nir_ixor(b, x_hi, y_hi));
}
+static nir_ssa_def *
+lower_ishl64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
+{
+ /* Implemented as
+ *
+ * uint64_t lshift(uint64_t x, int c)
+ * {
+ * if (c == 0) return x;
+ *
+ * uint32_t lo = LO(x), hi = HI(x);
+ *
+ * if (c < 32) {
+ * uint32_t lo_shifted = lo << c;
+ * uint32_t hi_shifted = hi << c;
+ * uint32_t lo_shifted_hi = lo >> abs(32 - c);
+ * return pack_64(lo_shifted, hi_shifted | lo_shifted_hi);
+ * } else {
+ * uint32_t lo_shifted_hi = lo << abs(32 - c);
+ * return pack_64(0, lo_shifted_hi);
+ * }
+ * }
+ */
+ 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 *reverse_count = nir_iabs(b, nir_iadd(b, y, nir_imm_int(b, -32)));
+ nir_ssa_def *lo_shifted = nir_ishl(b, x_lo, y);
+ nir_ssa_def *hi_shifted = nir_ishl(b, x_hi, y);
+ nir_ssa_def *lo_shifted_hi = nir_ushr(b, x_lo, reverse_count);
+
+ nir_ssa_def *res_if_lt_32 =
+ nir_pack_64_2x32_split(b, lo_shifted,
+ nir_ior(b, hi_shifted, lo_shifted_hi));
+ nir_ssa_def *res_if_ge_32 =
+ nir_pack_64_2x32_split(b, nir_imm_int(b, 0),
+ nir_ishl(b, x_lo, reverse_count));
+
+ return nir_bcsel(b,
+ nir_ieq(b, y, nir_imm_int(b, 0)), x,
+ nir_bcsel(b, nir_uge(b, y, nir_imm_int(b, 32)),
+ res_if_ge_32, res_if_lt_32));
+}
+
+static nir_ssa_def *
+lower_ishr64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
+{
+ /* Implemented as
+ *
+ * uint64_t arshift(uint64_t x, int c)
+ * {
+ * if (c == 0) return x;
+ *
+ * uint32_t lo = LO(x);
+ * int32_t hi = HI(x);
+ *
+ * if (c < 32) {
+ * uint32_t lo_shifted = lo >> c;
+ * uint32_t hi_shifted = hi >> c;
+ * uint32_t hi_shifted_lo = hi << abs(32 - c);
+ * return pack_64(hi_shifted, hi_shifted_lo | lo_shifted);
+ * } else {
+ * uint32_t hi_shifted = hi >> 31;
+ * uint32_t hi_shifted_lo = hi >> abs(32 - c);
+ * return pack_64(hi_shifted, hi_shifted_lo);
+ * }
+ * }
+ */
+ 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 *reverse_count = nir_iabs(b, nir_iadd(b, y, nir_imm_int(b, -32)));
+ nir_ssa_def *lo_shifted = nir_ushr(b, x_lo, y);
+ nir_ssa_def *hi_shifted = nir_ishr(b, x_hi, y);
+ nir_ssa_def *hi_shifted_lo = nir_ishl(b, x_hi, reverse_count);
+
+ nir_ssa_def *res_if_lt_32 =
+ nir_pack_64_2x32_split(b, nir_ior(b, lo_shifted, hi_shifted_lo),
+ hi_shifted);
+ nir_ssa_def *res_if_ge_32 =
+ nir_pack_64_2x32_split(b, nir_ishr(b, x_hi, reverse_count),
+ nir_ishr(b, x_hi, nir_imm_int(b, 31)));
+
+ return nir_bcsel(b,
+ nir_ieq(b, y, nir_imm_int(b, 0)), x,
+ nir_bcsel(b, nir_uge(b, y, nir_imm_int(b, 32)),
+ res_if_ge_32, res_if_lt_32));
+}
+
+static nir_ssa_def *
+lower_ushr64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
+{
+ /* Implemented as
+ *
+ * uint64_t rshift(uint64_t x, int c)
+ * {
+ * if (c == 0) return x;
+ *
+ * uint32_t lo = LO(x), hi = HI(x);
+ *
+ * if (c < 32) {
+ * uint32_t lo_shifted = lo >> c;
+ * uint32_t hi_shifted = hi >> c;
+ * uint32_t hi_shifted_lo = hi << abs(32 - c);
+ * return pack_64(hi_shifted, hi_shifted_lo | lo_shifted);
+ * } else {
+ * uint32_t hi_shifted_lo = hi >> abs(32 - c);
+ * return pack_64(0, hi_shifted_lo);
+ * }
+ * }
+ */
+
+ 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 *reverse_count = nir_iabs(b, nir_iadd(b, y, nir_imm_int(b, -32)));
+ nir_ssa_def *lo_shifted = nir_ushr(b, x_lo, y);
+ nir_ssa_def *hi_shifted = nir_ushr(b, x_hi, y);
+ nir_ssa_def *hi_shifted_lo = nir_ishl(b, x_hi, reverse_count);
+
+ nir_ssa_def *res_if_lt_32 =
+ nir_pack_64_2x32_split(b, nir_ior(b, lo_shifted, hi_shifted_lo),
+ hi_shifted);
+ nir_ssa_def *res_if_ge_32 =
+ nir_pack_64_2x32_split(b, nir_ushr(b, x_hi, reverse_count),
+ nir_imm_int(b, 0));
+
+ return nir_bcsel(b,
+ nir_ieq(b, y, nir_imm_int(b, 0)), x,
+ nir_bcsel(b, nir_uge(b, y, nir_imm_int(b, 32)),
+ res_if_ge_32, res_if_lt_32));
+}
+
static nir_ssa_def *
lower_iadd64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
{
return nir_bcsel(b, lower_int64_compare(b, nir_op_ilt, x, y), x, y);
}
+static nir_ssa_def *
+lower_mul_2x32_64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y,
+ bool sign_extend)
+{
+ nir_ssa_def *res_hi = sign_extend ? nir_imul_high(b, x, y)
+ : nir_umul_high(b, x, y);
+
+ return nir_pack_64_2x32_split(b, nir_imul(b, x, y), res_hi);
+}
+
static nir_ssa_def *
lower_imul64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
{
nir_ssa_def *y_lo = nir_unpack_64_2x32_split_x(b, y);
nir_ssa_def *y_hi = nir_unpack_64_2x32_split_y(b, y);
- nir_ssa_def *res_lo = nir_imul(b, x_lo, y_lo);
- nir_ssa_def *res_hi = nir_iadd(b, nir_umul_high(b, x_lo, y_lo),
+ nir_ssa_def *mul_lo = nir_umul_2x32_64(b, x_lo, y_lo);
+ nir_ssa_def *res_hi = nir_iadd(b, nir_unpack_64_2x32_split_y(b, mul_lo),
nir_iadd(b, nir_imul(b, x_lo, y_hi),
nir_imul(b, x_hi, y_lo)));
- return nir_pack_64_2x32_split(b, res_lo, res_hi);
+ return nir_pack_64_2x32_split(b, nir_unpack_64_2x32_split_x(b, mul_lo),
+ res_hi);
}
static nir_ssa_def *
* so we're guaranteed that we can add in two more 32-bit values
* without overflowing tmp.
*/
- nir_ssa_def *tmp =
- nir_pack_64_2x32_split(b, nir_imul(b, x32[i], y32[j]),
- nir_umul_high(b, x32[i], y32[j]));
+ nir_ssa_def *tmp = nir_umul_2x32_64(b, x32[i], y32[i]);
+
if (res[i + j])
tmp = nir_iadd(b, tmp, nir_u2u64(b, res[i + j]));
if (carry)
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_lower_int64_options
-opcode_to_options_mask(nir_op opcode)
+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);
+}
+
+nir_lower_int64_options
+nir_lower_int64_op_to_options_mask(nir_op opcode)
{
switch (opcode) {
case nir_op_imul:
return nir_lower_imul64;
+ case nir_op_imul_2x32_64:
+ case nir_op_umul_2x32_64:
+ return nir_lower_imul_2x32_64;
case nir_op_imul_high:
case nir_op_umul_high:
return nir_lower_imul_high64;
case nir_op_imod:
case nir_op_irem:
return nir_lower_divmod64;
+ case nir_op_b2i64:
+ case nir_op_i2b1:
+ case nir_op_i2i32:
+ case nir_op_i2i64:
+ case nir_op_u2u32:
+ case nir_op_u2u64:
+ case nir_op_bcsel:
+ return nir_lower_mov64;
+ case nir_op_ieq:
+ case nir_op_ine:
+ case nir_op_ult:
+ case nir_op_ilt:
+ case nir_op_uge:
+ case nir_op_ige:
+ return nir_lower_icmp64;
+ case nir_op_iadd:
+ case nir_op_isub:
+ return nir_lower_iadd64;
+ case nir_op_imin:
+ case nir_op_imax:
+ case nir_op_umin:
+ case nir_op_umax:
+ return nir_lower_minmax64;
+ case nir_op_iabs:
+ return nir_lower_iabs64;
+ case nir_op_ineg:
+ return nir_lower_ineg64;
+ case nir_op_iand:
+ case nir_op_ior:
+ case nir_op_ixor:
+ case nir_op_inot:
+ return nir_lower_logic64;
+ case nir_op_ishl:
+ 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;
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);
switch (alu->op) {
case nir_op_imul:
return lower_imul64(b, src[0], src[1]);
+ case nir_op_imul_2x32_64:
+ return lower_mul_2x32_64(b, src[0], src[1], true);
+ case nir_op_umul_2x32_64:
+ return lower_mul_2x32_64(b, src[0], src[1], false);
case nir_op_imul_high:
return lower_mul_high64(b, src[0], src[1], true);
case nir_op_umul_high:
return lower_imod64(b, src[0], src[1]);
case nir_op_irem:
return lower_irem64(b, src[0], src[1]);
+ case nir_op_b2i64:
+ return lower_b2i64(b, src[0]);
+ case nir_op_i2b1:
+ return lower_i2b(b, src[0]);
+ case nir_op_i2i8:
+ return lower_i2i8(b, src[0]);
+ case nir_op_i2i16:
+ return lower_i2i16(b, src[0]);
+ case nir_op_i2i32:
+ return lower_i2i32(b, src[0]);
+ case nir_op_i2i64:
+ return lower_i2i64(b, src[0]);
+ case nir_op_u2u8:
+ return lower_u2u8(b, src[0]);
+ case nir_op_u2u16:
+ return lower_u2u16(b, src[0]);
+ case nir_op_u2u32:
+ return lower_u2u32(b, src[0]);
+ case nir_op_u2u64:
+ return lower_u2u64(b, src[0]);
+ case nir_op_bcsel:
+ return lower_bcsel64(b, src[0], src[1], src[2]);
+ case nir_op_ieq:
+ case nir_op_ine:
+ case nir_op_ult:
+ case nir_op_ilt:
+ case nir_op_uge:
+ case nir_op_ige:
+ return lower_int64_compare(b, alu->op, src[0], src[1]);
+ case nir_op_iadd:
+ return lower_iadd64(b, src[0], src[1]);
+ case nir_op_isub:
+ return lower_isub64(b, src[0], src[1]);
+ case nir_op_imin:
+ return lower_imin64(b, src[0], src[1]);
+ case nir_op_imax:
+ return lower_imax64(b, src[0], src[1]);
+ case nir_op_umin:
+ return lower_umin64(b, src[0], src[1]);
+ case nir_op_umax:
+ return lower_umax64(b, src[0], src[1]);
+ case nir_op_iabs:
+ return lower_iabs64(b, src[0]);
+ case nir_op_ineg:
+ return lower_ineg64(b, src[0]);
+ case nir_op_iand:
+ return lower_iand64(b, src[0], src[1]);
+ case nir_op_ior:
+ return lower_ior64(b, src[0], src[1]);
+ case nir_op_ixor:
+ return lower_ixor64(b, src[0], src[1]);
+ case nir_op_inot:
+ return lower_inot64(b, src[0]);
+ case nir_op_ishl:
+ return lower_ishl64(b, src[0], src[1]);
+ case nir_op_ishr:
+ 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]);
default:
unreachable("Invalid ALU opcode to lower");
}
}
static bool
-lower_int64_impl(nir_function_impl *impl, nir_lower_int64_options options)
+should_lower_int64_alu_instr(const nir_instr *instr, const void *_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;
+ const nir_lower_int64_options options =
+ *(const nir_lower_int64_options *)_options;
- nir_alu_instr *alu = nir_instr_as_alu(instr);
- assert(alu->dest.dest.is_ssa);
- if (alu->dest.dest.ssa.bit_size != 64)
- continue;
+ if (instr->type != nir_instr_type_alu)
+ return false;
- if (!(options & opcode_to_options_mask(alu->op)))
- continue;
+ const nir_alu_instr *alu = nir_instr_as_alu(instr);
- b.cursor = nir_before_instr(instr);
-
- 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;
- }
+ 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;
+ default:
+ assert(alu->dest.dest.is_ssa);
+ if (alu->dest.dest.ssa.bit_size != 64)
+ return false;
+ break;
}
- if (progress)
- nir_metadata_preserve(impl, nir_metadata_none);
-
- 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