*/
#include "nir.h"
+#include "nir_builder.h"
#include "nir_control_flow_private.h"
#include "util/half_float.h"
#include <limits.h>
}
}
+nir_ssa_def *
+nir_instr_ssa_def(nir_instr *instr)
+{
+ switch (instr->type) {
+ case nir_instr_type_alu:
+ assert(nir_instr_as_alu(instr)->dest.dest.is_ssa);
+ return &nir_instr_as_alu(instr)->dest.dest.ssa;
+
+ case nir_instr_type_deref:
+ assert(nir_instr_as_deref(instr)->dest.is_ssa);
+ return &nir_instr_as_deref(instr)->dest.ssa;
+
+ case nir_instr_type_tex:
+ assert(nir_instr_as_tex(instr)->dest.is_ssa);
+ return &nir_instr_as_tex(instr)->dest.ssa;
+
+ case nir_instr_type_intrinsic: {
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+ if (nir_intrinsic_infos[intrin->intrinsic].has_dest) {
+ assert(intrin->dest.is_ssa);
+ return &intrin->dest.ssa;
+ } else {
+ return NULL;
+ }
+ }
+
+ case nir_instr_type_phi:
+ assert(nir_instr_as_phi(instr)->dest.is_ssa);
+ return &nir_instr_as_phi(instr)->dest.ssa;
+
+ case nir_instr_type_parallel_copy:
+ unreachable("Parallel copies are unsupported by this function");
+
+ case nir_instr_type_load_const:
+ return &nir_instr_as_load_const(instr)->def;
+
+ case nir_instr_type_ssa_undef:
+ return &nir_instr_as_ssa_undef(instr)->def;
+
+ case nir_instr_type_call:
+ case nir_instr_type_jump:
+ return NULL;
+ }
+
+ unreachable("Invalid instruction type");
+}
+
static bool
visit_src(nir_src *src, nir_foreach_src_cb cb, void *state)
{
return nir_foreach_dest(instr, visit_dest_indirect, &dest_state);
}
-int64_t
-nir_src_comp_as_int(nir_src src, unsigned comp)
-{
- assert(nir_src_is_const(src));
- nir_load_const_instr *load = nir_instr_as_load_const(src.ssa->parent_instr);
-
- assert(comp < load->def.num_components);
- switch (load->def.bit_size) {
- /* int1_t uses 0/-1 convention */
- case 1: return -(int)load->value[comp].b;
- case 8: return load->value[comp].i8;
- case 16: return load->value[comp].i16;
- case 32: return load->value[comp].i32;
- case 64: return load->value[comp].i64;
- default:
- unreachable("Invalid bit size");
- }
-}
-
-uint64_t
-nir_src_comp_as_uint(nir_src src, unsigned comp)
+nir_const_value
+nir_const_value_for_float(double f, unsigned bit_size)
{
- assert(nir_src_is_const(src));
- nir_load_const_instr *load = nir_instr_as_load_const(src.ssa->parent_instr);
+ nir_const_value v;
+ memset(&v, 0, sizeof(v));
- assert(comp < load->def.num_components);
- switch (load->def.bit_size) {
- case 1: return load->value[comp].b;
- case 8: return load->value[comp].u8;
- case 16: return load->value[comp].u16;
- case 32: return load->value[comp].u32;
- case 64: return load->value[comp].u64;
+ switch (bit_size) {
+ case 16:
+ v.u16 = _mesa_float_to_half(f);
+ break;
+ case 32:
+ v.f32 = f;
+ break;
+ case 64:
+ v.f64 = f;
+ break;
default:
unreachable("Invalid bit size");
}
-}
-
-bool
-nir_src_comp_as_bool(nir_src src, unsigned comp)
-{
- int64_t i = nir_src_comp_as_int(src, comp);
- /* Booleans of any size use 0/-1 convention */
- assert(i == 0 || i == -1);
-
- return i;
+ return v;
}
double
-nir_src_comp_as_float(nir_src src, unsigned comp)
+nir_const_value_as_float(nir_const_value value, unsigned bit_size)
{
- assert(nir_src_is_const(src));
- nir_load_const_instr *load = nir_instr_as_load_const(src.ssa->parent_instr);
-
- assert(comp < load->def.num_components);
- switch (load->def.bit_size) {
- case 16: return _mesa_half_to_float(load->value[comp].u16);
- case 32: return load->value[comp].f32;
- case 64: return load->value[comp].f64;
+ switch (bit_size) {
+ case 16: return _mesa_half_to_float(value.u16);
+ case 32: return value.f32;
+ case 64: return value.f64;
default:
unreachable("Invalid bit size");
}
}
-int64_t
-nir_src_as_int(nir_src src)
-{
- assert(nir_src_num_components(src) == 1);
- return nir_src_comp_as_int(src, 0);
-}
-
-uint64_t
-nir_src_as_uint(nir_src src)
-{
- assert(nir_src_num_components(src) == 1);
- return nir_src_comp_as_uint(src, 0);
-}
-
-bool
-nir_src_as_bool(nir_src src)
-{
- assert(nir_src_num_components(src) == 1);
- return nir_src_comp_as_bool(src, 0);
-}
-
-double
-nir_src_as_float(nir_src src)
-{
- assert(nir_src_num_components(src) == 1);
- return nir_src_comp_as_float(src, 0);
-}
-
nir_const_value *
nir_src_as_const_value(nir_src src)
{
return index;
}
+static nir_instr *
+cursor_next_instr(nir_cursor cursor)
+{
+ switch (cursor.option) {
+ case nir_cursor_before_block:
+ for (nir_block *block = cursor.block; block;
+ block = nir_block_cf_tree_next(block)) {
+ nir_instr *instr = nir_block_first_instr(block);
+ if (instr)
+ return instr;
+ }
+ return NULL;
+
+ case nir_cursor_after_block:
+ cursor.block = nir_block_cf_tree_next(cursor.block);
+ if (cursor.block == NULL)
+ return NULL;
+
+ cursor.option = nir_cursor_before_block;
+ return cursor_next_instr(cursor);
+
+ case nir_cursor_before_instr:
+ return cursor.instr;
+
+ case nir_cursor_after_instr:
+ if (nir_instr_next(cursor.instr))
+ return nir_instr_next(cursor.instr);
+
+ cursor.option = nir_cursor_after_block;
+ cursor.block = cursor.instr->block;
+ return cursor_next_instr(cursor);
+ }
+
+ unreachable("Inavlid cursor option");
+}
+
+static bool
+dest_is_ssa(nir_dest *dest, void *_state)
+{
+ return dest->is_ssa;
+}
+
+bool
+nir_function_impl_lower_instructions(nir_function_impl *impl,
+ nir_instr_filter_cb filter,
+ nir_lower_instr_cb lower,
+ void *cb_data)
+{
+ nir_builder b;
+ nir_builder_init(&b, impl);
+
+ nir_metadata preserved = nir_metadata_block_index |
+ nir_metadata_dominance;
+
+ bool progress = false;
+ nir_cursor iter = nir_before_cf_list(&impl->body);
+ nir_instr *instr;
+ while ((instr = cursor_next_instr(iter)) != NULL) {
+ if (filter && !filter(instr, cb_data)) {
+ iter = nir_after_instr(instr);
+ continue;
+ }
+
+ assert(nir_foreach_dest(instr, dest_is_ssa, NULL));
+ nir_ssa_def *old_def = nir_instr_ssa_def(instr);
+ if (old_def == NULL) {
+ iter = nir_after_instr(instr);
+ continue;
+ }
+
+ /* We're about to ask the callback to generate a replacement for instr.
+ * Save off the uses from instr's SSA def so we know what uses to
+ * rewrite later. If we use nir_ssa_def_rewrite_uses, it fails in the
+ * case where the generated replacement code uses the result of instr
+ * itself. If we use nir_ssa_def_rewrite_uses_after (which is the
+ * normal solution to this problem), it doesn't work well if control-
+ * flow is inserted as part of the replacement, doesn't handle cases
+ * where the replacement is something consumed by instr, and suffers
+ * from performance issues. This is the only way to 100% guarantee
+ * that we rewrite the correct set efficiently.
+ */
+ struct list_head old_uses, old_if_uses;
+ list_replace(&old_def->uses, &old_uses);
+ list_inithead(&old_def->uses);
+ list_replace(&old_def->if_uses, &old_if_uses);
+ list_inithead(&old_def->if_uses);
+
+ b.cursor = nir_after_instr(instr);
+ nir_ssa_def *new_def = lower(&b, instr, cb_data);
+ if (new_def && new_def != NIR_LOWER_INSTR_PROGRESS) {
+ assert(old_def != NULL);
+ if (new_def->parent_instr->block != instr->block)
+ preserved = nir_metadata_none;
+
+ nir_src new_src = nir_src_for_ssa(new_def);
+ list_for_each_entry_safe(nir_src, use_src, &old_uses, use_link)
+ nir_instr_rewrite_src(use_src->parent_instr, use_src, new_src);
+
+ list_for_each_entry_safe(nir_src, use_src, &old_if_uses, use_link)
+ nir_if_rewrite_condition(use_src->parent_if, new_src);
+
+ if (list_empty(&old_def->uses) && list_empty(&old_def->if_uses)) {
+ iter = nir_instr_remove(instr);
+ } else {
+ iter = nir_after_instr(instr);
+ }
+ progress = true;
+ } else {
+ /* We didn't end up lowering after all. Put the uses back */
+ if (old_def) {
+ list_replace(&old_uses, &old_def->uses);
+ list_replace(&old_if_uses, &old_def->if_uses);
+ }
+ iter = nir_after_instr(instr);
+
+ if (new_def == NIR_LOWER_INSTR_PROGRESS)
+ progress = true;
+ }
+ }
+
+ if (progress) {
+ nir_metadata_preserve(impl, preserved);
+ } else {
+#ifndef NDEBUG
+ impl->valid_metadata &= ~nir_metadata_not_properly_reset;
+#endif
+ }
+
+ return progress;
+}
+
+bool
+nir_shader_lower_instructions(nir_shader *shader,
+ nir_instr_filter_cb filter,
+ nir_lower_instr_cb lower,
+ void *cb_data)
+{
+ bool progress = false;
+
+ nir_foreach_function(function, shader) {
+ if (function->impl &&
+ nir_function_impl_lower_instructions(function->impl,
+ filter, lower, cb_data))
+ progress = true;
+ }
+
+ return progress;
+}
+
nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val)
{
return nir_intrinsic_load_invocation_id;
case SYSTEM_VALUE_FRAG_COORD:
return nir_intrinsic_load_frag_coord;
+ case SYSTEM_VALUE_POINT_COORD:
+ return nir_intrinsic_load_point_coord;
case SYSTEM_VALUE_FRONT_FACE:
return nir_intrinsic_load_front_face;
case SYSTEM_VALUE_SAMPLE_ID:
return nir_intrinsic_load_tess_level_outer;
case SYSTEM_VALUE_TESS_LEVEL_INNER:
return nir_intrinsic_load_tess_level_inner;
+ case SYSTEM_VALUE_TESS_LEVEL_OUTER_DEFAULT:
+ return nir_intrinsic_load_tess_level_outer_default;
+ case SYSTEM_VALUE_TESS_LEVEL_INNER_DEFAULT:
+ return nir_intrinsic_load_tess_level_inner_default;
case SYSTEM_VALUE_VERTICES_IN:
return nir_intrinsic_load_patch_vertices_in;
case SYSTEM_VALUE_HELPER_INVOCATION:
return nir_intrinsic_load_global_invocation_index;
case SYSTEM_VALUE_WORK_DIM:
return nir_intrinsic_load_work_dim;
+ case SYSTEM_VALUE_USER_DATA_AMD:
+ return nir_intrinsic_load_user_data_amd;
default:
unreachable("system value does not directly correspond to intrinsic");
}
return SYSTEM_VALUE_INVOCATION_ID;
case nir_intrinsic_load_frag_coord:
return SYSTEM_VALUE_FRAG_COORD;
+ case nir_intrinsic_load_point_coord:
+ return SYSTEM_VALUE_POINT_COORD;
case nir_intrinsic_load_front_face:
return SYSTEM_VALUE_FRONT_FACE;
case nir_intrinsic_load_sample_id:
return SYSTEM_VALUE_TESS_LEVEL_OUTER;
case nir_intrinsic_load_tess_level_inner:
return SYSTEM_VALUE_TESS_LEVEL_INNER;
+ case nir_intrinsic_load_tess_level_outer_default:
+ return SYSTEM_VALUE_TESS_LEVEL_OUTER_DEFAULT;
+ case nir_intrinsic_load_tess_level_inner_default:
+ return SYSTEM_VALUE_TESS_LEVEL_INNER_DEFAULT;
case nir_intrinsic_load_patch_vertices_in:
return SYSTEM_VALUE_VERTICES_IN;
case nir_intrinsic_load_helper_invocation:
return SYSTEM_VALUE_LOCAL_GROUP_SIZE;
case nir_intrinsic_load_global_invocation_id:
return SYSTEM_VALUE_GLOBAL_INVOCATION_ID;
+ case nir_intrinsic_load_user_data_amd:
+ return SYSTEM_VALUE_USER_DATA_AMD;
default:
unreachable("intrinsic doesn't produce a system value");
}
nir_rewrite_image_intrinsic(nir_intrinsic_instr *intrin, nir_ssa_def *src,
bool bindless)
{
+ enum gl_access_qualifier access = nir_intrinsic_access(intrin);
+
switch (intrin->intrinsic) {
#define CASE(op) \
case nir_intrinsic_image_deref_##op: \
CASE(load)
CASE(store)
CASE(atomic_add)
- CASE(atomic_min)
- CASE(atomic_max)
+ CASE(atomic_imin)
+ CASE(atomic_umin)
+ CASE(atomic_imax)
+ CASE(atomic_umax)
CASE(atomic_and)
CASE(atomic_or)
CASE(atomic_xor)
nir_intrinsic_set_image_dim(intrin, glsl_get_sampler_dim(deref->type));
nir_intrinsic_set_image_array(intrin, glsl_sampler_type_is_array(deref->type));
+ nir_intrinsic_set_access(intrin, access | var->data.image.access);
nir_intrinsic_set_format(intrin, var->data.image.format);
nir_instr_rewrite_src(&intrin->instr, &intrin->src[0],