From 6959818ec2472295e46e1cbe7526e42698cff849 Mon Sep 17 00:00:00 2001 From: Jubilee Young Date: Sat, 16 Dec 2023 01:37:00 -0800 Subject: [PATCH 1/4] Hoist fn absent --- compiler/rustc_abi/src/layout.rs | 30 ++++++++++++++++++------------ 1 file changed, 18 insertions(+), 12 deletions(-) diff --git a/compiler/rustc_abi/src/layout.rs b/compiler/rustc_abi/src/layout.rs index 5252472261f30..e73170c01f836 100644 --- a/compiler/rustc_abi/src/layout.rs +++ b/compiler/rustc_abi/src/layout.rs @@ -11,6 +11,24 @@ use crate::{ Variants, WrappingRange, }; +// A variant is absent if it's uninhabited and only has ZST fields. +// Present uninhabited variants only require space for their fields, +// but *not* an encoding of the discriminant (e.g., a tag value). +// See issue #49298 for more details on the need to leave space +// for non-ZST uninhabited data (mostly partial initialization). +fn absent<'a, FieldIdx, VariantIdx, F>(fields: &IndexSlice) -> bool +where + FieldIdx: Idx, + VariantIdx: Idx, + F: Deref> + fmt::Debug, +{ + let uninhabited = fields.iter().any(|f| f.abi.is_uninhabited()); + // We cannot ignore alignment; that might lead us to entirely discard a variant and + // produce an enum that is less aligned than it should be! + let is_1zst = fields.iter().all(|f| f.is_1zst()); + uninhabited && is_1zst +} + pub trait LayoutCalculator { type TargetDataLayoutRef: Borrow; @@ -168,18 +186,6 @@ pub trait LayoutCalculator { Scalar::Initialized { value, valid_range: WrappingRange::full(size) } }; - // A variant is absent if it's uninhabited and only has ZST fields. - // Present uninhabited variants only require space for their fields, - // but *not* an encoding of the discriminant (e.g., a tag value). - // See issue #49298 for more details on the need to leave space - // for non-ZST uninhabited data (mostly partial initialization). - let absent = |fields: &IndexSlice| { - let uninhabited = fields.iter().any(|f| f.abi.is_uninhabited()); - // We cannot ignore alignment; that might lead us to entirely discard a variant and - // produce an enum that is less aligned than it should be! - let is_1zst = fields.iter().all(|f| f.is_1zst()); - uninhabited && is_1zst - }; let (present_first, present_second) = { let mut present_variants = variants .iter_enumerated() From f116bc6e27042123ed25a43303615ba7d54a2f8e Mon Sep 17 00:00:00 2001 From: Jubilee Young Date: Sat, 16 Dec 2023 01:43:40 -0800 Subject: [PATCH 2/4] Manually inline scalar_unit --- compiler/rustc_abi/src/layout.rs | 10 +++------- 1 file changed, 3 insertions(+), 7 deletions(-) diff --git a/compiler/rustc_abi/src/layout.rs b/compiler/rustc_abi/src/layout.rs index e73170c01f836..9f84c1aeaaf14 100644 --- a/compiler/rustc_abi/src/layout.rs +++ b/compiler/rustc_abi/src/layout.rs @@ -180,12 +180,6 @@ pub trait LayoutCalculator { let dl = self.current_data_layout(); let dl = dl.borrow(); - let scalar_unit = |value: Primitive| { - let size = value.size(dl); - assert!(size.bits() <= 128); - Scalar::Initialized { value, valid_range: WrappingRange::full(size) } - }; - let (present_first, present_second) = { let mut present_variants = variants .iter_enumerated() @@ -678,7 +672,9 @@ pub trait LayoutCalculator { } if let Some((prim, offset)) = common_prim { let prim_scalar = if common_prim_initialized_in_all_variants { - scalar_unit(prim) + let size = prim.size(dl); + assert!(size.bits() <= 128); + Scalar::Initialized { value: prim, valid_range: WrappingRange::full(size) } } else { // Common prim might be uninit. Scalar::Union { value: prim } From b525f76bb50438ce5748635799c63421944b57c5 Mon Sep 17 00:00:00 2001 From: Jubilee Young Date: Sat, 16 Dec 2023 01:44:13 -0800 Subject: [PATCH 3/4] Extract fn layout_of_enum --- compiler/rustc_abi/src/layout.rs | 972 ++++++++++++++++--------------- 1 file changed, 496 insertions(+), 476 deletions(-) diff --git a/compiler/rustc_abi/src/layout.rs b/compiler/rustc_abi/src/layout.rs index 9f84c1aeaaf14..102bef9f6de37 100644 --- a/compiler/rustc_abi/src/layout.rs +++ b/compiler/rustc_abi/src/layout.rs @@ -291,482 +291,15 @@ pub trait LayoutCalculator { // structs. (We have also handled univariant enums // that allow representation optimization.) assert!(is_enum); - - // Until we've decided whether to use the tagged or - // niche filling LayoutS, we don't want to intern the - // variant layouts, so we can't store them in the - // overall LayoutS. Store the overall LayoutS - // and the variant LayoutSs here until then. - struct TmpLayout { - layout: LayoutS, - variants: IndexVec>, - } - - let calculate_niche_filling_layout = || -> Option> { - if dont_niche_optimize_enum { - return None; - } - - if variants.len() < 2 { - return None; - } - - let mut align = dl.aggregate_align; - let mut max_repr_align = repr.align; - let mut unadjusted_abi_align = align.abi; - - let mut variant_layouts = variants - .iter_enumerated() - .map(|(j, v)| { - let mut st = self.univariant(dl, v, repr, StructKind::AlwaysSized)?; - st.variants = Variants::Single { index: j }; - - align = align.max(st.align); - max_repr_align = max_repr_align.max(st.max_repr_align); - unadjusted_abi_align = unadjusted_abi_align.max(st.unadjusted_abi_align); - - Some(st) - }) - .collect::>>()?; - - let largest_variant_index = variant_layouts - .iter_enumerated() - .max_by_key(|(_i, layout)| layout.size.bytes()) - .map(|(i, _layout)| i)?; - - let all_indices = variants.indices(); - let needs_disc = - |index: VariantIdx| index != largest_variant_index && !absent(&variants[index]); - let niche_variants = all_indices.clone().find(|v| needs_disc(*v)).unwrap() - ..=all_indices.rev().find(|v| needs_disc(*v)).unwrap(); - - let count = - (niche_variants.end().index() as u128 - niche_variants.start().index() as u128) + 1; - - // Find the field with the largest niche - let (field_index, niche, (niche_start, niche_scalar)) = variants[largest_variant_index] - .iter() - .enumerate() - .filter_map(|(j, field)| Some((j, field.largest_niche?))) - .max_by_key(|(_, niche)| niche.available(dl)) - .and_then(|(j, niche)| Some((j, niche, niche.reserve(dl, count)?)))?; - let niche_offset = - niche.offset + variant_layouts[largest_variant_index].fields.offset(field_index); - let niche_size = niche.value.size(dl); - let size = variant_layouts[largest_variant_index].size.align_to(align.abi); - - let all_variants_fit = variant_layouts.iter_enumerated_mut().all(|(i, layout)| { - if i == largest_variant_index { - return true; - } - - layout.largest_niche = None; - - if layout.size <= niche_offset { - // This variant will fit before the niche. - return true; - } - - // Determine if it'll fit after the niche. - let this_align = layout.align.abi; - let this_offset = (niche_offset + niche_size).align_to(this_align); - - if this_offset + layout.size > size { - return false; - } - - // It'll fit, but we need to make some adjustments. - match layout.fields { - FieldsShape::Arbitrary { ref mut offsets, .. } => { - for offset in offsets.iter_mut() { - *offset += this_offset; - } - } - FieldsShape::Primitive | FieldsShape::Array { .. } | FieldsShape::Union(..) => { - panic!("Layout of fields should be Arbitrary for variants") - } - } - - // It can't be a Scalar or ScalarPair because the offset isn't 0. - if !layout.abi.is_uninhabited() { - layout.abi = Abi::Aggregate { sized: true }; - } - layout.size += this_offset; - - true - }); - - if !all_variants_fit { - return None; - } - - let largest_niche = Niche::from_scalar(dl, niche_offset, niche_scalar); - - let others_zst = variant_layouts - .iter_enumerated() - .all(|(i, layout)| i == largest_variant_index || layout.size == Size::ZERO); - let same_size = size == variant_layouts[largest_variant_index].size; - let same_align = align == variant_layouts[largest_variant_index].align; - - let abi = if variant_layouts.iter().all(|v| v.abi.is_uninhabited()) { - Abi::Uninhabited - } else if same_size && same_align && others_zst { - match variant_layouts[largest_variant_index].abi { - // When the total alignment and size match, we can use the - // same ABI as the scalar variant with the reserved niche. - Abi::Scalar(_) => Abi::Scalar(niche_scalar), - Abi::ScalarPair(first, second) => { - // Only the niche is guaranteed to be initialised, - // so use union layouts for the other primitive. - if niche_offset == Size::ZERO { - Abi::ScalarPair(niche_scalar, second.to_union()) - } else { - Abi::ScalarPair(first.to_union(), niche_scalar) - } - } - _ => Abi::Aggregate { sized: true }, - } - } else { - Abi::Aggregate { sized: true } - }; - - let layout = LayoutS { - variants: Variants::Multiple { - tag: niche_scalar, - tag_encoding: TagEncoding::Niche { - untagged_variant: largest_variant_index, - niche_variants, - niche_start, - }, - tag_field: 0, - variants: IndexVec::new(), - }, - fields: FieldsShape::Arbitrary { - offsets: [niche_offset].into(), - memory_index: [0].into(), - }, - abi, - largest_niche, - size, - align, - max_repr_align, - unadjusted_abi_align, - }; - - Some(TmpLayout { layout, variants: variant_layouts }) - }; - - let niche_filling_layout = calculate_niche_filling_layout(); - - let (mut min, mut max) = (i128::MAX, i128::MIN); - let discr_type = repr.discr_type(); - let bits = Integer::from_attr(dl, discr_type).size().bits(); - for (i, mut val) in discriminants { - if variants[i].iter().any(|f| f.abi.is_uninhabited()) { - continue; - } - if discr_type.is_signed() { - // sign extend the raw representation to be an i128 - val = (val << (128 - bits)) >> (128 - bits); - } - if val < min { - min = val; - } - if val > max { - max = val; - } - } - // We might have no inhabited variants, so pretend there's at least one. - if (min, max) == (i128::MAX, i128::MIN) { - min = 0; - max = 0; - } - assert!(min <= max, "discriminant range is {min}...{max}"); - let (min_ity, signed) = discr_range_of_repr(min, max); //Integer::repr_discr(tcx, ty, &repr, min, max); - - let mut align = dl.aggregate_align; - let mut max_repr_align = repr.align; - let mut unadjusted_abi_align = align.abi; - - let mut size = Size::ZERO; - - // We're interested in the smallest alignment, so start large. - let mut start_align = Align::from_bytes(256).unwrap(); - assert_eq!(Integer::for_align(dl, start_align), None); - - // repr(C) on an enum tells us to make a (tag, union) layout, - // so we need to grow the prefix alignment to be at least - // the alignment of the union. (This value is used both for - // determining the alignment of the overall enum, and the - // determining the alignment of the payload after the tag.) - let mut prefix_align = min_ity.align(dl).abi; - if repr.c() { - for fields in variants { - for field in fields { - prefix_align = prefix_align.max(field.align.abi); - } - } - } - - // Create the set of structs that represent each variant. - let mut layout_variants = variants - .iter_enumerated() - .map(|(i, field_layouts)| { - let mut st = self.univariant( - dl, - field_layouts, - repr, - StructKind::Prefixed(min_ity.size(), prefix_align), - )?; - st.variants = Variants::Single { index: i }; - // Find the first field we can't move later - // to make room for a larger discriminant. - for field_idx in st.fields.index_by_increasing_offset() { - let field = &field_layouts[FieldIdx::new(field_idx)]; - if !field.is_1zst() { - start_align = start_align.min(field.align.abi); - break; - } - } - size = cmp::max(size, st.size); - align = align.max(st.align); - max_repr_align = max_repr_align.max(st.max_repr_align); - unadjusted_abi_align = unadjusted_abi_align.max(st.unadjusted_abi_align); - Some(st) - }) - .collect::>>()?; - - // Align the maximum variant size to the largest alignment. - size = size.align_to(align.abi); - - // FIXME(oli-obk): deduplicate and harden these checks - if size.bytes() >= dl.obj_size_bound() { - return None; - } - - let typeck_ity = Integer::from_attr(dl, repr.discr_type()); - if typeck_ity < min_ity { - // It is a bug if Layout decided on a greater discriminant size than typeck for - // some reason at this point (based on values discriminant can take on). Mostly - // because this discriminant will be loaded, and then stored into variable of - // type calculated by typeck. Consider such case (a bug): typeck decided on - // byte-sized discriminant, but layout thinks we need a 16-bit to store all - // discriminant values. That would be a bug, because then, in codegen, in order - // to store this 16-bit discriminant into 8-bit sized temporary some of the - // space necessary to represent would have to be discarded (or layout is wrong - // on thinking it needs 16 bits) - panic!( - "layout decided on a larger discriminant type ({min_ity:?}) than typeck ({typeck_ity:?})" - ); - // However, it is fine to make discr type however large (as an optimisation) - // after this point – we’ll just truncate the value we load in codegen. - } - - // Check to see if we should use a different type for the - // discriminant. We can safely use a type with the same size - // as the alignment of the first field of each variant. - // We increase the size of the discriminant to avoid LLVM copying - // padding when it doesn't need to. This normally causes unaligned - // load/stores and excessive memcpy/memset operations. By using a - // bigger integer size, LLVM can be sure about its contents and - // won't be so conservative. - - // Use the initial field alignment - let mut ity = if repr.c() || repr.int.is_some() { - min_ity - } else { - Integer::for_align(dl, start_align).unwrap_or(min_ity) - }; - - // If the alignment is not larger than the chosen discriminant size, - // don't use the alignment as the final size. - if ity <= min_ity { - ity = min_ity; - } else { - // Patch up the variants' first few fields. - let old_ity_size = min_ity.size(); - let new_ity_size = ity.size(); - for variant in &mut layout_variants { - match variant.fields { - FieldsShape::Arbitrary { ref mut offsets, .. } => { - for i in offsets { - if *i <= old_ity_size { - assert_eq!(*i, old_ity_size); - *i = new_ity_size; - } - } - // We might be making the struct larger. - if variant.size <= old_ity_size { - variant.size = new_ity_size; - } - } - FieldsShape::Primitive | FieldsShape::Array { .. } | FieldsShape::Union(..) => { - panic!("encountered a non-arbitrary layout during enum layout") - } - } - } - } - - let tag_mask = ity.size().unsigned_int_max(); - let tag = Scalar::Initialized { - value: Primitive::Int(ity, signed), - valid_range: WrappingRange { - start: (min as u128 & tag_mask), - end: (max as u128 & tag_mask), - }, - }; - let mut abi = Abi::Aggregate { sized: true }; - - if layout_variants.iter().all(|v| v.abi.is_uninhabited()) { - abi = Abi::Uninhabited; - } else if tag.size(dl) == size { - // Make sure we only use scalar layout when the enum is entirely its - // own tag (i.e. it has no padding nor any non-ZST variant fields). - abi = Abi::Scalar(tag); - } else { - // Try to use a ScalarPair for all tagged enums. - // That's possible only if we can find a common primitive type for all variants. - let mut common_prim = None; - let mut common_prim_initialized_in_all_variants = true; - for (field_layouts, layout_variant) in iter::zip(variants, &layout_variants) { - let FieldsShape::Arbitrary { ref offsets, .. } = layout_variant.fields else { - panic!("encountered a non-arbitrary layout during enum layout"); - }; - // We skip *all* ZST here and later check if we are good in terms of alignment. - // This lets us handle some cases involving aligned ZST. - let mut fields = iter::zip(field_layouts, offsets).filter(|p| !p.0.is_zst()); - let (field, offset) = match (fields.next(), fields.next()) { - (None, None) => { - common_prim_initialized_in_all_variants = false; - continue; - } - (Some(pair), None) => pair, - _ => { - common_prim = None; - break; - } - }; - let prim = match field.abi { - Abi::Scalar(scalar) => { - common_prim_initialized_in_all_variants &= - matches!(scalar, Scalar::Initialized { .. }); - scalar.primitive() - } - _ => { - common_prim = None; - break; - } - }; - if let Some(pair) = common_prim { - // This is pretty conservative. We could go fancier - // by conflating things like i32 and u32, or even - // realising that (u8, u8) could just cohabit with - // u16 or even u32. - if pair != (prim, offset) { - common_prim = None; - break; - } - } else { - common_prim = Some((prim, offset)); - } - } - if let Some((prim, offset)) = common_prim { - let prim_scalar = if common_prim_initialized_in_all_variants { - let size = prim.size(dl); - assert!(size.bits() <= 128); - Scalar::Initialized { value: prim, valid_range: WrappingRange::full(size) } - } else { - // Common prim might be uninit. - Scalar::Union { value: prim } - }; - let pair = self.scalar_pair::(tag, prim_scalar); - let pair_offsets = match pair.fields { - FieldsShape::Arbitrary { ref offsets, ref memory_index } => { - assert_eq!(memory_index.raw, [0, 1]); - offsets - } - _ => panic!("encountered a non-arbitrary layout during enum layout"), - }; - if pair_offsets[FieldIdx::new(0)] == Size::ZERO - && pair_offsets[FieldIdx::new(1)] == *offset - && align == pair.align - && size == pair.size - { - // We can use `ScalarPair` only when it matches our - // already computed layout (including `#[repr(C)]`). - abi = pair.abi; - } - } - } - - // If we pick a "clever" (by-value) ABI, we might have to adjust the ABI of the - // variants to ensure they are consistent. This is because a downcast is - // semantically a NOP, and thus should not affect layout. - if matches!(abi, Abi::Scalar(..) | Abi::ScalarPair(..)) { - for variant in &mut layout_variants { - // We only do this for variants with fields; the others are not accessed anyway. - // Also do not overwrite any already existing "clever" ABIs. - if variant.fields.count() > 0 && matches!(variant.abi, Abi::Aggregate { .. }) { - variant.abi = abi; - // Also need to bump up the size and alignment, so that the entire value fits - // in here. - variant.size = cmp::max(variant.size, size); - variant.align.abi = cmp::max(variant.align.abi, align.abi); - } - } - } - - let largest_niche = Niche::from_scalar(dl, Size::ZERO, tag); - - let tagged_layout = LayoutS { - variants: Variants::Multiple { - tag, - tag_encoding: TagEncoding::Direct, - tag_field: 0, - variants: IndexVec::new(), - }, - fields: FieldsShape::Arbitrary { - offsets: [Size::ZERO].into(), - memory_index: [0].into(), - }, - largest_niche, - abi, - align, - size, - max_repr_align, - unadjusted_abi_align, - }; - - let tagged_layout = TmpLayout { layout: tagged_layout, variants: layout_variants }; - - let mut best_layout = match (tagged_layout, niche_filling_layout) { - (tl, Some(nl)) => { - // Pick the smaller layout; otherwise, - // pick the layout with the larger niche; otherwise, - // pick tagged as it has simpler codegen. - use cmp::Ordering::*; - let niche_size = |tmp_l: &TmpLayout| { - tmp_l.layout.largest_niche.map_or(0, |n| n.available(dl)) - }; - match (tl.layout.size.cmp(&nl.layout.size), niche_size(&tl).cmp(&niche_size(&nl))) { - (Greater, _) => nl, - (Equal, Less) => nl, - _ => tl, - } - } - (tl, None) => tl, - }; - - // Now we can intern the variant layouts and store them in the enum layout. - best_layout.layout.variants = match best_layout.layout.variants { - Variants::Multiple { tag, tag_encoding, tag_field, .. } => { - Variants::Multiple { tag, tag_encoding, tag_field, variants: best_layout.variants } - } - Variants::Single { .. } => { - panic!("encountered a single-variant enum during multi-variant layout") - } - }; - Some(best_layout.layout) + layout_of_enum( + self, + repr, + variants, + discr_range_of_repr, + discriminants, + dont_niche_optimize_enum, + dl, + ) } fn layout_of_union< @@ -874,6 +407,493 @@ pub trait LayoutCalculator { } } +fn layout_of_enum<'a, LC, FieldIdx: Idx, VariantIdx: Idx, F>( + layout_calc: &LC, + repr: &ReprOptions, + variants: &IndexSlice>, + discr_range_of_repr: impl Fn(i128, i128) -> (Integer, bool), + discriminants: impl Iterator, + dont_niche_optimize_enum: bool, + dl: &TargetDataLayout, +) -> Option> +where + LC: LayoutCalculator + ?Sized, + F: Deref> + fmt::Debug, +{ + // Until we've decided whether to use the tagged or + // niche filling LayoutS, we don't want to intern the + // variant layouts, so we can't store them in the + // overall LayoutS. Store the overall LayoutS + // and the variant LayoutSs here until then. + struct TmpLayout { + layout: LayoutS, + variants: IndexVec>, + } + + let calculate_niche_filling_layout = || -> Option> { + if dont_niche_optimize_enum { + return None; + } + + if variants.len() < 2 { + return None; + } + + let mut align = dl.aggregate_align; + let mut max_repr_align = repr.align; + let mut unadjusted_abi_align = align.abi; + + let mut variant_layouts = variants + .iter_enumerated() + .map(|(j, v)| { + let mut st = layout_calc.univariant(dl, v, repr, StructKind::AlwaysSized)?; + st.variants = Variants::Single { index: j }; + + align = align.max(st.align); + max_repr_align = max_repr_align.max(st.max_repr_align); + unadjusted_abi_align = unadjusted_abi_align.max(st.unadjusted_abi_align); + + Some(st) + }) + .collect::>>()?; + + let largest_variant_index = variant_layouts + .iter_enumerated() + .max_by_key(|(_i, layout)| layout.size.bytes()) + .map(|(i, _layout)| i)?; + + let all_indices = variants.indices(); + let needs_disc = + |index: VariantIdx| index != largest_variant_index && !absent(&variants[index]); + let niche_variants = all_indices.clone().find(|v| needs_disc(*v)).unwrap() + ..=all_indices.rev().find(|v| needs_disc(*v)).unwrap(); + + let count = + (niche_variants.end().index() as u128 - niche_variants.start().index() as u128) + 1; + + // Find the field with the largest niche + let (field_index, niche, (niche_start, niche_scalar)) = variants[largest_variant_index] + .iter() + .enumerate() + .filter_map(|(j, field)| Some((j, field.largest_niche?))) + .max_by_key(|(_, niche)| niche.available(dl)) + .and_then(|(j, niche)| Some((j, niche, niche.reserve(dl, count)?)))?; + let niche_offset = + niche.offset + variant_layouts[largest_variant_index].fields.offset(field_index); + let niche_size = niche.value.size(dl); + let size = variant_layouts[largest_variant_index].size.align_to(align.abi); + + let all_variants_fit = variant_layouts.iter_enumerated_mut().all(|(i, layout)| { + if i == largest_variant_index { + return true; + } + + layout.largest_niche = None; + + if layout.size <= niche_offset { + // This variant will fit before the niche. + return true; + } + + // Determine if it'll fit after the niche. + let this_align = layout.align.abi; + let this_offset = (niche_offset + niche_size).align_to(this_align); + + if this_offset + layout.size > size { + return false; + } + + // It'll fit, but we need to make some adjustments. + match layout.fields { + FieldsShape::Arbitrary { ref mut offsets, .. } => { + for offset in offsets.iter_mut() { + *offset += this_offset; + } + } + FieldsShape::Primitive | FieldsShape::Array { .. } | FieldsShape::Union(..) => { + panic!("Layout of fields should be Arbitrary for variants") + } + } + + // It can't be a Scalar or ScalarPair because the offset isn't 0. + if !layout.abi.is_uninhabited() { + layout.abi = Abi::Aggregate { sized: true }; + } + layout.size += this_offset; + + true + }); + + if !all_variants_fit { + return None; + } + + let largest_niche = Niche::from_scalar(dl, niche_offset, niche_scalar); + + let others_zst = variant_layouts + .iter_enumerated() + .all(|(i, layout)| i == largest_variant_index || layout.size == Size::ZERO); + let same_size = size == variant_layouts[largest_variant_index].size; + let same_align = align == variant_layouts[largest_variant_index].align; + + let abi = if variant_layouts.iter().all(|v| v.abi.is_uninhabited()) { + Abi::Uninhabited + } else if same_size && same_align && others_zst { + match variant_layouts[largest_variant_index].abi { + // When the total alignment and size match, we can use the + // same ABI as the scalar variant with the reserved niche. + Abi::Scalar(_) => Abi::Scalar(niche_scalar), + Abi::ScalarPair(first, second) => { + // Only the niche is guaranteed to be initialised, + // so use union layouts for the other primitive. + if niche_offset == Size::ZERO { + Abi::ScalarPair(niche_scalar, second.to_union()) + } else { + Abi::ScalarPair(first.to_union(), niche_scalar) + } + } + _ => Abi::Aggregate { sized: true }, + } + } else { + Abi::Aggregate { sized: true } + }; + + let layout = LayoutS { + variants: Variants::Multiple { + tag: niche_scalar, + tag_encoding: TagEncoding::Niche { + untagged_variant: largest_variant_index, + niche_variants, + niche_start, + }, + tag_field: 0, + variants: IndexVec::new(), + }, + fields: FieldsShape::Arbitrary { + offsets: [niche_offset].into(), + memory_index: [0].into(), + }, + abi, + largest_niche, + size, + align, + max_repr_align, + unadjusted_abi_align, + }; + + Some(TmpLayout { layout, variants: variant_layouts }) + }; + + let niche_filling_layout = calculate_niche_filling_layout(); + + let (mut min, mut max) = (i128::MAX, i128::MIN); + let discr_type = repr.discr_type(); + let bits = Integer::from_attr(dl, discr_type).size().bits(); + for (i, mut val) in discriminants { + if variants[i].iter().any(|f| f.abi.is_uninhabited()) { + continue; + } + if discr_type.is_signed() { + // sign extend the raw representation to be an i128 + val = (val << (128 - bits)) >> (128 - bits); + } + if val < min { + min = val; + } + if val > max { + max = val; + } + } + // We might have no inhabited variants, so pretend there's at least one. + if (min, max) == (i128::MAX, i128::MIN) { + min = 0; + max = 0; + } + assert!(min <= max, "discriminant range is {min}...{max}"); + let (min_ity, signed) = discr_range_of_repr(min, max); //Integer::repr_discr(tcx, ty, &repr, min, max); + + let mut align = dl.aggregate_align; + let mut max_repr_align = repr.align; + let mut unadjusted_abi_align = align.abi; + + let mut size = Size::ZERO; + + // We're interested in the smallest alignment, so start large. + let mut start_align = Align::from_bytes(256).unwrap(); + assert_eq!(Integer::for_align(dl, start_align), None); + + // repr(C) on an enum tells us to make a (tag, union) layout, + // so we need to grow the prefix alignment to be at least + // the alignment of the union. (This value is used both for + // determining the alignment of the overall enum, and the + // determining the alignment of the payload after the tag.) + let mut prefix_align = min_ity.align(dl).abi; + if repr.c() { + for fields in variants { + for field in fields { + prefix_align = prefix_align.max(field.align.abi); + } + } + } + + // Create the set of structs that represent each variant. + let mut layout_variants = variants + .iter_enumerated() + .map(|(i, field_layouts)| { + let mut st = layout_calc.univariant( + dl, + field_layouts, + repr, + StructKind::Prefixed(min_ity.size(), prefix_align), + )?; + st.variants = Variants::Single { index: i }; + // Find the first field we can't move later + // to make room for a larger discriminant. + for field_idx in st.fields.index_by_increasing_offset() { + let field = &field_layouts[FieldIdx::new(field_idx)]; + if !field.is_1zst() { + start_align = start_align.min(field.align.abi); + break; + } + } + size = cmp::max(size, st.size); + align = align.max(st.align); + max_repr_align = max_repr_align.max(st.max_repr_align); + unadjusted_abi_align = unadjusted_abi_align.max(st.unadjusted_abi_align); + Some(st) + }) + .collect::>>()?; + + // Align the maximum variant size to the largest alignment. + size = size.align_to(align.abi); + + // FIXME(oli-obk): deduplicate and harden these checks + if size.bytes() >= dl.obj_size_bound() { + return None; + } + + let typeck_ity = Integer::from_attr(dl, repr.discr_type()); + if typeck_ity < min_ity { + // It is a bug if Layout decided on a greater discriminant size than typeck for + // some reason at this point (based on values discriminant can take on). Mostly + // because this discriminant will be loaded, and then stored into variable of + // type calculated by typeck. Consider such case (a bug): typeck decided on + // byte-sized discriminant, but layout thinks we need a 16-bit to store all + // discriminant values. That would be a bug, because then, in codegen, in order + // to store this 16-bit discriminant into 8-bit sized temporary some of the + // space necessary to represent would have to be discarded (or layout is wrong + // on thinking it needs 16 bits) + panic!( + "layout decided on a larger discriminant type ({min_ity:?}) than typeck ({typeck_ity:?})" + ); + // However, it is fine to make discr type however large (as an optimisation) + // after this point – we’ll just truncate the value we load in codegen. + } + + // Check to see if we should use a different type for the + // discriminant. We can safely use a type with the same size + // as the alignment of the first field of each variant. + // We increase the size of the discriminant to avoid LLVM copying + // padding when it doesn't need to. This normally causes unaligned + // load/stores and excessive memcpy/memset operations. By using a + // bigger integer size, LLVM can be sure about its contents and + // won't be so conservative. + + // Use the initial field alignment + let mut ity = if repr.c() || repr.int.is_some() { + min_ity + } else { + Integer::for_align(dl, start_align).unwrap_or(min_ity) + }; + + // If the alignment is not larger than the chosen discriminant size, + // don't use the alignment as the final size. + if ity <= min_ity { + ity = min_ity; + } else { + // Patch up the variants' first few fields. + let old_ity_size = min_ity.size(); + let new_ity_size = ity.size(); + for variant in &mut layout_variants { + match variant.fields { + FieldsShape::Arbitrary { ref mut offsets, .. } => { + for i in offsets { + if *i <= old_ity_size { + assert_eq!(*i, old_ity_size); + *i = new_ity_size; + } + } + // We might be making the struct larger. + if variant.size <= old_ity_size { + variant.size = new_ity_size; + } + } + FieldsShape::Primitive | FieldsShape::Array { .. } | FieldsShape::Union(..) => { + panic!("encountered a non-arbitrary layout during enum layout") + } + } + } + } + + let tag_mask = ity.size().unsigned_int_max(); + let tag = Scalar::Initialized { + value: Primitive::Int(ity, signed), + valid_range: WrappingRange { + start: (min as u128 & tag_mask), + end: (max as u128 & tag_mask), + }, + }; + let mut abi = Abi::Aggregate { sized: true }; + + if layout_variants.iter().all(|v| v.abi.is_uninhabited()) { + abi = Abi::Uninhabited; + } else if tag.size(dl) == size { + // Make sure we only use scalar layout when the enum is entirely its + // own tag (i.e. it has no padding nor any non-ZST variant fields). + abi = Abi::Scalar(tag); + } else { + // Try to use a ScalarPair for all tagged enums. + // That's possible only if we can find a common primitive type for all variants. + let mut common_prim = None; + let mut common_prim_initialized_in_all_variants = true; + for (field_layouts, layout_variant) in iter::zip(variants, &layout_variants) { + let FieldsShape::Arbitrary { ref offsets, .. } = layout_variant.fields else { + panic!("encountered a non-arbitrary layout during enum layout"); + }; + // We skip *all* ZST here and later check if we are good in terms of alignment. + // This lets us handle some cases involving aligned ZST. + let mut fields = iter::zip(field_layouts, offsets).filter(|p| !p.0.is_zst()); + let (field, offset) = match (fields.next(), fields.next()) { + (None, None) => { + common_prim_initialized_in_all_variants = false; + continue; + } + (Some(pair), None) => pair, + _ => { + common_prim = None; + break; + } + }; + let prim = match field.abi { + Abi::Scalar(scalar) => { + common_prim_initialized_in_all_variants &= + matches!(scalar, Scalar::Initialized { .. }); + scalar.primitive() + } + _ => { + common_prim = None; + break; + } + }; + if let Some(pair) = common_prim { + // This is pretty conservative. We could go fancier + // by conflating things like i32 and u32, or even + // realising that (u8, u8) could just cohabit with + // u16 or even u32. + if pair != (prim, offset) { + common_prim = None; + break; + } + } else { + common_prim = Some((prim, offset)); + } + } + if let Some((prim, offset)) = common_prim { + let prim_scalar = if common_prim_initialized_in_all_variants { + let size = prim.size(dl); + assert!(size.bits() <= 128); + Scalar::Initialized { value: prim, valid_range: WrappingRange::full(size) } + } else { + // Common prim might be uninit. + Scalar::Union { value: prim } + }; + let pair = layout_calc.scalar_pair::(tag, prim_scalar); + let pair_offsets = match pair.fields { + FieldsShape::Arbitrary { ref offsets, ref memory_index } => { + assert_eq!(memory_index.raw, [0, 1]); + offsets + } + _ => panic!("encountered a non-arbitrary layout during enum layout"), + }; + if pair_offsets[FieldIdx::new(0)] == Size::ZERO + && pair_offsets[FieldIdx::new(1)] == *offset + && align == pair.align + && size == pair.size + { + // We can use `ScalarPair` only when it matches our + // already computed layout (including `#[repr(C)]`). + abi = pair.abi; + } + } + } + + // If we pick a "clever" (by-value) ABI, we might have to adjust the ABI of the + // variants to ensure they are consistent. This is because a downcast is + // semantically a NOP, and thus should not affect layout. + if matches!(abi, Abi::Scalar(..) | Abi::ScalarPair(..)) { + for variant in &mut layout_variants { + // We only do this for variants with fields; the others are not accessed anyway. + // Also do not overwrite any already existing "clever" ABIs. + if variant.fields.count() > 0 && matches!(variant.abi, Abi::Aggregate { .. }) { + variant.abi = abi; + // Also need to bump up the size and alignment, so that the entire value fits + // in here. + variant.size = cmp::max(variant.size, size); + variant.align.abi = cmp::max(variant.align.abi, align.abi); + } + } + } + + let largest_niche = Niche::from_scalar(dl, Size::ZERO, tag); + + let tagged_layout = LayoutS { + variants: Variants::Multiple { + tag, + tag_encoding: TagEncoding::Direct, + tag_field: 0, + variants: IndexVec::new(), + }, + fields: FieldsShape::Arbitrary { offsets: [Size::ZERO].into(), memory_index: [0].into() }, + largest_niche, + abi, + align, + size, + max_repr_align, + unadjusted_abi_align, + }; + + let tagged_layout = TmpLayout { layout: tagged_layout, variants: layout_variants }; + + let mut best_layout = match (tagged_layout, niche_filling_layout) { + (tl, Some(nl)) => { + // Pick the smaller layout; otherwise, + // pick the layout with the larger niche; otherwise, + // pick tagged as it has simpler codegen. + use cmp::Ordering::*; + let niche_size = |tmp_l: &TmpLayout| { + tmp_l.layout.largest_niche.map_or(0, |n| n.available(dl)) + }; + match (tl.layout.size.cmp(&nl.layout.size), niche_size(&tl).cmp(&niche_size(&nl))) { + (Greater, _) => nl, + (Equal, Less) => nl, + _ => tl, + } + } + (tl, None) => tl, + }; + + // Now we can intern the variant layouts and store them in the enum layout. + best_layout.layout.variants = match best_layout.layout.variants { + Variants::Multiple { tag, tag_encoding, tag_field, .. } => { + Variants::Multiple { tag, tag_encoding, tag_field, variants: best_layout.variants } + } + Variants::Single { .. } => { + panic!("encountered a single-variant enum during multi-variant layout") + } + }; + Some(best_layout.layout) +} + /// Determines towards which end of a struct layout optimizations will try to place the best niches. enum NicheBias { Start, From 8facfb23c6f6c8f338b80d21b0f01223684f0e44 Mon Sep 17 00:00:00 2001 From: Jubilee Young Date: Sat, 16 Dec 2023 01:56:55 -0800 Subject: [PATCH 4/4] Extract fn layout_of_struct --- compiler/rustc_abi/src/layout.rs | 232 +++++++++++++++++-------------- 1 file changed, 130 insertions(+), 102 deletions(-) diff --git a/compiler/rustc_abi/src/layout.rs b/compiler/rustc_abi/src/layout.rs index 102bef9f6de37..ec3ab828b7194 100644 --- a/compiler/rustc_abi/src/layout.rs +++ b/compiler/rustc_abi/src/layout.rs @@ -197,109 +197,37 @@ pub trait LayoutCalculator { None => VariantIdx::new(0), }; - let is_struct = !is_enum || - // Only one variant is present. - (present_second.is_none() && - // Representation optimizations are allowed. - !repr.inhibit_enum_layout_opt()); - if is_struct { - // Struct, or univariant enum equivalent to a struct. - // (Typechecking will reject discriminant-sizing attrs.) - - let v = present_first; - let kind = if is_enum || variants[v].is_empty() || always_sized { - StructKind::AlwaysSized - } else { - StructKind::MaybeUnsized - }; - - let mut st = self.univariant(dl, &variants[v], repr, kind)?; - st.variants = Variants::Single { index: v }; - - if is_unsafe_cell { - let hide_niches = |scalar: &mut _| match scalar { - Scalar::Initialized { value, valid_range } => { - *valid_range = WrappingRange::full(value.size(dl)) - } - // Already doesn't have any niches - Scalar::Union { .. } => {} - }; - match &mut st.abi { - Abi::Uninhabited => {} - Abi::Scalar(scalar) => hide_niches(scalar), - Abi::ScalarPair(a, b) => { - hide_niches(a); - hide_niches(b); - } - Abi::Vector { element, count: _ } => hide_niches(element), - Abi::Aggregate { sized: _ } => {} - } - st.largest_niche = None; - return Some(st); - } - - let (start, end) = scalar_valid_range; - match st.abi { - Abi::Scalar(ref mut scalar) | Abi::ScalarPair(ref mut scalar, _) => { - // Enlarging validity ranges would result in missed - // optimizations, *not* wrongly assuming the inner - // value is valid. e.g. unions already enlarge validity ranges, - // because the values may be uninitialized. - // - // Because of that we only check that the start and end - // of the range is representable with this scalar type. - - let max_value = scalar.size(dl).unsigned_int_max(); - if let Bound::Included(start) = start { - // FIXME(eddyb) this might be incorrect - it doesn't - // account for wrap-around (end < start) ranges. - assert!(start <= max_value, "{start} > {max_value}"); - scalar.valid_range_mut().start = start; - } - if let Bound::Included(end) = end { - // FIXME(eddyb) this might be incorrect - it doesn't - // account for wrap-around (end < start) ranges. - assert!(end <= max_value, "{end} > {max_value}"); - scalar.valid_range_mut().end = end; - } - - // Update `largest_niche` if we have introduced a larger niche. - let niche = Niche::from_scalar(dl, Size::ZERO, *scalar); - if let Some(niche) = niche { - match st.largest_niche { - Some(largest_niche) => { - // Replace the existing niche even if they're equal, - // because this one is at a lower offset. - if largest_niche.available(dl) <= niche.available(dl) { - st.largest_niche = Some(niche); - } - } - None => st.largest_niche = Some(niche), - } - } - } - _ => assert!( - start == Bound::Unbounded && end == Bound::Unbounded, - "nonscalar layout for layout_scalar_valid_range type: {st:#?}", - ), - } - - return Some(st); + // take the struct path if it is an actual struct + if !is_enum || + // or for optimizing univariant enums + (present_second.is_none() && !repr.inhibit_enum_layout_opt()) + { + layout_of_struct( + self, + repr, + variants, + is_enum, + is_unsafe_cell, + scalar_valid_range, + always_sized, + dl, + present_first, + ) + } else { + // At this point, we have handled all unions and + // structs. (We have also handled univariant enums + // that allow representation optimization.) + assert!(is_enum); + layout_of_enum( + self, + repr, + variants, + discr_range_of_repr, + discriminants, + dont_niche_optimize_enum, + dl, + ) } - - // At this point, we have handled all unions and - // structs. (We have also handled univariant enums - // that allow representation optimization.) - assert!(is_enum); - layout_of_enum( - self, - repr, - variants, - discr_range_of_repr, - discriminants, - dont_niche_optimize_enum, - dl, - ) } fn layout_of_union< @@ -407,6 +335,106 @@ pub trait LayoutCalculator { } } +/// single-variant enums are just structs, if you think about it +fn layout_of_struct<'a, LC, FieldIdx: Idx, VariantIdx: Idx, F>( + layout_calc: &LC, + repr: &ReprOptions, + variants: &IndexSlice>, + is_enum: bool, + is_unsafe_cell: bool, + scalar_valid_range: (Bound, Bound), + always_sized: bool, + dl: &TargetDataLayout, + present_first: VariantIdx, +) -> Option> +where + LC: LayoutCalculator + ?Sized, + F: Deref> + fmt::Debug, +{ + // Struct, or univariant enum equivalent to a struct. + // (Typechecking will reject discriminant-sizing attrs.) + + let v = present_first; + let kind = if is_enum || variants[v].is_empty() || always_sized { + StructKind::AlwaysSized + } else { + StructKind::MaybeUnsized + }; + + let mut st = layout_calc.univariant(dl, &variants[v], repr, kind)?; + st.variants = Variants::Single { index: v }; + + if is_unsafe_cell { + let hide_niches = |scalar: &mut _| match scalar { + Scalar::Initialized { value, valid_range } => { + *valid_range = WrappingRange::full(value.size(dl)) + } + // Already doesn't have any niches + Scalar::Union { .. } => {} + }; + match &mut st.abi { + Abi::Uninhabited => {} + Abi::Scalar(scalar) => hide_niches(scalar), + Abi::ScalarPair(a, b) => { + hide_niches(a); + hide_niches(b); + } + Abi::Vector { element, count: _ } => hide_niches(element), + Abi::Aggregate { sized: _ } => {} + } + st.largest_niche = None; + return Some(st); + } + + let (start, end) = scalar_valid_range; + match st.abi { + Abi::Scalar(ref mut scalar) | Abi::ScalarPair(ref mut scalar, _) => { + // Enlarging validity ranges would result in missed + // optimizations, *not* wrongly assuming the inner + // value is valid. e.g. unions already enlarge validity ranges, + // because the values may be uninitialized. + // + // Because of that we only check that the start and end + // of the range is representable with this scalar type. + + let max_value = scalar.size(dl).unsigned_int_max(); + if let Bound::Included(start) = start { + // FIXME(eddyb) this might be incorrect - it doesn't + // account for wrap-around (end < start) ranges. + assert!(start <= max_value, "{start} > {max_value}"); + scalar.valid_range_mut().start = start; + } + if let Bound::Included(end) = end { + // FIXME(eddyb) this might be incorrect - it doesn't + // account for wrap-around (end < start) ranges. + assert!(end <= max_value, "{end} > {max_value}"); + scalar.valid_range_mut().end = end; + } + + // Update `largest_niche` if we have introduced a larger niche. + let niche = Niche::from_scalar(dl, Size::ZERO, *scalar); + if let Some(niche) = niche { + match st.largest_niche { + Some(largest_niche) => { + // Replace the existing niche even if they're equal, + // because this one is at a lower offset. + if largest_niche.available(dl) <= niche.available(dl) { + st.largest_niche = Some(niche); + } + } + None => st.largest_niche = Some(niche), + } + } + } + _ => assert!( + start == Bound::Unbounded && end == Bound::Unbounded, + "nonscalar layout for layout_scalar_valid_range type: {st:#?}", + ), + } + + Some(st) +} + fn layout_of_enum<'a, LC, FieldIdx: Idx, VariantIdx: Idx, F>( layout_calc: &LC, repr: &ReprOptions,