#include "compiler/nir_types.h"
#include "compiler/shader_enums.h"
#include "compiler/shader_info.h"
+#define XXH_INLINE_ALL
+#include "util/xxhash.h"
#include <stdio.h>
#ifndef NDEBUG
* The values in this enum are carefully chosen so that the sized type is
* just the unsized type OR the number of bits.
*/
-typedef enum {
+typedef enum PACKED {
nir_type_invalid = 0, /* Not a valid type */
nir_type_int = 2,
nir_type_uint = 4,
return type & NIR_ALU_TYPE_SIZE_MASK;
}
-static inline unsigned
+static inline nir_alu_type
nir_alu_type_get_base_type(nir_alu_type type)
{
- return type & NIR_ALU_TYPE_BASE_TYPE_MASK;
+ return (nir_alu_type)(type & NIR_ALU_TYPE_BASE_TYPE_MASK);
}
static inline nir_alu_type
typedef struct {
const char *name;
- unsigned num_inputs;
+ uint8_t num_inputs;
/**
* The number of components in the output
* though output_size is zero; in that case, the inputs with a zero
* size act per-component, while the inputs with non-zero size don't.
*/
- unsigned output_size;
+ uint8_t output_size;
/**
* The type of vector that the instruction outputs. Note that the
/**
* The number of components in each input
*/
- unsigned input_sizes[NIR_MAX_VEC_COMPONENTS];
+ uint8_t input_sizes[NIR_MAX_VEC_COMPONENTS];
/**
* The type of vector that each input takes. Note that negate and
} nir_memory_semantics;
typedef enum {
+ NIR_SCOPE_NONE,
NIR_SCOPE_INVOCATION,
NIR_SCOPE_SUBGROUP,
NIR_SCOPE_WORKGROUP,
*/
NIR_INTRINSIC_MEMORY_SCOPE,
+ /**
+ * Value of nir_scope.
+ */
+ NIR_INTRINSIC_EXECUTION_SCOPE,
+
NIR_INTRINSIC_NUM_INDEX_FLAGS,
} nir_intrinsic_index_flag;
typedef struct {
const char *name;
- unsigned num_srcs; /** < number of register/SSA inputs */
+ uint8_t num_srcs; /** < number of register/SSA inputs */
/** number of components of each input register
*
* intrinsic consumes however many components are provided and it is not
* validated at all.
*/
- int src_components[NIR_INTRINSIC_MAX_INPUTS];
+ int8_t src_components[NIR_INTRINSIC_MAX_INPUTS];
bool has_dest;
* If this value is 0, the number of components is given by the
* num_components field of nir_intrinsic_instr.
*/
- unsigned dest_components;
+ uint8_t dest_components;
/** bitfield of legal bit sizes */
- unsigned dest_bit_sizes;
+ uint8_t dest_bit_sizes;
/** the number of constant indices used by the intrinsic */
- unsigned num_indices;
+ uint8_t num_indices;
/** indicates the usage of intr->const_index[n] */
- unsigned index_map[NIR_INTRINSIC_NUM_INDEX_FLAGS];
+ uint8_t index_map[NIR_INTRINSIC_NUM_INDEX_FLAGS];
/** semantic flags for calls to this intrinsic */
nir_intrinsic_semantic_flag flags;
extern const nir_intrinsic_info nir_intrinsic_infos[nir_num_intrinsics];
static inline unsigned
-nir_intrinsic_src_components(nir_intrinsic_instr *intr, unsigned srcn)
+nir_intrinsic_src_components(const nir_intrinsic_instr *intr, unsigned srcn)
{
const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
assert(srcn < info->num_srcs);
return intr->num_components;
}
+/**
+ * Helper to copy const_index[] from src to dst, without assuming they
+ * match in order.
+ */
+static inline void
+nir_intrinsic_copy_const_indices(nir_intrinsic_instr *dst, nir_intrinsic_instr *src)
+{
+ if (src->intrinsic == dst->intrinsic) {
+ memcpy(dst->const_index, src->const_index, sizeof(dst->const_index));
+ return;
+ }
+
+ const nir_intrinsic_info *src_info = &nir_intrinsic_infos[src->intrinsic];
+ const nir_intrinsic_info *dst_info = &nir_intrinsic_infos[dst->intrinsic];
+
+ for (unsigned i = 0; i < NIR_INTRINSIC_NUM_INDEX_FLAGS; i++) {
+ if (src_info->index_map[i] == 0)
+ continue;
+
+ /* require that dst instruction also uses the same const_index[]: */
+ assert(dst_info->index_map[i] > 0);
+
+ dst->const_index[dst_info->index_map[i] - 1] =
+ src->const_index[src_info->index_map[i] - 1];
+ }
+}
+
#define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
static inline type \
nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
INTRINSIC_IDX_ACCESSORS(memory_semantics, MEMORY_SEMANTICS, nir_memory_semantics)
INTRINSIC_IDX_ACCESSORS(memory_modes, MEMORY_MODES, nir_variable_mode)
INTRINSIC_IDX_ACCESSORS(memory_scope, MEMORY_SCOPE, nir_scope)
+INTRINSIC_IDX_ACCESSORS(execution_scope, EXECUTION_SCOPE, nir_scope)
static inline void
nir_intrinsic_set_align(nir_intrinsic_instr *intrin,
} nir_load_const_instr;
typedef enum {
+ /** Return from a function
+ *
+ * This instruction is a classic function return. It jumps to
+ * nir_function_impl::end_block. No return value is provided in this
+ * instruction. Instead, the function is expected to write any return
+ * data to a deref passed in from the caller.
+ */
nir_jump_return,
+
+ /** Break out of the inner-most loop
+ *
+ * This has the same semantics as C's "break" statement.
+ */
nir_jump_break,
+
+ /** Jump back to the top of the inner-most loop
+ *
+ * This has the same semantics as C's "continue" statement assuming that a
+ * NIR loop is implemented as "while (1) { body }".
+ */
nir_jump_continue,
} nir_jump_type;
* determine. Most passes shouldn't preserve this metadata type.
*/
nir_metadata_loop_analysis = 0x10,
+
+ /** All metadata
+ *
+ * This includes all nir_metadata flags except not_properly_reset. Passes
+ * which do not change the shader in any way should call
+ *
+ * nir_metadata_preserve(impl, nir_metadata_all);
+ */
+ nir_metadata_all = ~nir_metadata_not_properly_reset,
} nir_metadata;
typedef struct {
bool vectorize_io;
bool lower_to_scalar;
+ /**
+ * Whether nir_opt_vectorize should only create 16-bit 2D vectors.
+ */
+ bool vectorize_vec2_16bit;
+
/**
* Should the linker unify inputs_read/outputs_written between adjacent
* shader stages which are linked into a single program?
* to imul with masked inputs and iadd */
bool has_umad24;
- /* Whether to generate only scoped_memory_barrier intrinsics instead of the
- * set of memory barrier intrinsics based on GLSL.
+ /* Whether to generate only scoped_barrier intrinsics instead of the set of
+ * memory and control barrier intrinsics based on GLSL.
*/
- bool use_scoped_memory_barrier;
+ bool use_scoped_barrier;
/**
* Is this the Intel vec4 backend?
void nir_metadata_require(nir_function_impl *impl, nir_metadata required, ...);
/** dirties all but the preserved metadata */
void nir_metadata_preserve(nir_function_impl *impl, nir_metadata preserved);
+/** Preserves all metadata for the given shader */
+void nir_shader_preserve_all_metadata(nir_shader *shader);
/** creates an instruction with default swizzle/writemask/etc. with NULL registers */
nir_alu_instr *nir_alu_instr_create(nir_shader *shader, nir_op op);
void *state);
bool nir_foreach_dest(nir_instr *instr, nir_foreach_dest_cb cb, void *state);
bool nir_foreach_src(nir_instr *instr, nir_foreach_src_cb cb, void *state);
+bool nir_foreach_phi_src_leaving_block(nir_block *instr,
+ nir_foreach_src_cb cb,
+ void *state);
nir_const_value *nir_src_as_const_value(nir_src src);
*/
nir_address_format_32bit_index_offset,
+ /**
+ * An address format which is comprised of a vec3 where the first two
+ * components specify the buffer and the third is an offset.
+ */
+ nir_address_format_vec2_index_32bit_offset,
+
/**
* An address format which is a simple 32-bit offset.
*/
nir_address_format_bit_size(nir_address_format addr_format)
{
switch (addr_format) {
- case nir_address_format_32bit_global: return 32;
- case nir_address_format_64bit_global: return 64;
- case nir_address_format_64bit_bounded_global: return 32;
- case nir_address_format_32bit_index_offset: return 32;
- case nir_address_format_32bit_offset: return 32;
- case nir_address_format_logical: return 32;
+ case nir_address_format_32bit_global: return 32;
+ case nir_address_format_64bit_global: return 64;
+ case nir_address_format_64bit_bounded_global: return 32;
+ case nir_address_format_32bit_index_offset: return 32;
+ case nir_address_format_vec2_index_32bit_offset: return 32;
+ case nir_address_format_32bit_offset: return 32;
+ case nir_address_format_logical: return 32;
}
unreachable("Invalid address format");
}
nir_address_format_num_components(nir_address_format addr_format)
{
switch (addr_format) {
- case nir_address_format_32bit_global: return 1;
- case nir_address_format_64bit_global: return 1;
- case nir_address_format_64bit_bounded_global: return 4;
- case nir_address_format_32bit_index_offset: return 2;
- case nir_address_format_32bit_offset: return 1;
- case nir_address_format_logical: return 1;
+ case nir_address_format_32bit_global: return 1;
+ case nir_address_format_64bit_global: return 1;
+ case nir_address_format_64bit_bounded_global: return 4;
+ case nir_address_format_32bit_index_offset: return 2;
+ case nir_address_format_vec2_index_32bit_offset: return 3;
+ case nir_address_format_32bit_offset: return 1;
+ case nir_address_format_logical: return 1;
}
unreachable("Invalid address format");
}
bool nir_remove_dead_derefs(nir_shader *shader);
bool nir_remove_dead_derefs_impl(nir_function_impl *impl);
-bool nir_remove_dead_variables(nir_shader *shader, nir_variable_mode modes);
+bool nir_remove_dead_variables(nir_shader *shader, nir_variable_mode modes,
+ bool (*can_remove_var)(nir_variable *var));
bool nir_lower_variable_initializers(nir_shader *shader,
nir_variable_mode modes);
void nir_lower_io_to_scalar_early(nir_shader *shader, nir_variable_mode mask);
bool nir_lower_io_to_vector(nir_shader *shader, nir_variable_mode mask);
+bool nir_lower_fragcolor(nir_shader *shader);
void nir_lower_fragcoord_wtrans(nir_shader *shader);
void nir_lower_viewport_transform(nir_shader *shader);
bool nir_lower_uniforms_to_ubo(nir_shader *shader, int multiplier);
bool lower_subgroup_masks:1;
bool lower_shuffle:1;
bool lower_shuffle_to_32bit:1;
+ bool lower_shuffle_to_swizzle_amd:1;
bool lower_quad:1;
bool lower_quad_broadcast_dynamic:1;
bool lower_quad_broadcast_dynamic_to_const:1;
bool nir_lower_clip_fs(nir_shader *shader, unsigned ucp_enables,
bool use_clipdist_array);
bool nir_lower_clip_cull_distance_arrays(nir_shader *nir);
+bool nir_lower_clip_disable(nir_shader *shader, unsigned clip_plane_enable);
void nir_lower_point_size_mov(nir_shader *shader,
const gl_state_index16 *pointsize_state_tokens);
bool nir_lower_frexp(nir_shader *nir);
-void nir_lower_two_sided_color(nir_shader *shader);
+void nir_lower_two_sided_color(nir_shader *shader, bool face_sysval);
bool nir_lower_clamp_color_outputs(nir_shader *shader);
const nir_lower_wpos_ytransform_options *options);
bool nir_lower_wpos_center(nir_shader *shader, const bool for_sample_shading);
+bool nir_lower_wrmasks(nir_shader *shader, nir_instr_filter_cb cb, const void *data);
+
bool nir_lower_fb_read(nir_shader *shader);
typedef struct nir_lower_drawpixels_options {
nir_should_vectorize_mem_func callback,
nir_variable_mode robust_modes);
-void nir_schedule(nir_shader *shader, int threshold);
-
void nir_strip(nir_shader *shader);
void nir_sweep(nir_shader *shader);
return nir_variable_is_in_ubo(var) || nir_variable_is_in_ssbo(var);
}
+typedef struct nir_unsigned_upper_bound_config {
+ unsigned min_subgroup_size;
+ unsigned max_subgroup_size;
+ unsigned max_work_group_invocations;
+ unsigned max_work_group_count[3];
+ unsigned max_work_group_size[3];
+
+ uint32_t vertex_attrib_max[32];
+} nir_unsigned_upper_bound_config;
+
+uint32_t
+nir_unsigned_upper_bound(nir_shader *shader, struct hash_table *range_ht,
+ nir_ssa_scalar scalar,
+ const nir_unsigned_upper_bound_config *config);
+
+bool
+nir_addition_might_overflow(nir_shader *shader, struct hash_table *range_ht,
+ nir_ssa_scalar ssa, unsigned const_val,
+ const nir_unsigned_upper_bound_config *config);
+
#ifdef __cplusplus
} /* extern "C" */
#endif
projects / mesa.git / blobdiff