when getting an element pointer, use the best alignment

type we can figure out is safe to use

See #37
master
Andrew Kelley 2017-08-30 02:25:41 -04:00
parent 56cc62a9c7
commit 5482f9f9e8
2 changed files with 89 additions and 17 deletions

View File

@ -8404,7 +8404,7 @@ static IrInstruction *ir_analyze_cast(IrAnalyze *ira, IrInstruction *source_inst
if (!ir_emit_global_runtime_side_effect(ira, source_instr))
return ira->codegen->invalid_instruction;
uint64_t child_type_size = type_size(ira->codegen,
wanted_type->data.structure.fields[0].type_entry->data.pointer.child_type);
wanted_type->data.structure.fields[slice_ptr_index].type_entry->data.pointer.child_type);
if (actual_type->data.array.len % child_type_size == 0) {
return ir_resolve_cast(ira, source_instr, value, wanted_type, CastOpBytesToSlice, true);
} else {
@ -10870,6 +10870,15 @@ static TypeTableEntry *ir_analyze_instruction_var_ptr(IrAnalyze *ira, IrInstruct
var_ptr_instruction->is_volatile);
}
static TypeTableEntry *adjust_ptr_align(CodeGen *g, TypeTableEntry *ptr_type, uint32_t new_align) {
assert(ptr_type->id == TypeTableEntryIdPointer);
return get_pointer_to_type_extra(g,
ptr_type->data.pointer.child_type,
ptr_type->data.pointer.is_const, ptr_type->data.pointer.is_volatile,
new_align,
ptr_type->data.pointer.bit_offset, ptr_type->data.pointer.unaligned_bit_count);
}
static TypeTableEntry *ir_analyze_instruction_elem_ptr(IrAnalyze *ira, IrInstructionElemPtr *elem_ptr_instruction) {
IrInstruction *array_ptr = elem_ptr_instruction->array_ptr->other;
if (type_is_invalid(array_ptr->value.type))
@ -10888,6 +10897,9 @@ static TypeTableEntry *ir_analyze_instruction_elem_ptr(IrAnalyze *ira, IrInstruc
assert(ptr_type->id == TypeTableEntryIdPointer);
TypeTableEntry *array_type = ptr_type->data.pointer.child_type;
// At first return_type will be the pointer type we want to return, except with an optimistic alignment.
// We will adjust return_type's alignment before returning it.
TypeTableEntry *return_type;
if (type_is_invalid(array_type)) {
@ -10918,7 +10930,7 @@ static TypeTableEntry *ir_analyze_instruction_elem_ptr(IrAnalyze *ira, IrInstruc
} else if (array_type->id == TypeTableEntryIdPointer) {
return_type = array_type;
} else if (is_slice(array_type)) {
return_type = array_type->data.structure.fields[0].type_entry;
return_type = array_type->data.structure.fields[slice_ptr_index].type_entry;
} else if (array_type->id == TypeTableEntryIdArgTuple) {
ConstExprValue *ptr_val = ir_resolve_const(ira, array_ptr, UndefBad);
if (!ptr_val)
@ -10961,6 +10973,10 @@ static TypeTableEntry *ir_analyze_instruction_elem_ptr(IrAnalyze *ira, IrInstruc
return ira->codegen->builtin_types.entry_invalid;
bool safety_check_on = elem_ptr_instruction->safety_check_on;
ensure_complete_type(ira->codegen, return_type->data.pointer.child_type);
uint64_t elem_size = type_size(ira->codegen, return_type->data.pointer.child_type);
uint64_t abi_align = get_abi_alignment(ira->codegen, return_type->data.pointer.child_type);
uint64_t ptr_align = return_type->data.pointer.alignment;
if (instr_is_comptime(casted_elem_index)) {
uint64_t index = bigint_as_unsigned(&casted_elem_index->value.data.x_bigint);
if (array_type->id == TypeTableEntryIdArray) {
@ -10974,6 +10990,26 @@ static TypeTableEntry *ir_analyze_instruction_elem_ptr(IrAnalyze *ira, IrInstruc
safety_check_on = false;
}
{
// figure out the largest alignment possible
uint64_t chosen_align = abi_align;
if (ptr_align >= abi_align) {
while (ptr_align > abi_align) {
if ((index * elem_size) % ptr_align == 0) {
chosen_align = ptr_align;
break;
}
ptr_align >>= 1;
}
} else if (elem_size >= ptr_align && elem_size % ptr_align == 0) {
chosen_align = ptr_align;
} else {
// can't get here because guaranteed elem_size >= abi_align
zig_unreachable();
}
return_type = adjust_ptr_align(ira->codegen, return_type, chosen_align);
}
ConstExprValue *array_ptr_val;
if (array_ptr->value.special != ConstValSpecialRuntime &&
(array_ptr->value.data.x_ptr.mut != ConstPtrMutRuntimeVar || array_type->id == TypeTableEntryIdArray) &&
@ -11085,6 +11121,18 @@ static TypeTableEntry *ir_analyze_instruction_elem_ptr(IrAnalyze *ira, IrInstruc
}
}
} else {
// runtime known element index
if (ptr_align < abi_align) {
if (elem_size >= ptr_align && elem_size % ptr_align == 0) {
return_type = adjust_ptr_align(ira->codegen, return_type, ptr_align);
} else {
// can't get here because guaranteed elem_size >= abi_align
zig_unreachable();
}
} else {
return_type = adjust_ptr_align(ira->codegen, return_type, abi_align);
}
}
ir_build_elem_ptr_from(&ira->new_irb, &elem_ptr_instruction->base, array_ptr,
@ -14520,11 +14568,7 @@ static IrInstruction *ir_align_cast(IrAnalyze *ira, IrInstruction *target, uint3
uint32_t old_align_bytes;
if (target_type->id == TypeTableEntryIdPointer) {
result_type = get_pointer_to_type_extra(ira->codegen,
target_type->data.pointer.child_type,
target_type->data.pointer.is_const, target_type->data.pointer.is_volatile,
align_bytes,
target_type->data.pointer.bit_offset, target_type->data.pointer.unaligned_bit_count);
result_type = adjust_ptr_align(ira->codegen, target_type, align_bytes);
} else if (target_type->id == TypeTableEntryIdFn) {
FnTypeId fn_type_id = target_type->data.fn.fn_type_id;
old_align_bytes = fn_type_id.alignment;
@ -14535,11 +14579,7 @@ static IrInstruction *ir_align_cast(IrAnalyze *ira, IrInstruction *target, uint3
{
TypeTableEntry *ptr_type = target_type->data.maybe.child_type;
old_align_bytes = ptr_type->data.pointer.alignment;
TypeTableEntry *better_ptr_type = get_pointer_to_type_extra(ira->codegen,
ptr_type->data.pointer.child_type,
ptr_type->data.pointer.is_const, ptr_type->data.pointer.is_volatile,
align_bytes,
ptr_type->data.pointer.bit_offset, ptr_type->data.pointer.unaligned_bit_count);
TypeTableEntry *better_ptr_type = adjust_ptr_align(ira->codegen, ptr_type, align_bytes);
result_type = get_maybe_type(ira->codegen, better_ptr_type);
} else if (target_type->id == TypeTableEntryIdMaybe &&
@ -14553,11 +14593,7 @@ static IrInstruction *ir_align_cast(IrAnalyze *ira, IrInstruction *target, uint3
} else if (is_slice(target_type)) {
TypeTableEntry *slice_ptr_type = target_type->data.structure.fields[slice_ptr_index].type_entry;
old_align_bytes = slice_ptr_type->data.pointer.alignment;
TypeTableEntry *result_ptr_type = get_pointer_to_type_extra(ira->codegen,
slice_ptr_type->data.pointer.child_type,
slice_ptr_type->data.pointer.is_const, slice_ptr_type->data.pointer.is_volatile,
align_bytes,
slice_ptr_type->data.pointer.bit_offset, slice_ptr_type->data.pointer.unaligned_bit_count);
TypeTableEntry *result_ptr_type = adjust_ptr_align(ira->codegen, slice_ptr_type, align_bytes);
result_type = get_slice_type(ira->codegen, result_ptr_type);
} else {
ir_add_error(ira, target,

View File

@ -143,3 +143,39 @@ test "@ptrCast preserves alignment of bigger source" {
const ptr = @ptrCast(&u8, &x);
assert(@typeOf(ptr) == &align 16 u8);
}
test "compile-time known array index has best alignment possible" {
// take full advantage of over-alignment
var array align 4 = []u8 {1, 2, 3, 4};
assert(@typeOf(&array[0]) == &align 4 u8);
assert(@typeOf(&array[1]) == &u8);
assert(@typeOf(&array[2]) == &align 2 u8);
assert(@typeOf(&array[3]) == &u8);
// because align is too small but we still figure out to use 2
var bigger align 2 = []u64{1, 2, 3, 4};
assert(@typeOf(&bigger[0]) == &align 2 u64);
assert(@typeOf(&bigger[1]) == &align 2 u64);
assert(@typeOf(&bigger[2]) == &align 2 u64);
assert(@typeOf(&bigger[3]) == &align 2 u64);
// because pointer is align 2 and u32 align % 2 == 0 we can assume align 2
var smaller align 2 = []u32{1, 2, 3, 4};
testIndex(&smaller[0], 0, &align 2 u32);
testIndex(&smaller[0], 1, &align 2 u32);
testIndex(&smaller[0], 2, &align 2 u32);
testIndex(&smaller[0], 3, &align 2 u32);
// has to use ABI alignment because index known at runtime only
testIndex2(&array[0], 0, &u8);
testIndex2(&array[0], 1, &u8);
testIndex2(&array[0], 2, &u8);
testIndex2(&array[0], 3, &u8);
}
fn testIndex(smaller: &align 2 u32, index: usize, comptime T: type) {
assert(@typeOf(&smaller[index]) == T);
}
fn testIndex2(ptr: &align 4 u8, index: usize, comptime T: type) {
assert(@typeOf(&ptr[index]) == T);
}