Auto merge of #96591 - b-naber:transition-to-valtrees-in-type-system, r=lcnr

Use valtrees as the type-system representation for constant values

This is not quite ready yet, there are still some problems with pretty printing and symbol mangling and `deref_const` seems to not work correctly in all cases.

Mainly opening now for a perf-run (which should be good to go, despite the still existing problems).

r? `@oli-obk`

cc `@lcnr` `@RalfJung`

Author: bors

Diff for: compiler/rustc_codegen_cranelift/src/constant.rs

@@ -127,7 +127,7 @@ pub(crate) fn codegen_constant<'tcx>(
         ConstantKind::Val(val, ty) => return codegen_const_value(fx, val, ty),
     };
     let const_val = match const_.kind() {
-        ConstKind::Value(const_val) => const_val,
+        ConstKind::Value(valtree) => fx.tcx.valtree_to_const_val((const_.ty(), valtree)),
         ConstKind::Unevaluated(ty::Unevaluated { def, substs, promoted })
             if fx.tcx.is_static(def.did) =>
         {
@@ -468,9 +468,10 @@ pub(crate) fn mir_operand_get_const_val<'tcx>(
 ) -> Option<ConstValue<'tcx>> {
     match operand {
         Operand::Constant(const_) => match const_.literal {
-            ConstantKind::Ty(const_) => {
-                fx.monomorphize(const_).eval(fx.tcx, ParamEnv::reveal_all()).kind().try_to_value()
-            }
+            ConstantKind::Ty(const_) => fx
+                .monomorphize(const_)
+                .eval_for_mir(fx.tcx, ParamEnv::reveal_all())
+                .try_to_value(fx.tcx),
             ConstantKind::Val(val, _) => Some(val),
         },
         // FIXME(rust-lang/rust#85105): Casts like `IMM8 as u32` result in the const being stored

Diff for: compiler/rustc_codegen_ssa/src/debuginfo/type_names.rs

@@ -703,15 +703,19 @@ fn push_const_param<'tcx>(tcx: TyCtxt<'tcx>, ct: ty::Const<'tcx>, output: &mut S
                 // but we get a deterministic, virtually unique value for the constant.
                 let hcx = &mut tcx.create_stable_hashing_context();
                 let mut hasher = StableHasher::new();
-                hcx.while_hashing_spans(false, |hcx| ct.kind().hash_stable(hcx, &mut hasher));
+                let ct = ct.eval(tcx, ty::ParamEnv::reveal_all());
+                hcx.while_hashing_spans(false, |hcx| ct.to_valtree().hash_stable(hcx, &mut hasher));
                 // Let's only emit 64 bits of the hash value. That should be plenty for
                 // avoiding collisions and will make the emitted type names shorter.
-                let hash: u64 = hasher.finish();
+                // Note: Don't use `StableHashResult` impl of `u64` here directly, since that
+                // would lead to endianness problems.
+                let hash: u128 = hasher.finish();
+                let hash_short = (hash.to_le() as u64).to_le();
 
                 if cpp_like_debuginfo(tcx) {
-                    write!(output, "CONST${:x}", hash)
+                    write!(output, "CONST${:x}", hash_short)
                 } else {
-                    write!(output, "{{CONST#{:x}}}", hash)
+                    write!(output, "{{CONST#{:x}}}", hash_short)
                 }
             }
         },

Diff for: compiler/rustc_codegen_ssa/src/mir/constant.rs

@@ -38,7 +38,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                     self.cx.tcx().sess.span_err(constant.span, "erroneous constant encountered");
                     err
                 }),
-            ty::ConstKind::Value(value) => Ok(value),
+            ty::ConstKind::Value(val) => Ok(self.cx.tcx().valtree_to_const_val((ct.ty(), val))),
             err => span_bug!(
                 constant.span,
                 "encountered bad ConstKind after monomorphizing: {:?}",
@@ -58,14 +58,14 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         constant
             .map(|val| {
                 let field_ty = ty.builtin_index().unwrap();
-                let c = ty::Const::from_value(bx.tcx(), val, ty);
+                let c = mir::ConstantKind::from_value(val, ty);
                 let values: Vec<_> = bx
                     .tcx()
-                    .destructure_const(ty::ParamEnv::reveal_all().and(c))
+                    .destructure_mir_constant(ty::ParamEnv::reveal_all(), c)
                     .fields
                     .iter()
                     .map(|field| {
-                        if let Some(prim) = field.kind().try_to_scalar() {
+                        if let Some(prim) = field.try_to_scalar() {
                             let layout = bx.layout_of(field_ty);
                             let Abi::Scalar(scalar) = layout.abi else {
                                 bug!("from_const: invalid ByVal layout: {:#?}", layout);

Diff for: compiler/rustc_const_eval/src/const_eval/error.rs

@@ -140,6 +140,7 @@ impl<'tcx> ConstEvalErr<'tcx> {
     ///
     /// If `lint_root.is_some()` report it as a lint, else report it as a hard error.
     /// (Except that for some errors, we ignore all that -- see `must_error` below.)
+    #[instrument(skip(self, tcx, decorate, lint_root), level = "debug")]
     fn struct_generic(
         &self,
         tcx: TyCtxtAt<'tcx>,
@@ -190,6 +191,7 @@ impl<'tcx> ConstEvalErr<'tcx> {
             decorate(err);
         };
 
+        debug!("self.error: {:?}", self.error);
         // Special handling for certain errors
         match &self.error {
             // Don't emit a new diagnostic for these errors

Diff for: compiler/rustc_const_eval/src/const_eval/eval_queries.rs

@@ -196,7 +196,7 @@ pub(super) fn op_to_const<'tcx>(
 }
 
 #[instrument(skip(tcx), level = "debug")]
-fn turn_into_const_value<'tcx>(
+pub(crate) fn turn_into_const_value<'tcx>(
     tcx: TyCtxt<'tcx>,
     constant: ConstAlloc<'tcx>,
     key: ty::ParamEnvAnd<'tcx, GlobalId<'tcx>>,
@@ -222,6 +222,7 @@ fn turn_into_const_value<'tcx>(
     const_val
 }
 
+#[instrument(skip(tcx), level = "debug")]
 pub fn eval_to_const_value_raw_provider<'tcx>(
     tcx: TyCtxt<'tcx>,
     key: ty::ParamEnvAnd<'tcx, GlobalId<'tcx>>,
@@ -256,6 +257,7 @@ pub fn eval_to_const_value_raw_provider<'tcx>(
     tcx.eval_to_allocation_raw(key).map(|val| turn_into_const_value(tcx, val, key))
 }
 
+#[instrument(skip(tcx), level = "debug")]
 pub fn eval_to_allocation_raw_provider<'tcx>(
     tcx: TyCtxt<'tcx>,
     key: ty::ParamEnvAnd<'tcx, GlobalId<'tcx>>,

Diff for: compiler/rustc_const_eval/src/const_eval/mod.rs

@@ -1,12 +1,11 @@
 // Not in interpret to make sure we do not use private implementation details
 
-use std::convert::TryFrom;
-
 use rustc_hir::Mutability;
 use rustc_middle::mir;
 use rustc_middle::mir::interpret::{EvalToValTreeResult, GlobalId};
 use rustc_middle::ty::{self, TyCtxt};
 use rustc_span::{source_map::DUMMY_SP, symbol::Symbol};
+use rustc_target::abi::VariantIdx;
 
 use crate::interpret::{
     intern_const_alloc_recursive, ConstValue, InternKind, InterpCx, InterpResult, MemPlaceMeta,
@@ -40,7 +39,7 @@ pub(crate) fn const_caller_location(
 }
 
 // We forbid type-level constants that contain more than `VALTREE_MAX_NODES` nodes.
-const VALTREE_MAX_NODES: usize = 1000;
+const VALTREE_MAX_NODES: usize = 100000;
 
 pub(crate) enum ValTreeCreationError {
     NodesOverflow,
@@ -56,6 +55,8 @@ pub(crate) fn eval_to_valtree<'tcx>(
     cid: GlobalId<'tcx>,
 ) -> EvalToValTreeResult<'tcx> {
     let const_alloc = tcx.eval_to_allocation_raw(param_env.and(cid))?;
+
+    // FIXME Need to provide a span to `eval_to_valtree`
     let ecx = mk_eval_cx(
         tcx, DUMMY_SP, param_env,
         // It is absolutely crucial for soundness that
@@ -90,40 +91,81 @@ pub(crate) fn eval_to_valtree<'tcx>(
     }
 }
 
-/// This function should never fail for validated constants. However, it is also invoked from the
-/// pretty printer which might attempt to format invalid constants and in that case it might fail.
+/// Tries to destructure constants of type Array or Adt into the constants
+/// of its fields.
 pub(crate) fn try_destructure_const<'tcx>(
     tcx: TyCtxt<'tcx>,
-    param_env: ty::ParamEnv<'tcx>,
-    val: ty::Const<'tcx>,
-) -> InterpResult<'tcx, mir::DestructuredConst<'tcx>> {
-    trace!("destructure_const: {:?}", val);
-    let ecx = mk_eval_cx(tcx, DUMMY_SP, param_env, false);
-    let op = ecx.const_to_op(val, None)?;
-    // We go to `usize` as we cannot allocate anything bigger anyway.
-    let (field_count, variant, down) = match val.ty().kind() {
-        ty::Array(_, len) => (usize::try_from(len.eval_usize(tcx, param_env)).unwrap(), None, op),
-        // Checks if we have any variants, to avoid downcasting to a non-existing variant (when
-        // there are no variants `read_discriminant` successfully returns a non-existing variant
-        // index).
-        ty::Adt(def, _) if def.variants().is_empty() => throw_ub!(Unreachable),
-        ty::Adt(def, _) => {
-            let variant = ecx.read_discriminant(&op)?.1;
-            let down = ecx.operand_downcast(&op, variant)?;
-            (def.variant(variant).fields.len(), Some(variant), down)
-        }
-        ty::Tuple(substs) => (substs.len(), None, op),
-        _ => bug!("cannot destructure constant {:?}", val),
-    };
-    let fields = (0..field_count)
-        .map(|i| {
-            let field_op = ecx.operand_field(&down, i)?;
-            let val = op_to_const(&ecx, &field_op);
-            Ok(ty::Const::from_value(tcx, val, field_op.layout.ty))
-        })
-        .collect::<InterpResult<'tcx, Vec<_>>>()?;
-    let fields = tcx.arena.alloc_from_iter(fields);
-    Ok(mir::DestructuredConst { variant, fields })
+    const_: ty::Const<'tcx>,
+) -> Option<ty::DestructuredConst<'tcx>> {
+    if let ty::ConstKind::Value(valtree) = const_.kind() {
+        let branches = match valtree {
+            ty::ValTree::Branch(b) => b,
+            _ => return None,
+        };
+
+        let (fields, variant) = match const_.ty().kind() {
+            ty::Array(inner_ty, _) | ty::Slice(inner_ty) => {
+                // construct the consts for the elements of the array/slice
+                let field_consts = branches
+                    .iter()
+                    .map(|b| {
+                        tcx.mk_const(ty::ConstS { kind: ty::ConstKind::Value(*b), ty: *inner_ty })
+                    })
+                    .collect::<Vec<_>>();
+                debug!(?field_consts);
+
+                (field_consts, None)
+            }
+            ty::Adt(def, _) if def.variants().is_empty() => bug!("unreachable"),
+            ty::Adt(def, substs) => {
+                let variant_idx = if def.is_enum() {
+                    VariantIdx::from_u32(branches[0].unwrap_leaf().try_to_u32().ok()?)
+                } else {
+                    VariantIdx::from_u32(0)
+                };
+                let fields = &def.variant(variant_idx).fields;
+                let mut field_consts = Vec::with_capacity(fields.len());
+
+                // Note: First element inValTree corresponds to variant of enum
+                let mut valtree_idx = if def.is_enum() { 1 } else { 0 };
+                for field in fields {
+                    let field_ty = field.ty(tcx, substs);
+                    let field_valtree = branches[valtree_idx]; // first element of branches is variant
+                    let field_const = tcx.mk_const(ty::ConstS {
+                        kind: ty::ConstKind::Value(field_valtree),
+                        ty: field_ty,
+                    });
+                    field_consts.push(field_const);
+                    valtree_idx += 1;
+                }
+                debug!(?field_consts);
+
+                (field_consts, Some(variant_idx))
+            }
+            ty::Tuple(elem_tys) => {
+                let fields = elem_tys
+                    .iter()
+                    .enumerate()
+                    .map(|(i, elem_ty)| {
+                        let elem_valtree = branches[i];
+                        tcx.mk_const(ty::ConstS {
+                            kind: ty::ConstKind::Value(elem_valtree),
+                            ty: elem_ty,
+                        })
+                    })
+                    .collect::<Vec<_>>();
+
+                (fields, None)
+            }
+            _ => bug!("cannot destructure constant {:?}", const_),
+        };
+
+        let fields = tcx.arena.alloc_from_iter(fields.into_iter());
+
+        Some(ty::DestructuredConst { variant, fields })
+    } else {
+        None
+    }
 }
 
 #[instrument(skip(tcx), level = "debug")]
@@ -143,8 +185,8 @@ pub(crate) fn try_destructure_mir_constant<'tcx>(
             throw_ub!(Unreachable)
         }
         ty::Adt(def, _) => {
-            let variant = ecx.read_discriminant(&op).unwrap().1;
-            let down = ecx.operand_downcast(&op, variant).unwrap();
+            let variant = ecx.read_discriminant(&op)?.1;
+            let down = ecx.operand_downcast(&op, variant)?;
             (def.variants()[variant].fields.len(), Some(variant), down)
         }
         ty::Tuple(substs) => (substs.len(), None, op),
@@ -163,43 +205,6 @@ pub(crate) fn try_destructure_mir_constant<'tcx>(
     Ok(mir::DestructuredMirConstant { variant, fields })
 }
 
-#[instrument(skip(tcx), level = "debug")]
-pub(crate) fn deref_const<'tcx>(
-    tcx: TyCtxt<'tcx>,
-    param_env: ty::ParamEnv<'tcx>,
-    val: ty::Const<'tcx>,
-) -> ty::Const<'tcx> {
-    trace!("deref_const: {:?}", val);
-    let ecx = mk_eval_cx(tcx, DUMMY_SP, param_env, false);
-    let op = ecx.const_to_op(val, None).unwrap();
-    let mplace = ecx.deref_operand(&op).unwrap();
-    if let Some(alloc_id) = mplace.ptr.provenance {
-        assert_eq!(
-            tcx.get_global_alloc(alloc_id).unwrap().unwrap_memory().inner().mutability,
-            Mutability::Not,
-            "deref_const cannot be used with mutable allocations as \
-            that could allow pattern matching to observe mutable statics",
-        );
-    }
-
-    let ty = match mplace.meta {
-        MemPlaceMeta::None => mplace.layout.ty,
-        MemPlaceMeta::Poison => bug!("poison metadata in `deref_const`: {:#?}", mplace),
-        // In case of unsized types, figure out the real type behind.
-        MemPlaceMeta::Meta(scalar) => match mplace.layout.ty.kind() {
-            ty::Str => bug!("there's no sized equivalent of a `str`"),
-            ty::Slice(elem_ty) => tcx.mk_array(*elem_ty, scalar.to_machine_usize(&tcx).unwrap()),
-            _ => bug!(
-                "type {} should not have metadata, but had {:?}",
-                mplace.layout.ty,
-                mplace.meta
-            ),
-        },
-    };
-
-    tcx.mk_const(ty::ConstS { kind: ty::ConstKind::Value(op_to_const(&ecx, &mplace.into())), ty })
-}
-
 #[instrument(skip(tcx), level = "debug")]
 pub(crate) fn deref_mir_constant<'tcx>(
     tcx: TyCtxt<'tcx>,
@@ -213,14 +218,14 @@ pub(crate) fn deref_mir_constant<'tcx>(
         assert_eq!(
             tcx.get_global_alloc(alloc_id).unwrap().unwrap_memory().0.0.mutability,
             Mutability::Not,
-            "deref_const cannot be used with mutable allocations as \
+            "deref_mir_constant cannot be used with mutable allocations as \
             that could allow pattern matching to observe mutable statics",
         );
     }
 
     let ty = match mplace.meta {
         MemPlaceMeta::None => mplace.layout.ty,
-        MemPlaceMeta::Poison => bug!("poison metadata in `deref_const`: {:#?}", mplace),
+        MemPlaceMeta::Poison => bug!("poison metadata in `deref_mir_constant`: {:#?}", mplace),
         // In case of unsized types, figure out the real type behind.
         MemPlaceMeta::Meta(scalar) => match mplace.layout.ty.kind() {
             ty::Str => bug!("there's no sized equivalent of a `str`"),