Allow transmutes to produce OperandValues instead of always using allocas

LLVM can usually optimize these away, but especially for things like transmutes of newtypes it's silly to generate the `alloc`+`store`+`load` at all when it's actually a nop at LLVM level.

Author: scottmcm

compiler/rustc_codegen_ssa/src/mir/operand.rs

@@ -23,10 +23,26 @@ pub enum OperandValue<V> {
     /// to be valid for the operand's lifetime.
     /// The second value, if any, is the extra data (vtable or length)
     /// which indicates that it refers to an unsized rvalue.
+    ///
+    /// An `OperandValue` has this variant for types which are neither
+    /// `Immediate` nor `Pair`s. The backend value in this variant must be a
+    /// pointer to the *non*-immediate backend type. That pointee type is the
+    /// one returned by [`LayoutTypeMethods::backend_type`].
     Ref(V, Option<V>, Align),
-    /// A single LLVM value.
+    /// A single LLVM immediate value.
+    ///
+    /// An `OperandValue` *must* be this variant for any type for which
+    /// [`LayoutTypeMethods::is_backend_immediate`] returns `true`.
+    /// The backend value in this variant must be the *immediate* backend type,
+    /// as returned by [`LayoutTypeMethods::immediate_backend_type`].
     Immediate(V),
     /// A pair of immediate LLVM values. Used by fat pointers too.
+    ///
+    /// An `OperandValue` *must* be this variant for any type for which
+    /// [`LayoutTypeMethods::is_backend_scalar_pair`] returns `true`.
+    /// The backend values in this variant must be the *immediate* backend types,
+    /// as returned by [`LayoutTypeMethods::scalar_pair_element_backend_type`]
+    /// with `immediate: true`.
     Pair(V, V),
 }
 

compiler/rustc_codegen_ssa/src/mir/rvalue.rs

@@ -10,7 +10,7 @@ use crate::MemFlags;
 use rustc_middle::mir;
 use rustc_middle::mir::Operand;
 use rustc_middle::ty::cast::{CastTy, IntTy};
-use rustc_middle::ty::layout::{HasTyCtxt, LayoutOf};
+use rustc_middle::ty::layout::{HasTyCtxt, LayoutOf, TyAndLayout};
 use rustc_middle::ty::{self, adjustment::PointerCast, Instance, Ty, TyCtxt};
 use rustc_span::source_map::{Span, DUMMY_SP};
 use rustc_target::abi::{self, FIRST_VARIANT};
@@ -159,8 +159,8 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         debug_assert!(dst.layout.is_sized());
 
         if src.layout.size != dst.layout.size
-            || src.layout.abi == abi::Abi::Uninhabited
-            || dst.layout.abi == abi::Abi::Uninhabited
+            || src.layout.abi.is_uninhabited()
+            || dst.layout.abi.is_uninhabited()
         {
             // In all of these cases it's UB to run this transmute, but that's
             // known statically so might as well trap for it, rather than just
@@ -169,22 +169,20 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             return;
         }
 
-        let size_in_bytes = src.layout.size.bytes();
-        if size_in_bytes == 0 {
-            // Nothing to write
+        if let Some(val) = self.codegen_transmute_operand(bx, src, dst.layout) {
+            val.store(bx, dst);
             return;
         }
 
         match src.val {
-            OperandValue::Ref(src_llval, meta, src_align) => {
-                debug_assert_eq!(meta, None);
-                // For a place-to-place transmute, call `memcpy` directly so that
-                // both arguments get the best-available alignment information.
-                let bytes = bx.cx().const_usize(size_in_bytes);
-                let flags = MemFlags::empty();
-                bx.memcpy(dst.llval, dst.align, src_llval, src_align, bytes, flags);
+            OperandValue::Ref(..) => {
+                span_bug!(
+                    self.mir.span,
+                    "Operand path should have handled transmute \
+                    from `Ref` {src:?} to place {dst:?}"
+                );
             }
-            OperandValue::Immediate(_) | OperandValue::Pair(_, _) => {
+            OperandValue::Immediate(..) | OperandValue::Pair(..) => {
                 // When we have immediate(s), the alignment of the source is irrelevant,
                 // so we can store them using the destination's alignment.
                 let llty = bx.backend_type(src.layout);
@@ -194,6 +192,94 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         }
     }
 
+    /// Attempts to transmute an `OperandValue` to another `OperandValue`.
+    ///
+    /// Returns `None` for cases that can't work in that framework, such as for
+    /// `Immediate`->`Ref` that needs an `alloc` to get the location.
+    fn codegen_transmute_operand(
+        &mut self,
+        bx: &mut Bx,
+        operand: OperandRef<'tcx, Bx::Value>,
+        cast: TyAndLayout<'tcx>,
+    ) -> Option<OperandValue<Bx::Value>> {
+        // Callers already checked that the layout sizes match
+        debug_assert_eq!(operand.layout.size, cast.size);
+
+        let operand_kind = self.value_kind(operand.layout);
+        let cast_kind = self.value_kind(cast);
+
+        match operand.val {
+            OperandValue::Ref(ptr, meta, align) => {
+                debug_assert_eq!(meta, None);
+                debug_assert!(matches!(operand_kind, OperandValueKind::Ref));
+                let cast_bty = bx.backend_type(cast);
+                let cast_ptr = bx.pointercast(ptr, bx.type_ptr_to(cast_bty));
+                let fake_place = PlaceRef::new_sized_aligned(cast_ptr, cast, align);
+                Some(bx.load_operand(fake_place).val)
+            }
+            OperandValue::Immediate(imm) => {
+                let OperandValueKind::Immediate(in_scalar) = operand_kind else {
+                    bug!("Found {operand_kind:?} for operand {operand:?}");
+                };
+                if let OperandValueKind::Immediate(out_scalar) = cast_kind {
+                    let cast_bty = bx.backend_type(cast);
+                    Some(OperandValue::Immediate(Self::transmute_immediate(
+                        bx, imm, in_scalar, out_scalar, cast_bty,
+                    )))
+                } else {
+                    None
+                }
+            }
+            OperandValue::Pair(imm_a, imm_b) => {
+                let OperandValueKind::Pair(in_a, in_b) = operand_kind else {
+                    bug!("Found {operand_kind:?} for operand {operand:?}");
+                };
+                if let OperandValueKind::Pair(out_a, out_b) = cast_kind {
+                    let out_a_ibty = bx.scalar_pair_element_backend_type(cast, 0, false);
+                    let out_b_ibty = bx.scalar_pair_element_backend_type(cast, 1, false);
+                    Some(OperandValue::Pair(
+                        Self::transmute_immediate(bx, imm_a, in_a, out_a, out_a_ibty),
+                        Self::transmute_immediate(bx, imm_b, in_b, out_b, out_b_ibty),
+                    ))
+                } else {
+                    None
+                }
+            }
+        }
+    }
+
+    /// Transmutes one of the immediates from an [`OperandValue::Immediate`]
+    /// or an [`OperandValue::Pair`] to an immediate of the target type.
+    ///
+    /// `to_backend_ty` must be the *non*-immediate backend type (so it will be
+    /// `i8`, not `i1`, for `bool`-like types.)
+    fn transmute_immediate(
+        bx: &mut Bx,
+        mut imm: Bx::Value,
+        from_scalar: abi::Scalar,
+        to_scalar: abi::Scalar,
+        to_backend_ty: Bx::Type,
+    ) -> Bx::Value {
+        use abi::Primitive::*;
+        imm = bx.from_immediate(imm);
+        imm = match (from_scalar.primitive(), to_scalar.primitive()) {
+            (Int(..) | F32 | F64, Int(..) | F32 | F64) => bx.bitcast(imm, to_backend_ty),
+            (Pointer(..), Pointer(..)) => bx.pointercast(imm, to_backend_ty),
+            (Int(..), Pointer(..)) => bx.inttoptr(imm, to_backend_ty),
+            (Pointer(..), Int(..)) => bx.ptrtoint(imm, to_backend_ty),
+            (F32 | F64, Pointer(..)) => {
+                let int_imm = bx.bitcast(imm, bx.cx().type_isize());
+                bx.inttoptr(int_imm, to_backend_ty)
+            }
+            (Pointer(..), F32 | F64) => {
+                let int_imm = bx.ptrtoint(imm, bx.cx().type_isize());
+                bx.bitcast(int_imm, to_backend_ty)
+            }
+        };
+        imm = bx.to_immediate_scalar(imm, to_scalar);
+        imm
+    }
+
     pub fn codegen_rvalue_unsized(
         &mut self,
         bx: &mut Bx,
@@ -396,7 +482,9 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                         OperandValue::Immediate(newval)
                     }
                     mir::CastKind::Transmute => {
-                        bug!("Transmute operand {:?} in `codegen_rvalue_operand`", operand);
+                        self.codegen_transmute_operand(bx, operand, cast).unwrap_or_else(|| {
+                            bug!("Unsupported transmute-as-operand of {operand:?} to {cast:?}");
+                        })
                     }
                 };
                 OperandRef { val, layout: cast }
@@ -739,10 +827,36 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
 impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
     pub fn rvalue_creates_operand(&self, rvalue: &mir::Rvalue<'tcx>, span: Span) -> bool {
         match *rvalue {
-            mir::Rvalue::Cast(mir::CastKind::Transmute, ..) =>
-                // FIXME: Now that transmute is an Rvalue, it would be nice if
-                // it could create `Immediate`s for scalars, where possible.
-                false,
+            mir::Rvalue::Cast(mir::CastKind::Transmute, ref operand, cast_ty) => {
+                let operand_ty = operand.ty(self.mir, self.cx.tcx());
+                let cast_layout = self.cx.layout_of(self.monomorphize(cast_ty));
+                let operand_layout = self.cx.layout_of(self.monomorphize(operand_ty));
+                if operand_layout.size != cast_layout.size
+                    || operand_layout.abi.is_uninhabited()
+                    || cast_layout.abi.is_uninhabited()
+                {
+                    // Send UB cases to the full form so the operand version can
+                    // `bitcast` without worrying about malformed IR.
+                    return false;
+                }
+
+                match (self.value_kind(operand_layout), self.value_kind(cast_layout)) {
+                    // Can always load from a pointer as needed
+                    (OperandValueKind::Ref, _) => true,
+
+                    // Need to generate an `alloc` to get a pointer from an immediate
+                    (OperandValueKind::Immediate(..) | OperandValueKind::Pair(..), OperandValueKind::Ref) => false,
+
+                    // When we have scalar immediates, we can convert them as needed
+                    (OperandValueKind::Immediate(..), OperandValueKind::Immediate(..)) |
+                    (OperandValueKind::Pair(..), OperandValueKind::Pair(..)) => true,
+
+                    // Send mixings between scalars and pairs through the memory route
+                    // FIXME: Maybe this could use insertvalue/extractvalue instead?
+                    (OperandValueKind::Immediate(..), OperandValueKind::Pair(..)) |
+                    (OperandValueKind::Pair(..), OperandValueKind::Immediate(..)) => false,
+                }
+            }
             mir::Rvalue::Ref(..) |
             mir::Rvalue::CopyForDeref(..) |
             mir::Rvalue::AddressOf(..) |
@@ -767,4 +881,30 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
 
         // (*) this is only true if the type is suitable
     }
+
+    /// Gets which variant of [`OperandValue`] is expected for a particular type.
+    fn value_kind(&self, layout: TyAndLayout<'tcx>) -> OperandValueKind {
+        if self.cx.is_backend_immediate(layout) {
+            debug_assert!(!self.cx.is_backend_scalar_pair(layout));
+            OperandValueKind::Immediate(match layout.abi {
+                abi::Abi::Scalar(s) => s,
+                abi::Abi::Vector { element, .. } => element,
+                x => bug!("Couldn't translate {x:?} as backend immediate"),
+            })
+        } else if self.cx.is_backend_scalar_pair(layout) {
+            let abi::Abi::ScalarPair(s1, s2) = layout.abi else {
+                bug!("Couldn't translate {:?} as backend scalar pair", layout.abi)
+            };
+            OperandValueKind::Pair(s1, s2)
+        } else {
+            OperandValueKind::Ref
+        }
+    }
+}
+
+#[derive(Debug, Copy, Clone)]
+enum OperandValueKind {
+    Ref,
+    Immediate(abi::Scalar),
+    Pair(abi::Scalar, abi::Scalar),
 }

compiler/rustc_codegen_ssa/src/traits/type_.rs

@@ -100,11 +100,22 @@ pub trait DerivedTypeMethods<'tcx>: BaseTypeMethods<'tcx> + MiscMethods<'tcx> {
 impl<'tcx, T> DerivedTypeMethods<'tcx> for T where Self: BaseTypeMethods<'tcx> + MiscMethods<'tcx> {}
 
 pub trait LayoutTypeMethods<'tcx>: Backend<'tcx> {
+    /// The backend type used for a rust type when it's in memory,
+    /// such as when it's stack-allocated or when it's being loaded or stored.
     fn backend_type(&self, layout: TyAndLayout<'tcx>) -> Self::Type;
     fn cast_backend_type(&self, ty: &CastTarget) -> Self::Type;
     fn fn_decl_backend_type(&self, fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> Self::Type;
     fn fn_ptr_backend_type(&self, fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> Self::Type;
     fn reg_backend_type(&self, ty: &Reg) -> Self::Type;
+    /// The backend type used for a rust type when it's in an SSA register.
+    ///
+    /// For nearly all types this is the same as the [`Self::backend_type`], however
+    /// `bool` (and other `0`-or-`1` values) are kept as [`BaseTypeMethods::type_i1`]
+    /// in registers but as [`BaseTypeMethods::type_i8`] in memory.
+    ///
+    /// Converting values between the two different backend types is done using
+    /// [`from_immediate`](super::BuilderMethods::from_immediate) and
+    /// [`to_immediate_scalar`](super::BuilderMethods::to_immediate_scalar).
     fn immediate_backend_type(&self, layout: TyAndLayout<'tcx>) -> Self::Type;
     fn is_backend_immediate(&self, layout: TyAndLayout<'tcx>) -> bool;
     fn is_backend_scalar_pair(&self, layout: TyAndLayout<'tcx>) -> bool;