#include "nir_vla.h"
#include "util/half_float.h"
-#define HASH(hash, data) _mesa_fnv32_1a_accumulate((hash), (data))
+static bool
+src_is_ssa(nir_src *src, void *data)
+{
+ (void) data;
+ return src->is_ssa;
+}
+
+static bool
+dest_is_ssa(nir_dest *dest, void *data)
+{
+ (void) data;
+ return dest->is_ssa;
+}
+
+ASSERTED static inline bool
+instr_each_src_and_dest_is_ssa(const nir_instr *instr)
+{
+ if (!nir_foreach_dest((nir_instr *)instr, dest_is_ssa, NULL) ||
+ !nir_foreach_src((nir_instr *)instr, src_is_ssa, NULL))
+ return false;
+
+ return true;
+}
+
+/* This function determines if uses of an instruction can safely be rewritten
+ * to use another identical instruction instead. Note that this function must
+ * be kept in sync with hash_instr() and nir_instrs_equal() -- only
+ * instructions that pass this test will be handed on to those functions, and
+ * conversely they must handle everything that this function returns true for.
+ */
+static bool
+instr_can_rewrite(const nir_instr *instr)
+{
+ /* We only handle SSA. */
+ assert(instr_each_src_and_dest_is_ssa(instr));
+
+ switch (instr->type) {
+ case nir_instr_type_alu:
+ case nir_instr_type_deref:
+ case nir_instr_type_tex:
+ case nir_instr_type_load_const:
+ case nir_instr_type_phi:
+ return true;
+ case nir_instr_type_intrinsic:
+ return nir_intrinsic_can_reorder(nir_instr_as_intrinsic(instr));
+ case nir_instr_type_call:
+ case nir_instr_type_jump:
+ case nir_instr_type_ssa_undef:
+ return false;
+ case nir_instr_type_parallel_copy:
+ default:
+ unreachable("Invalid instruction type");
+ }
+
+ return false;
+}
+
+
+#define HASH(hash, data) XXH32(&(data), sizeof(data), hash)
static uint32_t
hash_src(uint32_t hash, const nir_src *src)
case nir_deref_type_cast:
hash = HASH(hash, instr->cast.ptr_stride);
+ hash = HASH(hash, instr->cast.align_mul);
+ hash = HASH(hash, instr->cast.align_offset);
break;
case nir_deref_type_var:
}
} else {
unsigned size = instr->def.num_components * sizeof(*instr->value);
- hash = _mesa_fnv32_1a_accumulate_block(hash, instr->value, size);
+ hash = XXH32(instr->value, size, hash);
}
return hash;
hash = HASH(hash, instr->dest.ssa.bit_size);
}
- hash = _mesa_fnv32_1a_accumulate_block(hash, instr->const_index,
- info->num_indices
- * sizeof(instr->const_index[0]));
+ hash = XXH32(instr->const_index, info->num_indices * sizeof(instr->const_index[0]), hash);
return hash;
}
for (unsigned j = 0; j < 2; ++j)
hash = HASH(hash, instr->tg4_offsets[i][j]);
hash = HASH(hash, instr->texture_index);
- hash = HASH(hash, instr->texture_array_size);
hash = HASH(hash, instr->sampler_index);
+ hash = HASH(hash, instr->texture_non_uniform);
+ hash = HASH(hash, instr->sampler_non_uniform);
return hash;
}
hash_instr(const void *data)
{
const nir_instr *instr = data;
- uint32_t hash = _mesa_fnv32_1a_offset_bias;
+ uint32_t hash = 0;
switch (instr->type) {
case nir_instr_type_alu:
}
bool
-nir_const_value_negative_equal(const nir_const_value *c1,
- const nir_const_value *c2,
- unsigned components,
- nir_alu_type base_type,
- unsigned bits)
+nir_const_value_negative_equal(nir_const_value c1,
+ nir_const_value c2,
+ nir_alu_type full_type)
{
- assert(base_type == nir_alu_type_get_base_type(base_type));
- assert(base_type != nir_type_invalid);
-
- /* This can occur for 1-bit Boolean values. */
- if (bits == 1)
- return false;
-
- switch (base_type) {
- case nir_type_float:
- switch (bits) {
- case 16:
- for (unsigned i = 0; i < components; i++) {
- if (_mesa_half_to_float(c1[i].u16) !=
- -_mesa_half_to_float(c2[i].u16)) {
- return false;
- }
- }
-
- return true;
-
- case 32:
- for (unsigned i = 0; i < components; i++) {
- if (c1[i].f32 != -c2[i].f32)
- return false;
- }
-
- return true;
-
- case 64:
- for (unsigned i = 0; i < components; i++) {
- if (c1[i].f64 != -c2[i].f64)
- return false;
- }
-
- return true;
-
- default:
- unreachable("unknown bit size");
- }
-
- break;
-
- case nir_type_int:
- case nir_type_uint:
- switch (bits) {
- case 8:
- for (unsigned i = 0; i < components; i++) {
- if (c1[i].i8 != -c2[i].i8)
- return false;
- }
-
- return true;
-
- case 16:
- for (unsigned i = 0; i < components; i++) {
- if (c1[i].i16 != -c2[i].i16)
- return false;
- }
-
- return true;
- break;
+ assert(nir_alu_type_get_base_type(full_type) != nir_type_invalid);
+ assert(nir_alu_type_get_type_size(full_type) != 0);
- case 32:
- for (unsigned i = 0; i < components; i++) {
- if (c1[i].i32 != -c2[i].i32)
- return false;
- }
+ switch (full_type) {
+ case nir_type_float16:
+ return _mesa_half_to_float(c1.u16) == -_mesa_half_to_float(c2.u16);
- return true;
+ case nir_type_float32:
+ return c1.f32 == -c2.f32;
- case 64:
- for (unsigned i = 0; i < components; i++) {
- if (c1[i].i64 != -c2[i].i64)
- return false;
- }
+ case nir_type_float64:
+ return c1.f64 == -c2.f64;
- return true;
+ case nir_type_int8:
+ case nir_type_uint8:
+ return c1.i8 == -c2.i8;
- default:
- unreachable("unknown bit size");
- }
+ case nir_type_int16:
+ case nir_type_uint16:
+ return c1.i16 == -c2.i16;
- break;
+ case nir_type_int32:
+ case nir_type_uint32:
+ return c1.i32 == -c2.i32;
- case nir_type_bool:
- return false;
+ case nir_type_int64:
+ case nir_type_uint64:
+ return c1.i64 == -c2.i64;
default:
break;
* This function does not detect the general case when \p alu1 and \p alu2 are
* SSA values that are the negations of each other (e.g., \p alu1 represents
* (a * b) and \p alu2 represents (-a * b)).
+ *
+ * \warning
+ * It is the responsibility of the caller to ensure that the component counts,
+ * write masks, and base types of the sources being compared are compatible.
*/
bool
nir_alu_srcs_negative_equal(const nir_alu_instr *alu1,
const nir_alu_instr *alu2,
unsigned src1, unsigned src2)
{
+#ifndef NDEBUG
+ for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++) {
+ assert(nir_alu_instr_channel_used(alu1, src1, i) ==
+ nir_alu_instr_channel_used(alu2, src2, i));
+ }
+
+ if (nir_op_infos[alu1->op].input_types[src1] == nir_type_float) {
+ assert(nir_op_infos[alu1->op].input_types[src1] ==
+ nir_op_infos[alu2->op].input_types[src2]);
+ } else {
+ assert(nir_op_infos[alu1->op].input_types[src1] == nir_type_int);
+ assert(nir_op_infos[alu2->op].input_types[src2] == nir_type_int);
+ }
+#endif
+
if (alu1->src[src1].abs != alu2->src[src2].abs)
return false;
if (const2 == NULL)
return false;
- /* FINISHME: Apply the swizzle? */
- return nir_const_value_negative_equal(const1,
- const2,
- nir_ssa_alu_instr_src_components(alu1, src1),
- nir_op_infos[alu1->op].input_types[src1],
- alu1->dest.dest.ssa.bit_size);
+ if (nir_src_bit_size(alu1->src[src1].src) !=
+ nir_src_bit_size(alu2->src[src2].src))
+ return false;
+
+ const nir_alu_type full_type = nir_op_infos[alu1->op].input_types[src1] |
+ nir_src_bit_size(alu1->src[src1].src);
+ for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++) {
+ if (nir_alu_instr_channel_used(alu1, src1, i) &&
+ !nir_const_value_negative_equal(const1[alu1->src[src1].swizzle[i]],
+ const2[alu2->src[src2].swizzle[i]],
+ full_type))
+ return false;
+ }
+
+ return true;
}
uint8_t alu1_swizzle[4] = {0};
* the same hash for (ignoring collisions, of course).
*/
-static bool
+bool
nir_instrs_equal(const nir_instr *instr1, const nir_instr *instr2)
{
+ assert(instr_can_rewrite(instr1) && instr_can_rewrite(instr2));
+
if (instr1->type != instr2->type)
return false;
break;
case nir_deref_type_cast:
- if (deref1->cast.ptr_stride != deref2->cast.ptr_stride)
+ if (deref1->cast.ptr_stride != deref2->cast.ptr_stride ||
+ deref1->cast.align_mul != deref2->cast.align_mul ||
+ deref1->cast.align_offset != deref2->cast.align_offset)
return false;
break;
tex1->is_new_style_shadow != tex2->is_new_style_shadow ||
tex1->component != tex2->component ||
tex1->texture_index != tex2->texture_index ||
- tex1->texture_array_size != tex2->texture_array_size ||
tex1->sampler_index != tex2->sampler_index) {
return false;
}
unreachable("All cases in the above switch should return");
}
-static bool
-src_is_ssa(nir_src *src, void *data)
-{
- (void) data;
- return src->is_ssa;
-}
-
-static bool
-dest_is_ssa(nir_dest *dest, void *data)
-{
- (void) data;
- return dest->is_ssa;
-}
-
-static inline bool
-instr_each_src_and_dest_is_ssa(nir_instr *instr)
-{
- if (!nir_foreach_dest(instr, dest_is_ssa, NULL) ||
- !nir_foreach_src(instr, src_is_ssa, NULL))
- return false;
-
- return true;
-}
-
-/* This function determines if uses of an instruction can safely be rewritten
- * to use another identical instruction instead. Note that this function must
- * be kept in sync with hash_instr() and nir_instrs_equal() -- only
- * instructions that pass this test will be handed on to those functions, and
- * conversely they must handle everything that this function returns true for.
- */
-
-static bool
-instr_can_rewrite(nir_instr *instr)
-{
- /* We only handle SSA. */
- assert(instr_each_src_and_dest_is_ssa(instr));
-
- switch (instr->type) {
- case nir_instr_type_alu:
- case nir_instr_type_deref:
- case nir_instr_type_tex:
- case nir_instr_type_load_const:
- case nir_instr_type_phi:
- return true;
- case nir_instr_type_intrinsic:
- return nir_intrinsic_can_reorder(nir_instr_as_intrinsic(instr));
- case nir_instr_type_call:
- case nir_instr_type_jump:
- case nir_instr_type_ssa_undef:
- return false;
- case nir_instr_type_parallel_copy:
- default:
- unreachable("Invalid instruction type");
- }
-
- return false;
-}
-
static nir_ssa_def *
nir_instr_get_dest_ssa_def(nir_instr *instr)
{