nir_ssa_def *d_lo = nir_unpack_64_2x32_split_x(b, d);
nir_ssa_def *d_hi = nir_unpack_64_2x32_split_y(b, d);
- nir_const_value v = { .u32 = { 0, 0, 0, 0 } };
- nir_ssa_def *q_lo = nir_build_imm(b, n->num_components, 32, v);
- nir_ssa_def *q_hi = nir_build_imm(b, n->num_components, 32, v);
+ nir_ssa_def *q_lo = nir_imm_zero(b, n->num_components, 32);
+ nir_ssa_def *q_hi = nir_imm_zero(b, n->num_components, 32);
nir_ssa_def *n_hi_before_if = n_hi;
nir_ssa_def *q_hi_before_if = q_hi;
return nir_bcsel(b, n_is_neg, nir_ineg(b, r), r);
}
+static nir_ssa_def *
+lower_extract(nir_builder *b, nir_op op, nir_ssa_def *x, nir_ssa_def *c)
+{
+ assert(op == nir_op_extract_u8 || op == nir_op_extract_i8 ||
+ op == nir_op_extract_u16 || op == nir_op_extract_i16);
+
+ const int chunk = nir_src_as_uint(nir_src_for_ssa(c));
+ const int chunk_bits =
+ (op == nir_op_extract_u8 || op == nir_op_extract_i8) ? 8 : 16;
+ const int num_chunks_in_32 = 32 / chunk_bits;
+
+ nir_ssa_def *extract32;
+ if (chunk < num_chunks_in_32) {
+ extract32 = nir_build_alu(b, op, nir_unpack_64_2x32_split_x(b, x),
+ nir_imm_int(b, chunk),
+ NULL, NULL);
+ } else {
+ extract32 = nir_build_alu(b, op, nir_unpack_64_2x32_split_y(b, x),
+ nir_imm_int(b, chunk - num_chunks_in_32),
+ NULL, NULL);
+ }
+
+ if (op == nir_op_extract_i8 || op == nir_op_extract_i16)
+ return lower_i2i64(b, extract32);
+ else
+ return lower_u2u64(b, extract32);
+}
+
+static nir_ssa_def *
+lower_ufind_msb64(nir_builder *b, nir_ssa_def *x)
+{
+
+ nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
+ nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
+ nir_ssa_def *lo_count = nir_ufind_msb(b, x_lo);
+ nir_ssa_def *hi_count = nir_ufind_msb(b, x_hi);
+ nir_ssa_def *valid_hi_bits = nir_ine(b, x_hi, nir_imm_int(b, 0));
+ nir_ssa_def *hi_res = nir_iadd(b, nir_imm_intN_t(b, 32, 32), hi_count);
+ return nir_bcsel(b, valid_hi_bits, hi_res, lo_count);
+}
+
nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode)
{
case nir_op_ishr:
case nir_op_ushr:
return nir_lower_shift64;
+ case nir_op_extract_u8:
+ case nir_op_extract_i8:
+ case nir_op_extract_u16:
+ case nir_op_extract_i16:
+ return nir_lower_extract64;
+ case nir_op_ufind_msb:
+ return nir_lower_ufind_msb64;
default:
return 0;
}
}
static nir_ssa_def *
-lower_int64_alu_instr(nir_builder *b, nir_alu_instr *alu)
+lower_int64_alu_instr(nir_builder *b, nir_instr *instr, void *_state)
{
+ nir_alu_instr *alu = nir_instr_as_alu(instr);
+
nir_ssa_def *src[4];
for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++)
src[i] = nir_ssa_for_alu_src(b, alu, i);
return lower_ishr64(b, src[0], src[1]);
case nir_op_ushr:
return lower_ushr64(b, src[0], src[1]);
+ case nir_op_extract_u8:
+ case nir_op_extract_i8:
+ case nir_op_extract_u16:
+ case nir_op_extract_i16:
+ return lower_extract(b, alu->op, src[0], src[1]);
+ case nir_op_ufind_msb:
+ return lower_ufind_msb64(b, src[0]);
+ break;
default:
unreachable("Invalid ALU opcode to lower");
}
}
static bool
-lower_int64_impl(nir_function_impl *impl, nir_lower_int64_options options)
-{
- nir_builder b;
- nir_builder_init(&b, impl);
-
- bool progress = false;
- nir_foreach_block(block, impl) {
- nir_foreach_instr_safe(instr, block) {
- if (instr->type != nir_instr_type_alu)
- continue;
-
- nir_alu_instr *alu = nir_instr_as_alu(instr);
- switch (alu->op) {
- case nir_op_i2b1:
- case nir_op_i2i32:
- case nir_op_u2u32:
- assert(alu->src[0].src.is_ssa);
- if (alu->src[0].src.ssa->bit_size != 64)
- continue;
- break;
- case nir_op_bcsel:
- assert(alu->src[1].src.is_ssa);
- assert(alu->src[2].src.is_ssa);
- assert(alu->src[1].src.ssa->bit_size ==
- alu->src[2].src.ssa->bit_size);
- if (alu->src[1].src.ssa->bit_size != 64)
- continue;
- break;
- case nir_op_ieq:
- case nir_op_ine:
- case nir_op_ult:
- case nir_op_ilt:
- case nir_op_uge:
- case nir_op_ige:
- assert(alu->src[0].src.is_ssa);
- assert(alu->src[1].src.is_ssa);
- assert(alu->src[0].src.ssa->bit_size ==
- alu->src[1].src.ssa->bit_size);
- if (alu->src[0].src.ssa->bit_size != 64)
- continue;
- break;
- default:
- assert(alu->dest.dest.is_ssa);
- if (alu->dest.dest.ssa.bit_size != 64)
- continue;
- break;
- }
-
- if (!(options & nir_lower_int64_op_to_options_mask(alu->op)))
- continue;
+should_lower_int64_alu_instr(const nir_instr *instr, const void *_options)
+{
+ const nir_lower_int64_options options =
+ *(const nir_lower_int64_options *)_options;
- b.cursor = nir_before_instr(instr);
+ if (instr->type != nir_instr_type_alu)
+ return false;
- nir_ssa_def *lowered = lower_int64_alu_instr(&b, alu);
- nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa,
- nir_src_for_ssa(lowered));
- nir_instr_remove(&alu->instr);
- progress = true;
- }
- }
+ const nir_alu_instr *alu = nir_instr_as_alu(instr);
- if (progress) {
- nir_metadata_preserve(impl, nir_metadata_none);
- } else {
-#ifndef NDEBUG
- impl->valid_metadata &= ~nir_metadata_not_properly_reset;
-#endif
+ switch (alu->op) {
+ case nir_op_i2b1:
+ case nir_op_i2i32:
+ case nir_op_u2u32:
+ assert(alu->src[0].src.is_ssa);
+ if (alu->src[0].src.ssa->bit_size != 64)
+ return false;
+ break;
+ case nir_op_bcsel:
+ assert(alu->src[1].src.is_ssa);
+ assert(alu->src[2].src.is_ssa);
+ assert(alu->src[1].src.ssa->bit_size ==
+ alu->src[2].src.ssa->bit_size);
+ if (alu->src[1].src.ssa->bit_size != 64)
+ return false;
+ break;
+ case nir_op_ieq:
+ case nir_op_ine:
+ case nir_op_ult:
+ case nir_op_ilt:
+ case nir_op_uge:
+ case nir_op_ige:
+ assert(alu->src[0].src.is_ssa);
+ assert(alu->src[1].src.is_ssa);
+ assert(alu->src[0].src.ssa->bit_size ==
+ alu->src[1].src.ssa->bit_size);
+ if (alu->src[0].src.ssa->bit_size != 64)
+ return false;
+ break;
+ case nir_op_ufind_msb:
+ assert(alu->src[0].src.is_ssa);
+ if (alu->src[0].src.ssa->bit_size != 64)
+ return false;
+ break;
+ default:
+ assert(alu->dest.dest.is_ssa);
+ if (alu->dest.dest.ssa.bit_size != 64)
+ return false;
+ break;
}
- return progress;
+ return (options & nir_lower_int64_op_to_options_mask(alu->op)) != 0;
}
bool
nir_lower_int64(nir_shader *shader, nir_lower_int64_options options)
{
- bool progress = false;
-
- nir_foreach_function(function, shader) {
- if (function->impl)
- progress |= lower_int64_impl(function->impl, options);
- }
-
- return progress;
+ return nir_shader_lower_instructions(shader,
+ should_lower_int64_alu_instr,
+ lower_int64_alu_instr,
+ &options);
}
projects / mesa.git / blobdiff