1 #![cfg_attr(feature = "nightly", feature(step_trait, rustc_attrs, min_specialization))]
4 #[cfg(feature = "nightly")]
6 use std::num::{NonZeroUsize, ParseIntError};
7 use std::ops::{Add, AddAssign, Mul, RangeInclusive, Sub};
10 use bitflags::bitflags;
11 #[cfg(feature = "nightly")]
12 use rustc_data_structures::stable_hasher::StableOrd;
13 use rustc_index::vec::{Idx, IndexVec};
14 #[cfg(feature = "nightly")]
15 use rustc_macros::HashStable_Generic;
16 #[cfg(feature = "nightly")]
17 use rustc_macros::{Decodable, Encodable};
21 pub use layout::LayoutCalculator;
23 /// Requirements for a `StableHashingContext` to be used in this crate.
24 /// This is a hack to allow using the `HashStable_Generic` derive macro
25 /// instead of implementing everything in `rustc_middle`.
26 pub trait HashStableContext {}
33 #[cfg_attr(feature = "nightly", derive(Encodable, Decodable, HashStable_Generic))]
34 pub struct ReprFlags: u8 {
36 const IS_SIMD = 1 << 1;
37 const IS_TRANSPARENT = 1 << 2;
38 // Internal only for now. If true, don't reorder fields.
39 const IS_LINEAR = 1 << 3;
40 // If true, the type's layout can be randomized using
41 // the seed stored in `ReprOptions.layout_seed`
42 const RANDOMIZE_LAYOUT = 1 << 4;
43 // Any of these flags being set prevent field reordering optimisation.
44 const IS_UNOPTIMISABLE = ReprFlags::IS_C.bits
45 | ReprFlags::IS_SIMD.bits
46 | ReprFlags::IS_LINEAR.bits;
50 #[derive(Copy, Clone, Debug, Eq, PartialEq)]
51 #[cfg_attr(feature = "nightly", derive(Encodable, Decodable, HashStable_Generic))]
52 pub enum IntegerType {
53 /// Pointer sized integer type, i.e. isize and usize. The field shows signedness, that
54 /// is, `Pointer(true)` is isize.
56 /// Fix sized integer type, e.g. i8, u32, i128 The bool field shows signedness, `Fixed(I8, false)` means `u8`
61 pub fn is_signed(&self) -> bool {
63 IntegerType::Pointer(b) => *b,
64 IntegerType::Fixed(_, b) => *b,
69 /// Represents the repr options provided by the user,
70 #[derive(Copy, Clone, Debug, Eq, PartialEq, Default)]
71 #[cfg_attr(feature = "nightly", derive(Encodable, Decodable, HashStable_Generic))]
72 pub struct ReprOptions {
73 pub int: Option<IntegerType>,
74 pub align: Option<Align>,
75 pub pack: Option<Align>,
77 /// The seed to be used for randomizing a type's layout
79 /// Note: This could technically be a `[u8; 16]` (a `u128`) which would
80 /// be the "most accurate" hash as it'd encompass the item and crate
81 /// hash without loss, but it does pay the price of being larger.
82 /// Everything's a tradeoff, a `u64` seed should be sufficient for our
83 /// purposes (primarily `-Z randomize-layout`)
84 pub field_shuffle_seed: u64,
89 pub fn simd(&self) -> bool {
90 self.flags.contains(ReprFlags::IS_SIMD)
94 pub fn c(&self) -> bool {
95 self.flags.contains(ReprFlags::IS_C)
99 pub fn packed(&self) -> bool {
104 pub fn transparent(&self) -> bool {
105 self.flags.contains(ReprFlags::IS_TRANSPARENT)
109 pub fn linear(&self) -> bool {
110 self.flags.contains(ReprFlags::IS_LINEAR)
113 /// Returns the discriminant type, given these `repr` options.
114 /// This must only be called on enums!
115 pub fn discr_type(&self) -> IntegerType {
116 self.int.unwrap_or(IntegerType::Pointer(true))
119 /// Returns `true` if this `#[repr()]` should inhabit "smart enum
120 /// layout" optimizations, such as representing `Foo<&T>` as a
122 pub fn inhibit_enum_layout_opt(&self) -> bool {
123 self.c() || self.int.is_some()
126 /// Returns `true` if this `#[repr()]` should inhibit struct field reordering
127 /// optimizations, such as with `repr(C)`, `repr(packed(1))`, or `repr(<int>)`.
128 pub fn inhibit_struct_field_reordering_opt(&self) -> bool {
129 if let Some(pack) = self.pack {
130 if pack.bytes() == 1 {
135 self.flags.intersects(ReprFlags::IS_UNOPTIMISABLE) || self.int.is_some()
138 /// Returns `true` if this type is valid for reordering and `-Z randomize-layout`
139 /// was enabled for its declaration crate
140 pub fn can_randomize_type_layout(&self) -> bool {
141 !self.inhibit_struct_field_reordering_opt()
142 && self.flags.contains(ReprFlags::RANDOMIZE_LAYOUT)
145 /// Returns `true` if this `#[repr()]` should inhibit union ABI optimisations.
146 pub fn inhibit_union_abi_opt(&self) -> bool {
151 /// Parsed [Data layout](https://llvm.org/docs/LangRef.html#data-layout)
152 /// for a target, which contains everything needed to compute layouts.
153 #[derive(Debug, PartialEq, Eq)]
154 pub struct TargetDataLayout {
156 pub i1_align: AbiAndPrefAlign,
157 pub i8_align: AbiAndPrefAlign,
158 pub i16_align: AbiAndPrefAlign,
159 pub i32_align: AbiAndPrefAlign,
160 pub i64_align: AbiAndPrefAlign,
161 pub i128_align: AbiAndPrefAlign,
162 pub f32_align: AbiAndPrefAlign,
163 pub f64_align: AbiAndPrefAlign,
164 pub pointer_size: Size,
165 pub pointer_align: AbiAndPrefAlign,
166 pub aggregate_align: AbiAndPrefAlign,
168 /// Alignments for vector types.
169 pub vector_align: Vec<(Size, AbiAndPrefAlign)>,
171 pub instruction_address_space: AddressSpace,
173 /// Minimum size of #[repr(C)] enums (default I32 bits)
174 pub c_enum_min_size: Integer,
177 impl Default for TargetDataLayout {
178 /// Creates an instance of `TargetDataLayout`.
179 fn default() -> TargetDataLayout {
180 let align = |bits| Align::from_bits(bits).unwrap();
183 i1_align: AbiAndPrefAlign::new(align(8)),
184 i8_align: AbiAndPrefAlign::new(align(8)),
185 i16_align: AbiAndPrefAlign::new(align(16)),
186 i32_align: AbiAndPrefAlign::new(align(32)),
187 i64_align: AbiAndPrefAlign { abi: align(32), pref: align(64) },
188 i128_align: AbiAndPrefAlign { abi: align(32), pref: align(64) },
189 f32_align: AbiAndPrefAlign::new(align(32)),
190 f64_align: AbiAndPrefAlign::new(align(64)),
191 pointer_size: Size::from_bits(64),
192 pointer_align: AbiAndPrefAlign::new(align(64)),
193 aggregate_align: AbiAndPrefAlign { abi: align(0), pref: align(64) },
195 (Size::from_bits(64), AbiAndPrefAlign::new(align(64))),
196 (Size::from_bits(128), AbiAndPrefAlign::new(align(128))),
198 instruction_address_space: AddressSpace::DATA,
199 c_enum_min_size: Integer::I32,
204 pub enum TargetDataLayoutErrors<'a> {
205 InvalidAddressSpace { addr_space: &'a str, cause: &'a str, err: ParseIntError },
206 InvalidBits { kind: &'a str, bit: &'a str, cause: &'a str, err: ParseIntError },
207 MissingAlignment { cause: &'a str },
208 InvalidAlignment { cause: &'a str, err: String },
209 InconsistentTargetArchitecture { dl: &'a str, target: &'a str },
210 InconsistentTargetPointerWidth { pointer_size: u64, target: u32 },
211 InvalidBitsSize { err: String },
214 impl TargetDataLayout {
215 /// Parse data layout from an [llvm data layout string](https://llvm.org/docs/LangRef.html#data-layout)
217 /// This function doesn't fill `c_enum_min_size` and it will always be `I32` since it can not be
218 /// determined from llvm string.
219 pub fn parse_from_llvm_datalayout_string<'a>(
221 ) -> Result<TargetDataLayout, TargetDataLayoutErrors<'a>> {
222 // Parse an address space index from a string.
223 let parse_address_space = |s: &'a str, cause: &'a str| {
224 s.parse::<u32>().map(AddressSpace).map_err(|err| {
225 TargetDataLayoutErrors::InvalidAddressSpace { addr_space: s, cause, err }
229 // Parse a bit count from a string.
230 let parse_bits = |s: &'a str, kind: &'a str, cause: &'a str| {
231 s.parse::<u64>().map_err(|err| TargetDataLayoutErrors::InvalidBits {
239 // Parse a size string.
240 let size = |s: &'a str, cause: &'a str| parse_bits(s, "size", cause).map(Size::from_bits);
242 // Parse an alignment string.
243 let align = |s: &[&'a str], cause: &'a str| {
245 return Err(TargetDataLayoutErrors::MissingAlignment { cause });
247 let align_from_bits = |bits| {
248 Align::from_bits(bits)
249 .map_err(|err| TargetDataLayoutErrors::InvalidAlignment { cause, err })
251 let abi = parse_bits(s[0], "alignment", cause)?;
252 let pref = s.get(1).map_or(Ok(abi), |pref| parse_bits(pref, "alignment", cause))?;
253 Ok(AbiAndPrefAlign { abi: align_from_bits(abi)?, pref: align_from_bits(pref)? })
256 let mut dl = TargetDataLayout::default();
257 let mut i128_align_src = 64;
258 for spec in input.split('-') {
259 let spec_parts = spec.split(':').collect::<Vec<_>>();
262 ["e"] => dl.endian = Endian::Little,
263 ["E"] => dl.endian = Endian::Big,
264 [p] if p.starts_with('P') => {
265 dl.instruction_address_space = parse_address_space(&p[1..], "P")?
267 ["a", ref a @ ..] => dl.aggregate_align = align(a, "a")?,
268 ["f32", ref a @ ..] => dl.f32_align = align(a, "f32")?,
269 ["f64", ref a @ ..] => dl.f64_align = align(a, "f64")?,
270 // FIXME(erikdesjardins): we should be parsing nonzero address spaces
271 // this will require replacing TargetDataLayout::{pointer_size,pointer_align}
272 // with e.g. `fn pointer_size_in(AddressSpace)`
273 [p @ "p", s, ref a @ ..] | [p @ "p0", s, ref a @ ..] => {
274 dl.pointer_size = size(s, p)?;
275 dl.pointer_align = align(a, p)?;
277 [s, ref a @ ..] if s.starts_with('i') => {
278 let Ok(bits) = s[1..].parse::<u64>() else {
279 size(&s[1..], "i")?; // For the user error.
282 let a = align(a, s)?;
284 1 => dl.i1_align = a,
285 8 => dl.i8_align = a,
286 16 => dl.i16_align = a,
287 32 => dl.i32_align = a,
288 64 => dl.i64_align = a,
291 if bits >= i128_align_src && bits <= 128 {
292 // Default alignment for i128 is decided by taking the alignment of
293 // largest-sized i{64..=128}.
294 i128_align_src = bits;
298 [s, ref a @ ..] if s.starts_with('v') => {
299 let v_size = size(&s[1..], "v")?;
300 let a = align(a, s)?;
301 if let Some(v) = dl.vector_align.iter_mut().find(|v| v.0 == v_size) {
305 // No existing entry, add a new one.
306 dl.vector_align.push((v_size, a));
308 _ => {} // Ignore everything else.
314 /// Returns exclusive upper bound on object size.
316 /// The theoretical maximum object size is defined as the maximum positive `isize` value.
317 /// This ensures that the `offset` semantics remain well-defined by allowing it to correctly
318 /// index every address within an object along with one byte past the end, along with allowing
319 /// `isize` to store the difference between any two pointers into an object.
321 /// The upper bound on 64-bit currently needs to be lower because LLVM uses a 64-bit integer
322 /// to represent object size in bits. It would need to be 1 << 61 to account for this, but is
323 /// currently conservatively bounded to 1 << 47 as that is enough to cover the current usable
324 /// address space on 64-bit ARMv8 and x86_64.
326 pub fn obj_size_bound(&self) -> u64 {
327 match self.pointer_size.bits() {
331 bits => panic!("obj_size_bound: unknown pointer bit size {}", bits),
336 pub fn ptr_sized_integer(&self) -> Integer {
337 match self.pointer_size.bits() {
341 bits => panic!("ptr_sized_integer: unknown pointer bit size {}", bits),
346 pub fn vector_align(&self, vec_size: Size) -> AbiAndPrefAlign {
347 for &(size, align) in &self.vector_align {
348 if size == vec_size {
352 // Default to natural alignment, which is what LLVM does.
353 // That is, use the size, rounded up to a power of 2.
354 AbiAndPrefAlign::new(Align::from_bytes(vec_size.bytes().next_power_of_two()).unwrap())
358 pub trait HasDataLayout {
359 fn data_layout(&self) -> &TargetDataLayout;
362 impl HasDataLayout for TargetDataLayout {
364 fn data_layout(&self) -> &TargetDataLayout {
369 /// Endianness of the target, which must match cfg(target-endian).
370 #[derive(Copy, Clone, PartialEq, Eq)]
377 pub fn as_str(&self) -> &'static str {
379 Self::Little => "little",
385 impl fmt::Debug for Endian {
386 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
387 f.write_str(self.as_str())
391 impl FromStr for Endian {
394 fn from_str(s: &str) -> Result<Self, Self::Err> {
396 "little" => Ok(Self::Little),
397 "big" => Ok(Self::Big),
398 _ => Err(format!(r#"unknown endian: "{}""#, s)),
403 /// Size of a type in bytes.
404 #[derive(Copy, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
405 #[cfg_attr(feature = "nightly", derive(Encodable, Decodable, HashStable_Generic))]
410 // Safety: Ord is implement as just comparing numerical values and numerical values
411 // are not changed by (de-)serialization.
412 #[cfg(feature = "nightly")]
413 unsafe impl StableOrd for Size {}
415 // This is debug-printed a lot in larger structs, don't waste too much space there
416 impl fmt::Debug for Size {
417 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
418 write!(f, "Size({} bytes)", self.bytes())
423 pub const ZERO: Size = Size { raw: 0 };
425 /// Rounds `bits` up to the next-higher byte boundary, if `bits` is
426 /// not a multiple of 8.
427 pub fn from_bits(bits: impl TryInto<u64>) -> Size {
428 let bits = bits.try_into().ok().unwrap();
429 // Avoid potential overflow from `bits + 7`.
430 Size { raw: bits / 8 + ((bits % 8) + 7) / 8 }
434 pub fn from_bytes(bytes: impl TryInto<u64>) -> Size {
435 let bytes: u64 = bytes.try_into().ok().unwrap();
440 pub fn bytes(self) -> u64 {
445 pub fn bytes_usize(self) -> usize {
446 self.bytes().try_into().unwrap()
450 pub fn bits(self) -> u64 {
452 fn overflow(bytes: u64) -> ! {
453 panic!("Size::bits: {} bytes in bits doesn't fit in u64", bytes)
456 self.bytes().checked_mul(8).unwrap_or_else(|| overflow(self.bytes()))
460 pub fn bits_usize(self) -> usize {
461 self.bits().try_into().unwrap()
465 pub fn align_to(self, align: Align) -> Size {
466 let mask = align.bytes() - 1;
467 Size::from_bytes((self.bytes() + mask) & !mask)
471 pub fn is_aligned(self, align: Align) -> bool {
472 let mask = align.bytes() - 1;
473 self.bytes() & mask == 0
477 pub fn checked_add<C: HasDataLayout>(self, offset: Size, cx: &C) -> Option<Size> {
478 let dl = cx.data_layout();
480 let bytes = self.bytes().checked_add(offset.bytes())?;
482 if bytes < dl.obj_size_bound() { Some(Size::from_bytes(bytes)) } else { None }
486 pub fn checked_mul<C: HasDataLayout>(self, count: u64, cx: &C) -> Option<Size> {
487 let dl = cx.data_layout();
489 let bytes = self.bytes().checked_mul(count)?;
490 if bytes < dl.obj_size_bound() { Some(Size::from_bytes(bytes)) } else { None }
493 /// Truncates `value` to `self` bits and then sign-extends it to 128 bits
494 /// (i.e., if it is negative, fill with 1's on the left).
496 pub fn sign_extend(self, value: u128) -> u128 {
497 let size = self.bits();
499 // Truncated until nothing is left.
503 let shift = 128 - size;
504 // Shift the unsigned value to the left, then shift back to the right as signed
505 // (essentially fills with sign bit on the left).
506 (((value << shift) as i128) >> shift) as u128
509 /// Truncates `value` to `self` bits.
511 pub fn truncate(self, value: u128) -> u128 {
512 let size = self.bits();
514 // Truncated until nothing is left.
517 let shift = 128 - size;
518 // Truncate (shift left to drop out leftover values, shift right to fill with zeroes).
519 (value << shift) >> shift
523 pub fn signed_int_min(&self) -> i128 {
524 self.sign_extend(1_u128 << (self.bits() - 1)) as i128
528 pub fn signed_int_max(&self) -> i128 {
529 i128::MAX >> (128 - self.bits())
533 pub fn unsigned_int_max(&self) -> u128 {
534 u128::MAX >> (128 - self.bits())
538 // Panicking addition, subtraction and multiplication for convenience.
539 // Avoid during layout computation, return `LayoutError` instead.
544 fn add(self, other: Size) -> Size {
545 Size::from_bytes(self.bytes().checked_add(other.bytes()).unwrap_or_else(|| {
546 panic!("Size::add: {} + {} doesn't fit in u64", self.bytes(), other.bytes())
554 fn sub(self, other: Size) -> Size {
555 Size::from_bytes(self.bytes().checked_sub(other.bytes()).unwrap_or_else(|| {
556 panic!("Size::sub: {} - {} would result in negative size", self.bytes(), other.bytes())
561 impl Mul<Size> for u64 {
564 fn mul(self, size: Size) -> Size {
569 impl Mul<u64> for Size {
572 fn mul(self, count: u64) -> Size {
573 match self.bytes().checked_mul(count) {
574 Some(bytes) => Size::from_bytes(bytes),
575 None => panic!("Size::mul: {} * {} doesn't fit in u64", self.bytes(), count),
580 impl AddAssign for Size {
582 fn add_assign(&mut self, other: Size) {
583 *self = *self + other;
587 #[cfg(feature = "nightly")]
590 fn steps_between(start: &Self, end: &Self) -> Option<usize> {
591 u64::steps_between(&start.bytes(), &end.bytes())
595 fn forward_checked(start: Self, count: usize) -> Option<Self> {
596 u64::forward_checked(start.bytes(), count).map(Self::from_bytes)
600 fn forward(start: Self, count: usize) -> Self {
601 Self::from_bytes(u64::forward(start.bytes(), count))
605 unsafe fn forward_unchecked(start: Self, count: usize) -> Self {
606 Self::from_bytes(u64::forward_unchecked(start.bytes(), count))
610 fn backward_checked(start: Self, count: usize) -> Option<Self> {
611 u64::backward_checked(start.bytes(), count).map(Self::from_bytes)
615 fn backward(start: Self, count: usize) -> Self {
616 Self::from_bytes(u64::backward(start.bytes(), count))
620 unsafe fn backward_unchecked(start: Self, count: usize) -> Self {
621 Self::from_bytes(u64::backward_unchecked(start.bytes(), count))
625 /// Alignment of a type in bytes (always a power of two).
626 #[derive(Copy, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
627 #[cfg_attr(feature = "nightly", derive(Encodable, Decodable, HashStable_Generic))]
632 // This is debug-printed a lot in larger structs, don't waste too much space there
633 impl fmt::Debug for Align {
634 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
635 write!(f, "Align({} bytes)", self.bytes())
640 pub const ONE: Align = Align { pow2: 0 };
641 pub const MAX: Align = Align { pow2: 29 };
644 pub fn from_bits(bits: u64) -> Result<Align, String> {
645 Align::from_bytes(Size::from_bits(bits).bytes())
649 pub fn from_bytes(align: u64) -> Result<Align, String> {
650 // Treat an alignment of 0 bytes like 1-byte alignment.
652 return Ok(Align::ONE);
656 fn not_power_of_2(align: u64) -> String {
657 format!("`{}` is not a power of 2", align)
661 fn too_large(align: u64) -> String {
662 format!("`{}` is too large", align)
665 let mut bytes = align;
666 let mut pow2: u8 = 0;
667 while (bytes & 1) == 0 {
672 return Err(not_power_of_2(align));
674 if pow2 > Self::MAX.pow2 {
675 return Err(too_large(align));
682 pub fn bytes(self) -> u64 {
687 pub fn bits(self) -> u64 {
691 /// Computes the best alignment possible for the given offset
692 /// (the largest power of two that the offset is a multiple of).
694 /// N.B., for an offset of `0`, this happens to return `2^64`.
696 pub fn max_for_offset(offset: Size) -> Align {
697 Align { pow2: offset.bytes().trailing_zeros() as u8 }
700 /// Lower the alignment, if necessary, such that the given offset
701 /// is aligned to it (the offset is a multiple of the alignment).
703 pub fn restrict_for_offset(self, offset: Size) -> Align {
704 self.min(Align::max_for_offset(offset))
708 /// A pair of alignments, ABI-mandated and preferred.
709 #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
710 #[cfg_attr(feature = "nightly", derive(HashStable_Generic))]
712 pub struct AbiAndPrefAlign {
717 impl AbiAndPrefAlign {
719 pub fn new(align: Align) -> AbiAndPrefAlign {
720 AbiAndPrefAlign { abi: align, pref: align }
724 pub fn min(self, other: AbiAndPrefAlign) -> AbiAndPrefAlign {
725 AbiAndPrefAlign { abi: self.abi.min(other.abi), pref: self.pref.min(other.pref) }
729 pub fn max(self, other: AbiAndPrefAlign) -> AbiAndPrefAlign {
730 AbiAndPrefAlign { abi: self.abi.max(other.abi), pref: self.pref.max(other.pref) }
734 /// Integers, also used for enum discriminants.
735 #[derive(Copy, Clone, PartialEq, Eq, PartialOrd, Ord, Hash, Debug)]
736 #[cfg_attr(feature = "nightly", derive(Encodable, Decodable, HashStable_Generic))]
748 pub fn size(self) -> Size {
750 I8 => Size::from_bytes(1),
751 I16 => Size::from_bytes(2),
752 I32 => Size::from_bytes(4),
753 I64 => Size::from_bytes(8),
754 I128 => Size::from_bytes(16),
758 /// Gets the Integer type from an IntegerType.
759 pub fn from_attr<C: HasDataLayout>(cx: &C, ity: IntegerType) -> Integer {
760 let dl = cx.data_layout();
763 IntegerType::Pointer(_) => dl.ptr_sized_integer(),
764 IntegerType::Fixed(x, _) => x,
768 pub fn align<C: HasDataLayout>(self, cx: &C) -> AbiAndPrefAlign {
769 let dl = cx.data_layout();
776 I128 => dl.i128_align,
780 /// Returns the largest signed value that can be represented by this Integer.
782 pub fn signed_max(self) -> i128 {
784 I8 => i8::MAX as i128,
785 I16 => i16::MAX as i128,
786 I32 => i32::MAX as i128,
787 I64 => i64::MAX as i128,
792 /// Finds the smallest Integer type which can represent the signed value.
794 pub fn fit_signed(x: i128) -> Integer {
796 -0x0000_0000_0000_0080..=0x0000_0000_0000_007f => I8,
797 -0x0000_0000_0000_8000..=0x0000_0000_0000_7fff => I16,
798 -0x0000_0000_8000_0000..=0x0000_0000_7fff_ffff => I32,
799 -0x8000_0000_0000_0000..=0x7fff_ffff_ffff_ffff => I64,
804 /// Finds the smallest Integer type which can represent the unsigned value.
806 pub fn fit_unsigned(x: u128) -> Integer {
808 0..=0x0000_0000_0000_00ff => I8,
809 0..=0x0000_0000_0000_ffff => I16,
810 0..=0x0000_0000_ffff_ffff => I32,
811 0..=0xffff_ffff_ffff_ffff => I64,
816 /// Finds the smallest integer with the given alignment.
817 pub fn for_align<C: HasDataLayout>(cx: &C, wanted: Align) -> Option<Integer> {
818 let dl = cx.data_layout();
820 [I8, I16, I32, I64, I128].into_iter().find(|&candidate| {
821 wanted == candidate.align(dl).abi && wanted.bytes() == candidate.size().bytes()
825 /// Find the largest integer with the given alignment or less.
826 pub fn approximate_align<C: HasDataLayout>(cx: &C, wanted: Align) -> Integer {
827 let dl = cx.data_layout();
829 // FIXME(eddyb) maybe include I128 in the future, when it works everywhere.
830 for candidate in [I64, I32, I16] {
831 if wanted >= candidate.align(dl).abi && wanted.bytes() >= candidate.size().bytes() {
838 // FIXME(eddyb) consolidate this and other methods that find the appropriate
839 // `Integer` given some requirements.
841 pub fn from_size(size: Size) -> Result<Self, String> {
843 8 => Ok(Integer::I8),
844 16 => Ok(Integer::I16),
845 32 => Ok(Integer::I32),
846 64 => Ok(Integer::I64),
847 128 => Ok(Integer::I128),
848 _ => Err(format!("rust does not support integers with {} bits", size.bits())),
853 /// Fundamental unit of memory access and layout.
854 #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
855 #[cfg_attr(feature = "nightly", derive(HashStable_Generic))]
857 /// The `bool` is the signedness of the `Integer` type.
859 /// One would think we would not care about such details this low down,
860 /// but some ABIs are described in terms of C types and ISAs where the
861 /// integer arithmetic is done on {sign,zero}-extended registers, e.g.
862 /// a negative integer passed by zero-extension will appear positive in
863 /// the callee, and most operations on it will produce the wrong values.
867 Pointer(AddressSpace),
871 pub fn size<C: HasDataLayout>(self, cx: &C) -> Size {
872 let dl = cx.data_layout();
875 Int(i, _) => i.size(),
876 F32 => Size::from_bits(32),
877 F64 => Size::from_bits(64),
878 // FIXME(erikdesjardins): ignoring address space is technically wrong, pointers in
879 // different address spaces can have different sizes
880 // (but TargetDataLayout doesn't currently parse that part of the DL string)
881 Pointer(_) => dl.pointer_size,
885 pub fn align<C: HasDataLayout>(self, cx: &C) -> AbiAndPrefAlign {
886 let dl = cx.data_layout();
889 Int(i, _) => i.align(dl),
892 // FIXME(erikdesjardins): ignoring address space is technically wrong, pointers in
893 // different address spaces can have different alignments
894 // (but TargetDataLayout doesn't currently parse that part of the DL string)
895 Pointer(_) => dl.pointer_align,
900 /// Inclusive wrap-around range of valid values, that is, if
901 /// start > end, it represents `start..=MAX`,
902 /// followed by `0..=end`.
904 /// That is, for an i8 primitive, a range of `254..=2` means following
907 /// 254 (-2), 255 (-1), 0, 1, 2
909 /// This is intended specifically to mirror LLVM’s `!range` metadata semantics.
910 #[derive(Clone, Copy, PartialEq, Eq, Hash)]
911 #[cfg_attr(feature = "nightly", derive(HashStable_Generic))]
912 pub struct WrappingRange {
918 pub fn full(size: Size) -> Self {
919 Self { start: 0, end: size.unsigned_int_max() }
922 /// Returns `true` if `v` is contained in the range.
924 pub fn contains(&self, v: u128) -> bool {
925 if self.start <= self.end {
926 self.start <= v && v <= self.end
928 self.start <= v || v <= self.end
932 /// Returns `self` with replaced `start`
934 pub fn with_start(mut self, start: u128) -> Self {
939 /// Returns `self` with replaced `end`
941 pub fn with_end(mut self, end: u128) -> Self {
946 /// Returns `true` if `size` completely fills the range.
948 pub fn is_full_for(&self, size: Size) -> bool {
949 let max_value = size.unsigned_int_max();
950 debug_assert!(self.start <= max_value && self.end <= max_value);
951 self.start == (self.end.wrapping_add(1) & max_value)
955 impl fmt::Debug for WrappingRange {
956 fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
957 if self.start > self.end {
958 write!(fmt, "(..={}) | ({}..)", self.end, self.start)?;
960 write!(fmt, "{}..={}", self.start, self.end)?;
966 /// Information about one scalar component of a Rust type.
967 #[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
968 #[cfg_attr(feature = "nightly", derive(HashStable_Generic))]
973 // FIXME(eddyb) always use the shortest range, e.g., by finding
974 // the largest space between two consecutive valid values and
975 // taking everything else as the (shortest) valid range.
976 valid_range: WrappingRange,
979 /// Even for unions, we need to use the correct registers for the kind of
980 /// values inside the union, so we keep the `Primitive` type around. We
981 /// also use it to compute the size of the scalar.
982 /// However, unions never have niches and even allow undef,
983 /// so there is no `valid_range`.
990 pub fn is_bool(&self) -> bool {
993 Scalar::Initialized {
994 value: Int(I8, false),
995 valid_range: WrappingRange { start: 0, end: 1 }
1000 /// Get the primitive representation of this type, ignoring the valid range and whether the
1001 /// value is allowed to be undefined (due to being a union).
1002 pub fn primitive(&self) -> Primitive {
1004 Scalar::Initialized { value, .. } | Scalar::Union { value } => value,
1008 pub fn align(self, cx: &impl HasDataLayout) -> AbiAndPrefAlign {
1009 self.primitive().align(cx)
1012 pub fn size(self, cx: &impl HasDataLayout) -> Size {
1013 self.primitive().size(cx)
1017 pub fn to_union(&self) -> Self {
1018 Self::Union { value: self.primitive() }
1022 pub fn valid_range(&self, cx: &impl HasDataLayout) -> WrappingRange {
1024 Scalar::Initialized { valid_range, .. } => valid_range,
1025 Scalar::Union { value } => WrappingRange::full(value.size(cx)),
1030 /// Allows the caller to mutate the valid range. This operation will panic if attempted on a union.
1031 pub fn valid_range_mut(&mut self) -> &mut WrappingRange {
1033 Scalar::Initialized { valid_range, .. } => valid_range,
1034 Scalar::Union { .. } => panic!("cannot change the valid range of a union"),
1038 /// Returns `true` if all possible numbers are valid, i.e `valid_range` covers the whole layout
1040 pub fn is_always_valid<C: HasDataLayout>(&self, cx: &C) -> bool {
1042 Scalar::Initialized { valid_range, .. } => valid_range.is_full_for(self.size(cx)),
1043 Scalar::Union { .. } => true,
1047 /// Returns `true` if this type can be left uninit.
1049 pub fn is_uninit_valid(&self) -> bool {
1051 Scalar::Initialized { .. } => false,
1052 Scalar::Union { .. } => true,
1057 /// Describes how the fields of a type are located in memory.
1058 #[derive(PartialEq, Eq, Hash, Clone, Debug)]
1059 #[cfg_attr(feature = "nightly", derive(HashStable_Generic))]
1060 pub enum FieldsShape {
1061 /// Scalar primitives and `!`, which never have fields.
1064 /// All fields start at no offset. The `usize` is the field count.
1065 Union(NonZeroUsize),
1067 /// Array/vector-like placement, with all fields of identical types.
1068 Array { stride: Size, count: u64 },
1070 /// Struct-like placement, with precomputed offsets.
1072 /// Fields are guaranteed to not overlap, but note that gaps
1073 /// before, between and after all the fields are NOT always
1074 /// padding, and as such their contents may not be discarded.
1075 /// For example, enum variants leave a gap at the start,
1076 /// where the discriminant field in the enum layout goes.
1078 /// Offsets for the first byte of each field,
1079 /// ordered to match the source definition order.
1080 /// This vector does not go in increasing order.
1081 // FIXME(eddyb) use small vector optimization for the common case.
1084 /// Maps source order field indices to memory order indices,
1085 /// depending on how the fields were reordered (if at all).
1086 /// This is a permutation, with both the source order and the
1087 /// memory order using the same (0..n) index ranges.
1089 /// Note that during computation of `memory_index`, sometimes
1090 /// it is easier to operate on the inverse mapping (that is,
1091 /// from memory order to source order), and that is usually
1092 /// named `inverse_memory_index`.
1094 // FIXME(eddyb) build a better abstraction for permutations, if possible.
1095 // FIXME(camlorn) also consider small vector optimization here.
1096 memory_index: Vec<u32>,
1102 pub fn count(&self) -> usize {
1104 FieldsShape::Primitive => 0,
1105 FieldsShape::Union(count) => count.get(),
1106 FieldsShape::Array { count, .. } => count.try_into().unwrap(),
1107 FieldsShape::Arbitrary { ref offsets, .. } => offsets.len(),
1112 pub fn offset(&self, i: usize) -> Size {
1114 FieldsShape::Primitive => {
1115 unreachable!("FieldsShape::offset: `Primitive`s have no fields")
1117 FieldsShape::Union(count) => {
1120 "tried to access field {} of union with {} fields",
1126 FieldsShape::Array { stride, count } => {
1127 let i = u64::try_from(i).unwrap();
1131 FieldsShape::Arbitrary { ref offsets, .. } => offsets[i],
1136 pub fn memory_index(&self, i: usize) -> usize {
1138 FieldsShape::Primitive => {
1139 unreachable!("FieldsShape::memory_index: `Primitive`s have no fields")
1141 FieldsShape::Union(_) | FieldsShape::Array { .. } => i,
1142 FieldsShape::Arbitrary { ref memory_index, .. } => memory_index[i].try_into().unwrap(),
1146 /// Gets source indices of the fields by increasing offsets.
1148 pub fn index_by_increasing_offset<'a>(&'a self) -> impl Iterator<Item = usize> + 'a {
1149 let mut inverse_small = [0u8; 64];
1150 let mut inverse_big = vec![];
1151 let use_small = self.count() <= inverse_small.len();
1153 // We have to write this logic twice in order to keep the array small.
1154 if let FieldsShape::Arbitrary { ref memory_index, .. } = *self {
1156 for i in 0..self.count() {
1157 inverse_small[memory_index[i] as usize] = i as u8;
1160 inverse_big = vec![0; self.count()];
1161 for i in 0..self.count() {
1162 inverse_big[memory_index[i] as usize] = i as u32;
1167 (0..self.count()).map(move |i| match *self {
1168 FieldsShape::Primitive | FieldsShape::Union(_) | FieldsShape::Array { .. } => i,
1169 FieldsShape::Arbitrary { .. } => {
1171 inverse_small[i] as usize
1173 inverse_big[i] as usize
1180 /// An identifier that specifies the address space that some operation
1181 /// should operate on. Special address spaces have an effect on code generation,
1182 /// depending on the target and the address spaces it implements.
1183 #[derive(Copy, Clone, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)]
1184 #[cfg_attr(feature = "nightly", derive(HashStable_Generic))]
1185 pub struct AddressSpace(pub u32);
1188 /// The default address space, corresponding to data space.
1189 pub const DATA: Self = AddressSpace(0);
1192 /// Describes how values of the type are passed by target ABIs,
1193 /// in terms of categories of C types there are ABI rules for.
1194 #[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
1195 #[cfg_attr(feature = "nightly", derive(HashStable_Generic))]
1200 ScalarPair(Scalar, Scalar),
1206 /// If true, the size is exact, otherwise it's only a lower bound.
1212 /// Returns `true` if the layout corresponds to an unsized type.
1214 pub fn is_unsized(&self) -> bool {
1216 Abi::Uninhabited | Abi::Scalar(_) | Abi::ScalarPair(..) | Abi::Vector { .. } => false,
1217 Abi::Aggregate { sized } => !sized,
1222 pub fn is_sized(&self) -> bool {
1226 /// Returns `true` if this is a single signed integer scalar
1228 pub fn is_signed(&self) -> bool {
1230 Abi::Scalar(scal) => match scal.primitive() {
1231 Primitive::Int(_, signed) => signed,
1234 _ => panic!("`is_signed` on non-scalar ABI {:?}", self),
1238 /// Returns `true` if this is an uninhabited type
1240 pub fn is_uninhabited(&self) -> bool {
1241 matches!(*self, Abi::Uninhabited)
1244 /// Returns `true` is this is a scalar type
1246 pub fn is_scalar(&self) -> bool {
1247 matches!(*self, Abi::Scalar(_))
1251 #[derive(PartialEq, Eq, Hash, Clone, Debug)]
1252 #[cfg_attr(feature = "nightly", derive(HashStable_Generic))]
1253 pub enum Variants<V: Idx> {
1254 /// Single enum variants, structs/tuples, unions, and all non-ADTs.
1255 Single { index: V },
1257 /// Enum-likes with more than one inhabited variant: each variant comes with
1258 /// a *discriminant* (usually the same as the variant index but the user can
1259 /// assign explicit discriminant values). That discriminant is encoded
1260 /// as a *tag* on the machine. The layout of each variant is
1261 /// a struct, and they all have space reserved for the tag.
1262 /// For enums, the tag is the sole field of the layout.
1265 tag_encoding: TagEncoding<V>,
1267 variants: IndexVec<V, LayoutS<V>>,
1271 #[derive(PartialEq, Eq, Hash, Clone, Debug)]
1272 #[cfg_attr(feature = "nightly", derive(HashStable_Generic))]
1273 pub enum TagEncoding<V: Idx> {
1274 /// The tag directly stores the discriminant, but possibly with a smaller layout
1275 /// (so converting the tag to the discriminant can require sign extension).
1278 /// Niche (values invalid for a type) encoding the discriminant:
1279 /// Discriminant and variant index coincide.
1280 /// The variant `untagged_variant` contains a niche at an arbitrary
1281 /// offset (field `tag_field` of the enum), which for a variant with
1282 /// discriminant `d` is set to
1283 /// `(d - niche_variants.start).wrapping_add(niche_start)`.
1285 /// For example, `Option<(usize, &T)>` is represented such that
1286 /// `None` has a null pointer for the second tuple field, and
1287 /// `Some` is the identity function (with a non-null reference).
1288 Niche { untagged_variant: V, niche_variants: RangeInclusive<V>, niche_start: u128 },
1291 #[derive(Clone, Copy, PartialEq, Eq, Hash, Debug)]
1292 #[cfg_attr(feature = "nightly", derive(HashStable_Generic))]
1295 pub value: Primitive,
1296 pub valid_range: WrappingRange,
1300 pub fn from_scalar<C: HasDataLayout>(cx: &C, offset: Size, scalar: Scalar) -> Option<Self> {
1301 let Scalar::Initialized { value, valid_range } = scalar else { return None };
1302 let niche = Niche { offset, value, valid_range };
1303 if niche.available(cx) > 0 { Some(niche) } else { None }
1306 pub fn available<C: HasDataLayout>(&self, cx: &C) -> u128 {
1307 let Self { value, valid_range: v, .. } = *self;
1308 let size = value.size(cx);
1309 assert!(size.bits() <= 128);
1310 let max_value = size.unsigned_int_max();
1312 // Find out how many values are outside the valid range.
1313 let niche = v.end.wrapping_add(1)..v.start;
1314 niche.end.wrapping_sub(niche.start) & max_value
1317 pub fn reserve<C: HasDataLayout>(&self, cx: &C, count: u128) -> Option<(u128, Scalar)> {
1320 let Self { value, valid_range: v, .. } = *self;
1321 let size = value.size(cx);
1322 assert!(size.bits() <= 128);
1323 let max_value = size.unsigned_int_max();
1325 let niche = v.end.wrapping_add(1)..v.start;
1326 let available = niche.end.wrapping_sub(niche.start) & max_value;
1327 if count > available {
1331 // Extend the range of valid values being reserved by moving either `v.start` or `v.end` bound.
1332 // Given an eventual `Option<T>`, we try to maximize the chance for `None` to occupy the niche of zero.
1333 // This is accomplished by preferring enums with 2 variants(`count==1`) and always taking the shortest path to niche zero.
1334 // Having `None` in niche zero can enable some special optimizations.
1336 // Bound selection criteria:
1337 // 1. Select closest to zero given wrapping semantics.
1338 // 2. Avoid moving past zero if possible.
1340 // In practice this means that enums with `count > 1` are unlikely to claim niche zero, since they have to fit perfectly.
1341 // If niche zero is already reserved, the selection of bounds are of little interest.
1342 let move_start = |v: WrappingRange| {
1343 let start = v.start.wrapping_sub(count) & max_value;
1344 Some((start, Scalar::Initialized { value, valid_range: v.with_start(start) }))
1346 let move_end = |v: WrappingRange| {
1347 let start = v.end.wrapping_add(1) & max_value;
1348 let end = v.end.wrapping_add(count) & max_value;
1349 Some((start, Scalar::Initialized { value, valid_range: v.with_end(end) }))
1351 let distance_end_zero = max_value - v.end;
1352 if v.start > v.end {
1353 // zero is unavailable because wrapping occurs
1355 } else if v.start <= distance_end_zero {
1356 if count <= v.start {
1359 // moved past zero, use other bound
1363 let end = v.end.wrapping_add(count) & max_value;
1364 let overshot_zero = (1..=v.end).contains(&end);
1366 // moved past zero, use other bound
1375 #[derive(PartialEq, Eq, Hash, Clone)]
1376 #[cfg_attr(feature = "nightly", derive(HashStable_Generic))]
1377 pub struct LayoutS<V: Idx> {
1378 /// Says where the fields are located within the layout.
1379 pub fields: FieldsShape,
1381 /// Encodes information about multi-variant layouts.
1382 /// Even with `Multiple` variants, a layout still has its own fields! Those are then
1383 /// shared between all variants. One of them will be the discriminant,
1384 /// but e.g. generators can have more.
1386 /// To access all fields of this layout, both `fields` and the fields of the active variant
1387 /// must be taken into account.
1388 pub variants: Variants<V>,
1390 /// The `abi` defines how this data is passed between functions, and it defines
1391 /// value restrictions via `valid_range`.
1393 /// Note that this is entirely orthogonal to the recursive structure defined by
1394 /// `variants` and `fields`; for example, `ManuallyDrop<Result<isize, isize>>` has
1395 /// `Abi::ScalarPair`! So, even with non-`Aggregate` `abi`, `fields` and `variants`
1396 /// have to be taken into account to find all fields of this layout.
1399 /// The leaf scalar with the largest number of invalid values
1400 /// (i.e. outside of its `valid_range`), if it exists.
1401 pub largest_niche: Option<Niche>,
1403 pub align: AbiAndPrefAlign,
1407 impl<V: Idx> LayoutS<V> {
1408 pub fn scalar<C: HasDataLayout>(cx: &C, scalar: Scalar) -> Self {
1409 let largest_niche = Niche::from_scalar(cx, Size::ZERO, scalar);
1410 let size = scalar.size(cx);
1411 let align = scalar.align(cx);
1413 variants: Variants::Single { index: V::new(0) },
1414 fields: FieldsShape::Primitive,
1415 abi: Abi::Scalar(scalar),
1423 impl<V: Idx> fmt::Debug for LayoutS<V> {
1424 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
1425 // This is how `Layout` used to print before it become
1426 // `Interned<LayoutS>`. We print it like this to avoid having to update
1427 // expected output in a lot of tests.
1428 let LayoutS { size, align, abi, fields, largest_niche, variants } = self;
1429 f.debug_struct("Layout")
1430 .field("size", size)
1431 .field("align", align)
1433 .field("fields", fields)
1434 .field("largest_niche", largest_niche)
1435 .field("variants", variants)
1440 #[derive(Copy, Clone, PartialEq, Eq, Debug)]
1441 pub enum PointerKind {
1442 /// Most general case, we know no restrictions to tell LLVM.
1445 /// `&T` where `T` contains no `UnsafeCell`, is `dereferenceable`, `noalias` and `readonly`.
1448 /// `&mut T` which is `dereferenceable` and `noalias` but not `readonly`.
1451 /// `&mut !Unpin`, which is `dereferenceable` but neither `noalias` nor `readonly`.
1452 UniqueBorrowedPinned,
1454 /// `Box<T>`, which is `noalias` (even on return types, unlike the above) but neither `readonly`
1455 /// nor `dereferenceable`.
1459 /// Note that this information is advisory only, and backends are free to ignore it.
1460 /// It can only be used to encode potential optimizations, but no critical information.
1461 #[derive(Copy, Clone, Debug)]
1462 pub struct PointeeInfo {
1465 pub safe: Option<PointerKind>,
1468 /// Used in `might_permit_raw_init` to indicate the kind of initialisation
1469 /// that is checked to be valid
1470 #[derive(Copy, Clone, Debug, PartialEq, Eq)]
1473 UninitMitigated0x01Fill,
1476 impl<V: Idx> LayoutS<V> {
1477 /// Returns `true` if the layout corresponds to an unsized type.
1478 pub fn is_unsized(&self) -> bool {
1479 self.abi.is_unsized()
1482 pub fn is_sized(&self) -> bool {
1486 /// Returns `true` if the type is a ZST and not unsized.
1487 pub fn is_zst(&self) -> bool {
1489 Abi::Scalar(_) | Abi::ScalarPair(..) | Abi::Vector { .. } => false,
1490 Abi::Uninhabited => self.size.bytes() == 0,
1491 Abi::Aggregate { sized } => sized && self.size.bytes() == 0,
1496 #[derive(Copy, Clone, Debug)]
1497 pub enum StructKind {
1498 /// A tuple, closure, or univariant which cannot be coerced to unsized.
1500 /// A univariant, the last field of which may be coerced to unsized.
1502 /// A univariant, but with a prefix of an arbitrary size & alignment (e.g., enum tag).
1503 Prefixed(Size, Align),