- const bool field_row_major =
- is_dereferenced_thing_row_major(field_deref);
-
- ralloc_free(field_deref);
-
- unsigned field_align = 0;
-
- if (packing == GLSL_INTERFACE_PACKING_STD430)
- field_align = type->std430_base_alignment(field_row_major);
- else
- field_align = type->std140_base_alignment(field_row_major);
-
- intra_struct_offset = glsl_align(intra_struct_offset, field_align);
-
- if (strcmp(struct_type->fields.structure[i].name,
- deref_record->field) == 0)
- break;
-
- if (packing == GLSL_INTERFACE_PACKING_STD430)
- intra_struct_offset += type->std430_size(field_row_major);
- else
- intra_struct_offset += type->std140_size(field_row_major);
-
- /* If the field just examined was itself a structure, apply rule
- * #9:
- *
- * "The structure may have padding at the end; the base offset
- * of the member following the sub-structure is rounded up to
- * the next multiple of the base alignment of the structure."
- */
- if (type->without_array()->is_record()) {
- intra_struct_offset = glsl_align(intra_struct_offset,
- field_align);
-
- }
- }
-
- *const_offset += intra_struct_offset;
- deref = deref_record->record->as_dereference();
- break;
- }
-
- case ir_type_swizzle: {
- ir_swizzle *deref_swizzle = (ir_swizzle *) deref;
-
- assert(deref_swizzle->mask.num_components == 1);
+ *const_offset = ubo_var->Offset;
- *const_offset += deref_swizzle->mask.x * sizeof(int);
- deref = deref_swizzle->val->as_dereference();
- break;
- }
-
- default:
- assert(!"not reached");
- deref = NULL;
- break;
- }
- }
+ setup_buffer_access(mem_ctx, var, deref, offset, const_offset, row_major,
+ matrix_columns, packing);
}
void
if (!var || !var->is_in_buffer_block())
return;
- mem_ctx = ralloc_parent(shader->ir);
+ void *mem_ctx = ralloc_parent(shader->ir);
ir_rvalue *offset = NULL;
unsigned const_offset;
int matrix_columns;
unsigned packing = var->get_interface_type()->interface_packing;
+ this->buffer_access_type =
+ var->is_in_shader_storage_block() ?
+ ssbo_load_access : ubo_load_access;
+
/* Compute the offset to the start if the dereference as well as other
* information we need to configure the write
*/
- setup_for_load_or_store(var, deref,
+ setup_for_load_or_store(mem_ctx, var, deref,
&offset, &const_offset,
&row_major, &matrix_columns,
packing);
base_ir->insert_before(assign(load_offset, offset));
deref = new(mem_ctx) ir_dereference_variable(load_var);
- emit_access(false, deref, load_offset, const_offset,
+ emit_access(mem_ctx, false, deref, load_offset, const_offset,
row_major, matrix_columns, packing, 0);
*rvalue = deref;
}
ir_expression *
-lower_ubo_reference_visitor::ubo_load(const glsl_type *type,
+lower_ubo_reference_visitor::ubo_load(void *mem_ctx,
+ const glsl_type *type,
ir_rvalue *offset)
{
ir_rvalue *block_ref = this->uniform_block->clone(mem_ctx, NULL);
}
ir_call *
-lower_ubo_reference_visitor::ssbo_store(ir_rvalue *deref,
+lower_ubo_reference_visitor::ssbo_store(void *mem_ctx,
+ ir_rvalue *deref,
ir_rvalue *offset,
unsigned write_mask)
{
}
ir_call *
-lower_ubo_reference_visitor::ssbo_load(const struct glsl_type *type,
+lower_ubo_reference_visitor::ssbo_load(void *mem_ctx,
+ const struct glsl_type *type,
ir_rvalue *offset)
{
exec_list sig_params;
return new(mem_ctx) ir_call(sig, deref_result, &call_params);
}
-static inline int
-writemask_for_size(unsigned n)
-{
- return ((1 << n) - 1);
-}
-
-/**
- * Takes a deref and recursively calls itself to break the deref down to the
- * point that the reads or writes generated are contiguous scalars or vectors.
- */
void
-lower_ubo_reference_visitor::emit_access(bool is_write,
- ir_dereference *deref,
- ir_variable *base_offset,
- unsigned int deref_offset,
- bool row_major,
- int matrix_columns,
- unsigned packing,
- unsigned write_mask)
+lower_ubo_reference_visitor::insert_buffer_access(void *mem_ctx,
+ ir_dereference *deref,
+ const glsl_type *type,
+ ir_rvalue *offset,
+ unsigned mask,
+ int channel)
{
- if (deref->type->is_record()) {
- unsigned int field_offset = 0;
-
- for (unsigned i = 0; i < deref->type->length; i++) {
- const struct glsl_struct_field *field =
- &deref->type->fields.structure[i];
- ir_dereference *field_deref =
- new(mem_ctx) ir_dereference_record(deref->clone(mem_ctx, NULL),
- field->name);
-
- field_offset =
- glsl_align(field_offset,
- field->type->std140_base_alignment(row_major));
-
- emit_access(is_write, field_deref, base_offset,
- deref_offset + field_offset,
- row_major, 1, packing,
- writemask_for_size(field_deref->type->vector_elements));
-
- field_offset += field->type->std140_size(row_major);
- }
- return;
- }
-
- if (deref->type->is_array()) {
- unsigned array_stride = packing == GLSL_INTERFACE_PACKING_STD430 ?
- deref->type->fields.array->std430_array_stride(row_major) :
- glsl_align(deref->type->fields.array->std140_size(row_major), 16);
-
- for (unsigned i = 0; i < deref->type->length; i++) {
- ir_constant *element = new(mem_ctx) ir_constant(i);
- ir_dereference *element_deref =
- new(mem_ctx) ir_dereference_array(deref->clone(mem_ctx, NULL),
- element);
- emit_access(is_write, element_deref, base_offset,
- deref_offset + i * array_stride,
- row_major, 1, packing,
- writemask_for_size(element_deref->type->vector_elements));
- }
- return;
- }
-
- if (deref->type->is_matrix()) {
- for (unsigned i = 0; i < deref->type->matrix_columns; i++) {
- ir_constant *col = new(mem_ctx) ir_constant(i);
- ir_dereference *col_deref =
- new(mem_ctx) ir_dereference_array(deref->clone(mem_ctx, NULL), col);
-
- if (row_major) {
- /* For a row-major matrix, the next column starts at the next
- * element.
- */
- int size_mul = deref->type->is_double() ? 8 : 4;
- emit_access(is_write, col_deref, base_offset,
- deref_offset + i * size_mul,
- row_major, deref->type->matrix_columns, packing,
- writemask_for_size(col_deref->type->vector_elements));
- } else {
- int size_mul;
-
- /* std430 doesn't round up vec2 size to a vec4 size */
- if (packing == GLSL_INTERFACE_PACKING_STD430 &&
- deref->type->vector_elements == 2 &&
- !deref->type->is_double()) {
- size_mul = 8;
- } else {
- /* std140 always rounds the stride of arrays (and matrices) to a
- * vec4, so matrices are always 16 between columns/rows. With
- * doubles, they will be 32 apart when there are more than 2 rows.
- *
- * For both std140 and std430, if the member is a
- * three-'component vector with components consuming N basic
- * machine units, the base alignment is 4N. For vec4, base
- * alignment is 4N.
- */
- size_mul = (deref->type->is_double() &&
- deref->type->vector_elements > 2) ? 32 : 16;
- }
-
- emit_access(is_write, col_deref, base_offset,
- deref_offset + i * size_mul,
- row_major, deref->type->matrix_columns, packing,
- writemask_for_size(col_deref->type->vector_elements));
- }
- }
- return;
+ switch (this->buffer_access_type) {
+ case ubo_load_access:
+ base_ir->insert_before(assign(deref->clone(mem_ctx, NULL),
+ ubo_load(mem_ctx, type, offset),
+ mask));
+ break;
+ case ssbo_load_access: {
+ ir_call *load_ssbo = ssbo_load(mem_ctx, type, offset);
+ base_ir->insert_before(load_ssbo);
+ ir_rvalue *value = load_ssbo->return_deref->as_rvalue()->clone(mem_ctx, NULL);
+ ir_assignment *assignment =
+ assign(deref->clone(mem_ctx, NULL), value, mask);
+ base_ir->insert_before(assignment);
+ break;
}
-
- assert(deref->type->is_scalar() || deref->type->is_vector());
-
- if (!row_major) {
- ir_rvalue *offset =
- add(base_offset, new(mem_ctx) ir_constant(deref_offset));
- if (is_write)
- base_ir->insert_after(ssbo_store(deref, offset, write_mask));
- else {
- if (!this->is_shader_storage) {
- base_ir->insert_before(assign(deref->clone(mem_ctx, NULL),
- ubo_load(deref->type, offset)));
- } else {
- ir_call *load_ssbo = ssbo_load(deref->type, offset);
- base_ir->insert_before(load_ssbo);
- ir_rvalue *value = load_ssbo->return_deref->as_rvalue()->clone(mem_ctx, NULL);
- base_ir->insert_before(assign(deref->clone(mem_ctx, NULL), value));
- }
- }
- } else {
- unsigned N = deref->type->is_double() ? 8 : 4;
-
- /* We're dereffing a column out of a row-major matrix, so we
- * gather the vector from each stored row.
- */
- assert(deref->type->base_type == GLSL_TYPE_FLOAT ||
- deref->type->base_type == GLSL_TYPE_DOUBLE);
- /* Matrices, row_major or not, are stored as if they were
- * arrays of vectors of the appropriate size in std140.
- * Arrays have their strides rounded up to a vec4, so the
- * matrix stride is always 16. However a double matrix may either be 16
- * or 32 depending on the number of columns.
- */
- assert(matrix_columns <= 4);
- unsigned matrix_stride = 0;
- /* Matrix stride for std430 mat2xY matrices are not rounded up to
- * vec4 size. From OpenGL 4.3 spec, section 7.6.2.2 "Standard Uniform
- * Block Layout":
- *
- * "2. If the member is a two- or four-component vector with components
- * consuming N basic machine units, the base alignment is 2N or 4N,
- * respectively." [...]
- * "4. If the member is an array of scalars or vectors, the base alignment
- * and array stride are set to match the base alignment of a single array
- * element, according to rules (1), (2), and (3), and rounded up to the
- * base alignment of a vec4." [...]
- * "7. If the member is a row-major matrix with C columns and R rows, the
- * matrix is stored identically to an array of R row vectors with C
- * components each, according to rule (4)." [...]
- * "When using the std430 storage layout, shader storage blocks will be
- * laid out in buffer storage identically to uniform and shader storage
- * blocks using the std140 layout, except that the base alignment and
- * stride of arrays of scalars and vectors in rule 4 and of structures in
- * rule 9 are not rounded up a multiple of the base alignment of a vec4."
- */
- if (packing == GLSL_INTERFACE_PACKING_STD430 && matrix_columns == 2)
- matrix_stride = 2 * N;
- else
- matrix_stride = glsl_align(matrix_columns * N, 16);
-
- const glsl_type *deref_type = deref->type->base_type == GLSL_TYPE_FLOAT ?
- glsl_type::float_type : glsl_type::double_type;
-
- for (unsigned i = 0; i < deref->type->vector_elements; i++) {
- ir_rvalue *chan_offset =
- add(base_offset,
- new(mem_ctx) ir_constant(deref_offset + i * matrix_stride));
- if (is_write) {
- /* If the component is not in the writemask, then don't
- * store any value.
- */
- if (!((1 << i) & write_mask))
- continue;
-
- base_ir->insert_after(ssbo_store(swizzle(deref, i, 1), chan_offset, 1));
- } else {
- if (!this->is_shader_storage) {
- base_ir->insert_before(assign(deref->clone(mem_ctx, NULL),
- ubo_load(deref_type, chan_offset),
- (1U << i)));
- } else {
- ir_call *load_ssbo = ssbo_load(deref_type, chan_offset);
- base_ir->insert_before(load_ssbo);
- ir_rvalue *value = load_ssbo->return_deref->as_rvalue()->clone(mem_ctx, NULL);
- base_ir->insert_before(assign(deref->clone(mem_ctx, NULL),
- value,
- (1U << i)));
- }
- }
+ case ssbo_store_access:
+ if (channel >= 0) {
+ base_ir->insert_after(ssbo_store(mem_ctx,
+ swizzle(deref, channel, 1),
+ offset, 1));
+ } else {
+ base_ir->insert_after(ssbo_store(mem_ctx, deref, offset, mask));
}
+ break;
+ default:
+ unreachable("invalid buffer_access_type in insert_buffer_access");
}
}
void
-lower_ubo_reference_visitor::write_to_memory(ir_dereference *deref,
+lower_ubo_reference_visitor::write_to_memory(void *mem_ctx,
+ ir_dereference *deref,
ir_variable *var,
ir_variable *write_var,
unsigned write_mask)
int matrix_columns;
unsigned packing = var->get_interface_type()->interface_packing;
+ this->buffer_access_type = ssbo_store_access;
+
/* Compute the offset to the start if the dereference as well as other
* information we need to configure the write
*/
- setup_for_load_or_store(var, deref,
+ setup_for_load_or_store(mem_ctx, var, deref,
&offset, &const_offset,
&row_major, &matrix_columns,
packing);
base_ir->insert_before(assign(write_offset, offset));
deref = new(mem_ctx) ir_dereference_variable(write_var);
- emit_access(true, deref, write_offset, const_offset,
+ emit_access(mem_ctx, true, deref, write_offset, const_offset,
row_major, matrix_columns, packing, write_mask);
}
}
ir_expression *
-lower_ubo_reference_visitor::emit_ssbo_get_buffer_size()
+lower_ubo_reference_visitor::emit_ssbo_get_buffer_size(void *mem_ctx)
{
ir_rvalue *block_ref = this->uniform_block->clone(mem_ctx, NULL);
return new(mem_ctx) ir_expression(ir_unop_get_buffer_size,
ir_dereference *deref,
ir_variable *var)
{
- mem_ctx = ralloc_parent(*rvalue);
+ void *mem_ctx = ralloc_parent(*rvalue);
ir_rvalue *base_offset = NULL;
unsigned const_offset;
unsigned packing = var->get_interface_type()->interface_packing;
int unsized_array_stride = calculate_unsized_array_stride(deref, packing);
+ this->buffer_access_type = ssbo_unsized_array_length_access;
+
/* Compute the offset to the start if the dereference as well as other
* information we need to calculate the length.
*/
- setup_for_load_or_store(var, deref,
+ setup_for_load_or_store(mem_ctx, var, deref,
&base_offset, &const_offset,
&row_major, &matrix_columns,
packing);
/* array.length() =
* max((buffer_object_size - offset_of_array) / stride_of_array, 0)
*/
- ir_expression *buffer_size = emit_ssbo_get_buffer_size();
+ ir_expression *buffer_size = emit_ssbo_get_buffer_size(mem_ctx);
ir_expression *offset_of_array = new(mem_ctx)
ir_expression(ir_binop_add, base_offset,
return;
ir_variable *var = ir->lhs->variable_referenced();
- if (!var || !var->is_in_buffer_block())
+ if (!var || !var->is_in_shader_storage_block())
return;
/* We have a write to a buffer variable, so declare a temporary and rewrite
* the assignment so that the temporary is the LHS.
*/
- mem_ctx = ralloc_parent(shader->ir);
+ void *mem_ctx = ralloc_parent(shader->ir);
const glsl_type *type = rvalue->type;
ir_variable *write_var = new(mem_ctx) ir_variable(type,
ir->lhs = new(mem_ctx) ir_dereference_variable(write_var);
/* Now we have to write the value assigned to the temporary back to memory */
- write_to_memory(deref, var, write_var, ir->write_mask);
+ write_to_memory(mem_ctx, deref, var, write_var, ir->write_mask);
+ progress = true;
+}
+
+static bool
+is_buffer_backed_variable(ir_variable *var)
+{
+ return var->is_in_buffer_block() ||
+ var->data.mode == ir_var_shader_shared;
+}
+
+bool
+lower_ubo_reference_visitor::check_for_buffer_array_copy(ir_assignment *ir)
+{
+ if (!ir || !ir->lhs || !ir->rhs)
+ return false;
+
+ /* LHS and RHS must be arrays
+ * FIXME: arrays of arrays?
+ */
+ if (!ir->lhs->type->is_array() || !ir->rhs->type->is_array())
+ return false;
+
+ /* RHS must be a buffer-backed variable. This is what can cause the problem
+ * since it would lead to a series of loads that need to live until we
+ * see the writes to the LHS.
+ */
+ ir_variable *rhs_var = ir->rhs->variable_referenced();
+ if (!rhs_var || !is_buffer_backed_variable(rhs_var))
+ return false;
+
+ /* Split the array copy into individual element copies to reduce
+ * register pressure
+ */
+ ir_dereference *rhs_deref = ir->rhs->as_dereference();
+ if (!rhs_deref)
+ return false;
+
+ ir_dereference *lhs_deref = ir->lhs->as_dereference();
+ if (!lhs_deref)
+ return false;
+
+ assert(lhs_deref->type->length == rhs_deref->type->length);
+ void *mem_ctx = ralloc_parent(shader->ir);
+
+ for (unsigned i = 0; i < lhs_deref->type->length; i++) {
+ ir_dereference *lhs_i =
+ new(mem_ctx) ir_dereference_array(lhs_deref->clone(mem_ctx, NULL),
+ new(mem_ctx) ir_constant(i));
+
+ ir_dereference *rhs_i =
+ new(mem_ctx) ir_dereference_array(rhs_deref->clone(mem_ctx, NULL),
+ new(mem_ctx) ir_constant(i));
+ ir->insert_after(assign(lhs_i, rhs_i));
+ }
+
+ ir->remove();
progress = true;
+ return true;
}
+bool
+lower_ubo_reference_visitor::check_for_buffer_struct_copy(ir_assignment *ir)
+{
+ if (!ir || !ir->lhs || !ir->rhs)
+ return false;
+
+ /* LHS and RHS must be records */
+ if (!ir->lhs->type->is_record() || !ir->rhs->type->is_record())
+ return false;
+
+ /* RHS must be a buffer-backed variable. This is what can cause the problem
+ * since it would lead to a series of loads that need to live until we
+ * see the writes to the LHS.
+ */
+ ir_variable *rhs_var = ir->rhs->variable_referenced();
+ if (!rhs_var || !is_buffer_backed_variable(rhs_var))
+ return false;
+
+ /* Split the struct copy into individual element copies to reduce
+ * register pressure
+ */
+ ir_dereference *rhs_deref = ir->rhs->as_dereference();
+ if (!rhs_deref)
+ return false;
+
+ ir_dereference *lhs_deref = ir->lhs->as_dereference();
+ if (!lhs_deref)
+ return false;
+
+ assert(lhs_deref->type->record_compare(rhs_deref->type));
+ void *mem_ctx = ralloc_parent(shader->ir);
+
+ for (unsigned i = 0; i < lhs_deref->type->length; i++) {
+ const char *field_name = lhs_deref->type->fields.structure[i].name;
+ ir_dereference *lhs_field =
+ new(mem_ctx) ir_dereference_record(lhs_deref->clone(mem_ctx, NULL),
+ field_name);
+ ir_dereference *rhs_field =
+ new(mem_ctx) ir_dereference_record(rhs_deref->clone(mem_ctx, NULL),
+ field_name);
+ ir->insert_after(assign(lhs_field, rhs_field));
+ }
+
+ ir->remove();
+ progress = true;
+ return true;
+}
ir_visitor_status
lower_ubo_reference_visitor::visit_enter(ir_assignment *ir)
{
+ /* Array and struct copies could involve large amounts of load/store
+ * operations. To improve register pressure we want to special-case
+ * these and split them into individual element copies.
+ * This way we avoid emitting all the loads for the RHS first and
+ * all the writes for the LHS second and register usage is more
+ * efficient.
+ */
+ if (check_for_buffer_array_copy(ir))
+ return visit_continue_with_parent;
+
+ if (check_for_buffer_struct_copy(ir))
+ return visit_continue_with_parent;
+
check_ssbo_unsized_array_length_assignment(ir);
check_for_ssbo_store(ir);
return rvalue_visit(ir);
/* Compute the offset to the start if the dereference and the
* block index
*/
- mem_ctx = ralloc_parent(shader->ir);
+ void *mem_ctx = ralloc_parent(shader->ir);
ir_rvalue *offset = NULL;
unsigned const_offset;
int matrix_columns;
unsigned packing = var->get_interface_type()->interface_packing;
- setup_for_load_or_store(var, deref,
+ this->buffer_access_type = ssbo_atomic_access;
+
+ setup_for_load_or_store(mem_ctx, var, deref,
&offset, &const_offset,
&row_major, &matrix_columns,
packing);
sig->is_intrinsic = true;
char func_name[64];
- sprintf(func_name, "%s_internal", ir->callee_name());
+ sprintf(func_name, "%s_ssbo", ir->callee_name());
ir_function *f = new(mem_ctx) ir_function(func_name);
f->add_signature(sig);
ir_call *
lower_ubo_reference_visitor::check_for_ssbo_atomic_intrinsic(ir_call *ir)
{
+ exec_list& params = ir->actual_parameters;
+
+ if (params.length() < 2 || params.length() > 3)
+ return ir;
+
+ ir_rvalue *rvalue =
+ ((ir_instruction *) params.get_head())->as_rvalue();
+ if (!rvalue)
+ return ir;
+
+ ir_variable *var = rvalue->variable_referenced();
+ if (!var || !var->is_in_shader_storage_block())
+ return ir;
+
const char *callee = ir->callee_name();
- if (!strcmp("__intrinsic_ssbo_atomic_add", callee) ||
- !strcmp("__intrinsic_ssbo_atomic_min", callee) ||
- !strcmp("__intrinsic_ssbo_atomic_max", callee) ||
- !strcmp("__intrinsic_ssbo_atomic_and", callee) ||
- !strcmp("__intrinsic_ssbo_atomic_or", callee) ||
- !strcmp("__intrinsic_ssbo_atomic_xor", callee) ||
- !strcmp("__intrinsic_ssbo_atomic_exchange", callee) ||
- !strcmp("__intrinsic_ssbo_atomic_comp_swap", callee)) {
+ if (!strcmp("__intrinsic_atomic_add", callee) ||
+ !strcmp("__intrinsic_atomic_min", callee) ||
+ !strcmp("__intrinsic_atomic_max", callee) ||
+ !strcmp("__intrinsic_atomic_and", callee) ||
+ !strcmp("__intrinsic_atomic_or", callee) ||
+ !strcmp("__intrinsic_atomic_xor", callee) ||
+ !strcmp("__intrinsic_atomic_exchange", callee) ||
+ !strcmp("__intrinsic_atomic_comp_swap", callee)) {
return lower_ssbo_atomic_intrinsic(ir);
}
} /* unnamed namespace */
void
-lower_ubo_reference(struct gl_shader *shader, exec_list *instructions)
+lower_ubo_reference(struct gl_shader *shader)
{
lower_ubo_reference_visitor v(shader);
*/
do {
v.progress = false;
- visit_list_elements(&v, instructions);
+ visit_list_elements(&v, shader->ir);
} while (v.progress);
}