ctz + clz

This commit is contained in:
Auguste Rame 2021-07-25 22:43:52 -04:00
parent ecca829bcb
commit c619b85f67
4 changed files with 148 additions and 24 deletions

View File

@ -1907,9 +1907,11 @@ struct ZigLLVMFnKey {
union {
struct {
uint32_t bit_count;
uint32_t vector_len; // 0 means not a vector
} ctz;
struct {
uint32_t bit_count;
uint32_t vector_len; // 0 means not a vector
} clz;
struct {
uint32_t bit_count;

View File

@ -7883,9 +7883,11 @@ bool type_id_eql(TypeId const *a, TypeId const *b) {
uint32_t zig_llvm_fn_key_hash(ZigLLVMFnKey const *x) {
switch (x->id) {
case ZigLLVMFnIdCtz:
return (uint32_t)(x->data.ctz.bit_count) * (uint32_t)810453934;
return (uint32_t)(x->data.ctz.bit_count) * (uint32_t)810453934 +
(uint32_t)(x->data.ctz.vector_len) * (((uint32_t)x->id << 5) + 1025);
case ZigLLVMFnIdClz:
return (uint32_t)(x->data.clz.bit_count) * (uint32_t)2428952817;
return (uint32_t)(x->data.clz.bit_count) * (uint32_t)2428952817 +
(uint32_t)(x->data.clz.vector_len) * (((uint32_t)x->id << 5) + 1025);
case ZigLLVMFnIdPopCount:
return (uint32_t)(x->data.pop_count.bit_count) * (uint32_t)101195049 +
(uint32_t)(x->data.pop_count.vector_len) * (((uint32_t)x->id << 5) + 1025);

View File

@ -15945,54 +15945,154 @@ static Stage1AirInst *ir_analyze_instruction_optional_unwrap_ptr(IrAnalyze *ira,
}
static Stage1AirInst *ir_analyze_instruction_ctz(IrAnalyze *ira, Stage1ZirInstCtz *instruction) {
Error err;
ZigType *int_type = ir_resolve_int_type(ira, instruction->type->child);
if (type_is_invalid(int_type))
return ira->codegen->invalid_inst_gen;
Stage1AirInst *op = ir_implicit_cast(ira, instruction->op->child, int_type);
Stage1AirInst *uncasted_op = instruction->op->child;
if (type_is_invalid(uncasted_op->value->type))
return ira->codegen->invalid_inst_gen;
uint32_t vector_len = UINT32_MAX; // means not a vector
if (uncasted_op->value->type->id == ZigTypeIdArray) {
bool can_be_vec_elem;
if ((err = is_valid_vector_elem_type(ira->codegen, uncasted_op->value->type->data.array.child_type,
&can_be_vec_elem)))
{
return ira->codegen->invalid_inst_gen;
}
if (can_be_vec_elem) {
vector_len = uncasted_op->value->type->data.array.len;
}
} else if (uncasted_op->value->type->id == ZigTypeIdVector) {
vector_len = uncasted_op->value->type->data.vector.len;
}
bool is_vector = (vector_len != UINT32_MAX);
ZigType *op_type = is_vector ? get_vector_type(ira->codegen, vector_len, int_type) : int_type;
Stage1AirInst *op = ir_implicit_cast(ira, uncasted_op, op_type);
if (type_is_invalid(op->value->type))
return ira->codegen->invalid_inst_gen;
if (int_type->data.integral.bit_count == 0)
return ir_const_unsigned(ira, instruction->base.scope, instruction->base.source_node, 0);
ZigType *smallest_type = get_smallest_unsigned_int_type(ira->codegen, int_type->data.integral.bit_count);
if (instr_is_comptime(op)) {
ZigValue *val = ir_resolve_const(ira, op, UndefOk);
if (val == nullptr)
return ira->codegen->invalid_inst_gen;
if (val->special == ConstValSpecialUndef)
return ir_const_undef(ira, instruction->base.scope, instruction->base.source_node, ira->codegen->builtin_types.entry_num_lit_int);
size_t result_usize = bigint_ctz(&op->value->data.x_bigint, int_type->data.integral.bit_count);
return ir_const_unsigned(ira, instruction->base.scope, instruction->base.source_node, result_usize);
if (is_vector) {
ZigType *smallest_vec_type = get_vector_type(ira->codegen, vector_len, smallest_type);
Stage1AirInst *result = ir_const(ira, instruction->base.scope, instruction->base.source_node, smallest_vec_type);
expand_undef_array(ira->codegen, val);
result->value->data.x_array.data.s_none.elements = ira->codegen->pass1_arena->allocate<ZigValue>(smallest_vec_type->data.vector.len);
for (unsigned i = 0; i < smallest_vec_type->data.vector.len; i += 1) {
ZigValue *op_elem_val = &val->data.x_array.data.s_none.elements[i];
if ((err = ir_resolve_const_val(ira->codegen, ira->new_irb.exec, instruction->base.source_node,
op_elem_val, UndefOk)))
{
return ira->codegen->invalid_inst_gen;
}
ZigValue *result_elem_val = &result->value->data.x_array.data.s_none.elements[i];
result_elem_val->type = smallest_type;
result_elem_val->special = op_elem_val->special;
if (op_elem_val->special == ConstValSpecialUndef)
continue;
size_t value = bigint_ctz(&op_elem_val->data.x_bigint, int_type->data.integral.bit_count);
bigint_init_unsigned(&result->value->data.x_array.data.s_none.elements[i].data.x_bigint, value);
}
return result;
} else {
size_t result_usize = bigint_ctz(&op->value->data.x_bigint, int_type->data.integral.bit_count);
return ir_const_unsigned(ira, instruction->base.scope, instruction->base.source_node, result_usize);
}
}
ZigType *return_type = get_smallest_unsigned_int_type(ira->codegen, int_type->data.integral.bit_count);
ZigType *return_type = is_vector ? get_vector_type(ira->codegen, vector_len, smallest_type) : smallest_type;
return ir_build_ctz_gen(ira, instruction->base.scope, instruction->base.source_node, return_type, op);
}
static Stage1AirInst *ir_analyze_instruction_clz(IrAnalyze *ira, Stage1ZirInstClz *instruction) {
Error err;
ZigType *int_type = ir_resolve_int_type(ira, instruction->type->child);
if (type_is_invalid(int_type))
return ira->codegen->invalid_inst_gen;
Stage1AirInst *op = ir_implicit_cast(ira, instruction->op->child, int_type);
Stage1AirInst *uncasted_op = instruction->op->child;
if (type_is_invalid(uncasted_op->value->type))
return ira->codegen->invalid_inst_gen;
uint32_t vector_len = UINT32_MAX; // means not a vector
if (uncasted_op->value->type->id == ZigTypeIdArray) {
bool can_be_vec_elem;
if ((err = is_valid_vector_elem_type(ira->codegen, uncasted_op->value->type->data.array.child_type,
&can_be_vec_elem)))
{
return ira->codegen->invalid_inst_gen;
}
if (can_be_vec_elem) {
vector_len = uncasted_op->value->type->data.array.len;
}
} else if (uncasted_op->value->type->id == ZigTypeIdVector) {
vector_len = uncasted_op->value->type->data.vector.len;
}
bool is_vector = (vector_len != UINT32_MAX);
ZigType *op_type = is_vector ? get_vector_type(ira->codegen, vector_len, int_type) : int_type;
Stage1AirInst *op = ir_implicit_cast(ira, uncasted_op, op_type);
if (type_is_invalid(op->value->type))
return ira->codegen->invalid_inst_gen;
if (int_type->data.integral.bit_count == 0)
return ir_const_unsigned(ira, instruction->base.scope, instruction->base.source_node, 0);
ZigType *smallest_type = get_smallest_unsigned_int_type(ira->codegen, int_type->data.integral.bit_count);
if (instr_is_comptime(op)) {
ZigValue *val = ir_resolve_const(ira, op, UndefOk);
if (val == nullptr)
return ira->codegen->invalid_inst_gen;
if (val->special == ConstValSpecialUndef)
return ir_const_undef(ira, instruction->base.scope, instruction->base.source_node, ira->codegen->builtin_types.entry_num_lit_int);
size_t result_usize = bigint_clz(&op->value->data.x_bigint, int_type->data.integral.bit_count);
return ir_const_unsigned(ira, instruction->base.scope, instruction->base.source_node, result_usize);
if (is_vector) {
ZigType *smallest_vec_type = get_vector_type(ira->codegen, vector_len, smallest_type);
Stage1AirInst *result = ir_const(ira, instruction->base.scope, instruction->base.source_node, smallest_vec_type);
expand_undef_array(ira->codegen, val);
result->value->data.x_array.data.s_none.elements = ira->codegen->pass1_arena->allocate<ZigValue>(smallest_vec_type->data.vector.len);
for (unsigned i = 0; i < smallest_vec_type->data.vector.len; i += 1) {
ZigValue *op_elem_val = &val->data.x_array.data.s_none.elements[i];
if ((err = ir_resolve_const_val(ira->codegen, ira->new_irb.exec, instruction->base.source_node,
op_elem_val, UndefOk)))
{
return ira->codegen->invalid_inst_gen;
}
ZigValue *result_elem_val = &result->value->data.x_array.data.s_none.elements[i];
result_elem_val->type = smallest_type;
result_elem_val->special = op_elem_val->special;
if (op_elem_val->special == ConstValSpecialUndef)
continue;
size_t value = bigint_clz(&op_elem_val->data.x_bigint, int_type->data.integral.bit_count);
bigint_init_unsigned(&result->value->data.x_array.data.s_none.elements[i].data.x_bigint, value);
}
return result;
} else {
size_t result_usize = bigint_clz(&op->value->data.x_bigint, int_type->data.integral.bit_count);
return ir_const_unsigned(ira, instruction->base.scope, instruction->base.source_node, result_usize);
}
}
ZigType *return_type = get_smallest_unsigned_int_type(ira->codegen, int_type->data.integral.bit_count);
ZigType *return_type = is_vector ? get_vector_type(ira->codegen, vector_len, smallest_type) : smallest_type;
return ir_build_clz_gen(ira, instruction->base.scope, instruction->base.source_node, return_type, op);
}

View File

@ -123,16 +123,27 @@ test "@clz" {
}
fn testClz() !void {
try expect(clz(u8, 0b10001010) == 0);
try expect(clz(u8, 0b00001010) == 4);
try expect(clz(u8, 0b00011010) == 3);
try expect(clz(u8, 0b00000000) == 8);
try expect(clz(u128, 0xffffffffffffffff) == 64);
try expect(clz(u128, 0x10000000000000000) == 63);
try expect(@clz(u8, 0b10001010) == 0);
try expect(@clz(u8, 0b00001010) == 4);
try expect(@clz(u8, 0b00011010) == 3);
try expect(@clz(u8, 0b00000000) == 8);
try expect(@clz(u128, 0xffffffffffffffff) == 64);
try expect(@clz(u128, 0x10000000000000000) == 63);
}
fn clz(comptime T: type, x: T) usize {
return @clz(T, x);
test "@clz vectors" {
try testClzVectors();
comptime try testClzVectors();
}
fn testClzVectors() !void {
@setEvalBranchQuota(10_000);
try expectEqual(@clz(u8, @splat(64, @as(u8, 0b10001010))), @splat(64, @as(u4, 0)));
try expectEqual(@clz(u8, @splat(64, @as(u8, 0b00001010))), @splat(64, @as(u4, 4)));
try expectEqual(@clz(u8, @splat(64, @as(u8, 0b00011010))), @splat(64, @as(u4, 3)));
try expectEqual(@clz(u8, @splat(64, @as(u8, 0b00000000))), @splat(64, @as(u4, 8)));
try expectEqual(@clz(u128, @splat(64, @as(u128, 0xffffffffffffffff))), @splat(64, @as(u8, 64)));
try expectEqual(@clz(u128, @splat(64, @as(u128, 0x10000000000000000))), @splat(64, @as(u8, 63)));
}
test "@ctz" {
@ -141,14 +152,23 @@ test "@ctz" {
}
fn testCtz() !void {
try expect(ctz(u8, 0b10100000) == 5);
try expect(ctz(u8, 0b10001010) == 1);
try expect(ctz(u8, 0b00000000) == 8);
try expect(ctz(u16, 0b00000000) == 16);
try expect(@ctz(u8, 0b10100000) == 5);
try expect(@ctz(u8, 0b10001010) == 1);
try expect(@ctz(u8, 0b00000000) == 8);
try expect(@ctz(u16, 0b00000000) == 16);
}
fn ctz(comptime T: type, x: T) usize {
return @ctz(T, x);
test "@ctz vectors" {
try testClzVectors();
comptime try testClzVectors();
}
fn testCtzVectors() !void {
@setEvalBranchQuota(10_000);
try expectEqual(@ctz(u8, @splat(64, @as(u8, 0b10100000))), @splat(64, @as(u4, 5)));
try expectEqual(@ctz(u8, @splat(64, @as(u8, 0b10001010))), @splat(64, @as(u4, 1)));
try expectEqual(@ctz(u8, @splat(64, @as(u8, 0b00000000))), @splat(64, @as(u4, 8)));
try expectEqual(@ctz(u16, @splat(64, @as(u16, 0b00000000))), @splat(64, @as(u5, 16)));
}
test "assignment operators" {