Merge pull request #992 from hash-org/impl-repeat-expr

implement repeat expressions

compiler/hash-ast-utils/src/pretty/mod.rs

@@ -280,6 +280,21 @@ where
         self.write("]")
     }
 
+    type RepeatExprRet = ();
+
+    fn visit_repeat_expr(
+        &mut self,
+        node: ast::AstNodeRef<ast::RepeatExpr>,
+    ) -> Result<Self::RepeatExprRet, Self::Error> {
+        let ast::RepeatExpr { subject, repeat } = node.body();
+
+        self.write("[")?;
+        self.visit_expr(subject.ast_ref())?;
+        self.write("; ")?;
+        self.visit_expr(repeat.ast_ref())?;
+        self.write("]")
+    }
+
     type TupleLitEntryRet = ();
 
     fn visit_tuple_lit_entry(

compiler/hash-ast-utils/src/tree.rs

@@ -858,6 +858,23 @@ impl AstVisitor for AstTreePrinter {
         ))
     }
 
+    type RepeatExprRet = TreeNode;
+
+    fn visit_repeat_expr(
+        &self,
+        node: ast::AstNodeRef<ast::RepeatExpr>,
+    ) -> Result<Self::RepeatExprRet, Self::Error> {
+        let walk::RepeatExpr { subject, repeat } = walk::walk_repeat_expr(self, node)?;
+
+        Ok(TreeNode::branch(
+            "repeat_expr",
+            vec![
+                TreeNode::branch("subject", vec![subject]),
+                TreeNode::branch("repeat", vec![repeat]),
+            ],
+        ))
+    }
+
     type StructDefRet = TreeNode;
     fn visit_struct_def(
         &self,

compiler/hash-ast/src/ast.rs

@@ -442,6 +442,11 @@ impl<T> AstNodes<T> {
         SpanMap::span_of(self.id)
     }
 
+    /// Insert an item into the [AstNodes] at a particular index.
+    pub fn insert(&mut self, item: AstNode<T>, index: usize) {
+        self.nodes.insert(index, item);
+    }
+
     /// Merge two [AstNodes] together, this will append the nodes of the
     /// other [AstNodes] to this one, and then return the new [AstNodes].
     ///
@@ -2235,6 +2240,17 @@ define_tree! {
         pub index_expr: Child!(Expr),
     }
 
+    /// A repeat expression `[x; 5]`.
+    #[derive(Debug, PartialEq, Clone)]
+    #[node]
+    pub struct RepeatExpr {
+        /// The subject that is being repeated.
+        pub subject: Child!(Expr),
+
+        /// The constant specifying the number of repeats of the subject.
+        pub repeat: Child!(Expr),
+    }
+
     /// An expression.
     #[derive(Debug, PartialEq, Clone)]
     #[node]
@@ -2315,6 +2331,10 @@ define_tree! {
         /// Expression to index a subject e.g. `arr[x]`
         Index(IndexExpr),
 
+        /// An expression that specifies that an operand is to be
+        /// repeated a certain number of times, e.g. `[1; 5]`.
+        Repeat(RepeatExpr),
+
         /// An expression that captures a variable or a pattern being assigned
         /// to a right hand-side expression such as `x = 3`.
         Assign(AssignExpr),

compiler/hash-codegen-llvm/src/translation/builder.rs

@@ -39,7 +39,7 @@ use llvm_sys::core as llvm;
 
 use super::{
     abi::ExtendedFnAbiMethods, layouts::ExtendedLayoutMethods, ty::ExtendedTyBuilderMethods,
-    LLVMBuilder,
+    LLVMBuilder, EMPTY_NAME,
 };
 use crate::misc::{
     AtomicOrderingWrapper, FloatPredicateWrapper, IntPredicateWrapper, MetadataTypeKind,
@@ -71,7 +71,7 @@ impl<'a, 'b, 'm> LLVMBuilder<'a, 'b, 'm> {
 
         // Create the PHI value, and then add all of the incoming values.
         let ty: BasicTypeEnum = ty.try_into().unwrap();
-        let phi = self.builder.build_phi(ty, "phi");
+        let phi = self.builder.build_phi(ty, "");
 
         // @@Efficiency: patch inkwell to allow to provide these references directly...
         let blocks_and_values = blocks
@@ -464,10 +464,21 @@ impl<'a, 'b, 'm> BlockBuilderMethods<'a, 'b> for LLVMBuilder<'a, 'b, 'm> {
     }
 
     fn unchecked_uadd(&mut self, lhs: Self::Value, rhs: Self::Value) -> Self::Value {
-        let lhs = lhs.into_int_value();
-        let rhs = rhs.into_int_value();
-
-        self.builder.build_int_nuw_add(lhs, rhs, "").into()
+        // @@PatchInkwell: we need to potentially support `phi_value` from
+        // the lhs, so we can't use the standard builder method here...
+        //
+        // let lhs = lhs.into_int_value();
+        // let rhs = rhs.into_int_value();
+        // self.builder.build_int_nuw_add(lhs, rhs, "").into()
+        //
+        unsafe {
+            AnyValueEnum::new(llvm::LLVMBuildNUWAdd(
+                self.builder.as_mut_ptr(),
+                lhs.as_value_ref(),
+                rhs.as_value_ref(),
+                EMPTY_NAME,
+            ))
+        }
     }
 
     fn unchecked_ssub(&mut self, lhs: Self::Value, rhs: Self::Value) -> Self::Value {
@@ -579,10 +590,20 @@ impl<'a, 'b, 'm> BlockBuilderMethods<'a, 'b> for LLVMBuilder<'a, 'b, 'm> {
 
     fn icmp(&mut self, op: IntComparisonKind, lhs: Self::Value, rhs: Self::Value) -> Self::Value {
         let op = IntPredicateWrapper::from(op).0;
-        let lhs = lhs.into_int_value();
-        let rhs = rhs.into_int_value();
 
-        self.builder.build_int_compare(op, lhs, rhs, "").into()
+        // @@PatchInkwell: `built_int_compare` doesn't support passing in `phi` values
+        // because of the `IntMathType` bound on the operands.
+        unsafe {
+            let value = llvm::LLVMBuildICmp(
+                self.builder.as_mut_ptr(),
+                op.into(),
+                lhs.as_value_ref(),
+                rhs.as_value_ref(),
+                EMPTY_NAME,
+            );
+
+            AnyValueEnum::new(value)
+        }
     }
 
     fn fcmp(&mut self, op: RealComparisonKind, lhs: Self::Value, rhs: Self::Value) -> Self::Value {
@@ -751,8 +772,6 @@ impl<'a, 'b, 'm> BlockBuilderMethods<'a, 'b> for LLVMBuilder<'a, 'b, 'm> {
     ) {
         let zero = self.const_usize(0);
         let count = self.const_usize(count as u64);
-        let start = destination.project_index(self, zero).value;
-        let end = destination.project_index(self, count).value;
 
         // setup the blocks to write the operand to...
         let header_bb = self.append_sibling_block("repeat_loop_header");
@@ -764,37 +783,20 @@ impl<'a, 'b, 'm> BlockBuilderMethods<'a, 'b> for LLVMBuilder<'a, 'b, 'm> {
 
         let mut header_builder = Self::build(self.ctx, header_bb);
 
-        let basic_value: BasicValueEnum = start.try_into().unwrap();
+        let zero_val: BasicValueEnum = zero.try_into().unwrap();
         let current: AnyValueEnum = header_builder
-            .phi(
-                self.ctx.ty_of_value(start),
-                &[&basic_value as &dyn BasicValue],
-                &[self.basic_block()],
-            )
+            .phi(self.ctx.ty_of_value(zero), &[&zero_val], &[self.basic_block()])
             .into();
 
-        let keep_going = header_builder.icmp(IntComparisonKind::Ne, current, end);
+        let keep_going = header_builder.icmp(IntComparisonKind::Ult, current, count);
         header_builder.conditional_branch(keep_going, body_bb, next_bb);
 
         let mut body_builder = Self::build(self.ctx, body_bb);
+        let dest = destination.project_index(&mut body_builder, current);
+        operand.value.store(&mut body_builder, dest);
 
-        let field_info = destination.info.field(self.ctx.layouts(), 0);
-        let field_size = field_info.size();
-
-        let alignment = destination.alignment.restrict_to(field_size);
-
-        // now we want to emit a `store` for the value we are writing
-        operand.value.store(
-            &mut body_builder,
-            PlaceRef { value: current, extra: None, info: operand.info, alignment },
-        );
-
-        // Compute the "next" value...
-        let ty = self.backend_ty_from_info(operand.info);
-        let next: BasicValueEnum = body_builder
-            .bounded_get_element_pointer(ty, current, &[self.const_usize(1)])
-            .try_into()
-            .unwrap();
+        let next: BasicValueEnum =
+            body_builder.unchecked_uadd(current, self.const_usize(1)).try_into().unwrap();
 
         // and branch back to the header
         body_builder.branch(header_bb);
@@ -1013,13 +1015,29 @@ impl<'a, 'b, 'm> BlockBuilderMethods<'a, 'b> for LLVMBuilder<'a, 'b, 'm> {
         indices: &[Self::Value],
     ) -> Self::Value {
         let ty: BasicTypeEnum = ty.try_into().unwrap();
-        let indices = indices.iter().map(|i| i.into_int_value()).collect::<Vec<_>>();
 
         // ## Safety: If the `indices` are invalid or out of bounds, LLVM
         // is likely to segfault, which is noted by Inkwell and thus labelled
         // as `unsafe`.
         unsafe {
-            self.builder.build_in_bounds_gep(ty, ptr.into_pointer_value(), &indices, "").into()
+            let pointee = ptr.into_pointer_value();
+            let mut index_values: Vec<llvm_sys::prelude::LLVMValueRef> =
+                indices.iter().map(|val| val.as_value_ref()).collect();
+
+            // @@PatchInkwell, we can't use `self.builder.build_in_bounds_gep()` because
+            // again of the possibility of the `indices` to be non-int values.
+            //
+            // We need to patch inkwell to support this.
+            let value = llvm::LLVMBuildInBoundsGEP2(
+                self.builder.as_mut_ptr(),
+                ty.as_type_ref(),
+                pointee.as_value_ref(),
+                index_values.as_mut_ptr(),
+                index_values.len() as u32,
+                EMPTY_NAME,
+            );
+
+            AnyValueEnum::new(value)
         }
     }
 

compiler/hash-codegen-llvm/src/translation/mod.rs

@@ -2,6 +2,8 @@
 //! logic of transforming generated Hash IR into LLVM IR so that it can be
 //! compiled by LLVM into a native executable with the specified target triple.
 
+use std::ffi::{c_char, CStr};
+
 use hash_codegen::{
     backend::CodeGenStorage,
     layout::compute::LayoutComputer,
@@ -25,6 +27,13 @@ mod misc;
 mod statics;
 pub(crate) mod ty;
 
+/// An empty c-string.
+const EMPTY_C_STR: &CStr = unsafe { CStr::from_bytes_with_nul_unchecked(b"\0") };
+
+/// Used to fill in names where LLVM requires a name for an instruction, or some
+/// value. We don't care about the names, so we just use an empty c-string.
+pub(crate) const EMPTY_NAME: *const c_char = EMPTY_C_STR.as_ptr();
+
 /// A [Builder] is defined as being a context that is used to implement
 /// all of the specified builder methods.
 pub struct LLVMBuilder<'a, 'b, 'm> {

compiler/hash-codegen/src/lower/rvalue.rs

@@ -17,7 +17,7 @@ use super::{
     FnBuilder,
 };
 use crate::{
-    common::{CheckedOp, IntComparisonKind, TypeKind},
+    common::{CheckedOp, IntComparisonKind, MemFlags, TypeKind},
     traits::{
         builder::BlockBuilderMethods, constants::ConstValueBuilderMethods, layout::LayoutMethods,
         ty::TypeBuilderMethods,
@@ -147,6 +147,41 @@ impl<'a, 'b, Builder: BlockBuilderMethods<'a, 'b>> FnBuilder<'a, 'b, Builder> {
                 let operand = self.codegen_operand(builder, operand);
                 operand.value.store(builder, destination);
             }
+            ir::RValue::Repeat(op, repeat) => {
+                let op = self.codegen_operand(builder, op);
+
+                // We don't write anything into the destination if it is a ZST... i.e. being
+                // ignored!
+                if destination.info.is_zst() {
+                    return;
+                }
+
+                // Optimisation for immediates, we can use llvm.memset.p0i8.* to
+                // initialise the memory.
+                if let OperandValue::Immediate(val) = op.value {
+                    let zero = builder.const_usize(0);
+                    let start = destination.project_index(builder, zero).value;
+                    let size = builder.const_usize(destination.info.size().bytes());
+
+                    // If it is zero, we just fill the array with zeroes...
+                    if self.ctx.const_to_optional_u128(val, false) == Some(0) {
+                        let fill = self.ctx.const_u8(0);
+                        builder.memset(start, fill, size, destination.alignment, MemFlags::empty());
+                        return;
+                    }
+
+                    // If the type is the size of a `i8`, we just memset it with the value...
+                    let v = builder.value_from_immediate(val);
+                    if self.ctx.ty_of_value(v) == self.ctx.type_i8() {
+                        builder.memset(start, v, size, destination.alignment, MemFlags::empty());
+                        return;
+                    }
+                }
+
+                // Slow-path: we actually need to generate code to write the value into the
+                // array...
+                builder.write_operand_repeatedly(op, *repeat, destination)
+            }
             ir::RValue::Aggregate(kind, operands) => {
                 let destination = match *kind {
                     ir::AggregateKind::Enum(_, variant) => {
@@ -335,7 +370,7 @@ impl<'a, 'b, Builder: BlockBuilderMethods<'a, 'b>> FnBuilder<'a, 'b, Builder> {
                 let value = OperandValue::Immediate(new_value);
                 OperandRef { value, info: cast_ty }
             }
-            ir::RValue::Aggregate(_, _) => {
+            ir::RValue::Repeat(_, _) | ir::RValue::Aggregate(_, _) => {
                 // This is only called if the aggregate value is a ZST, so we just
                 // create a new ZST operand...
                 OperandRef::zst(builder.layout_of(self.ty_of_rvalue(rvalue)))
@@ -546,7 +581,7 @@ impl<'a, 'b, Builder: BlockBuilderMethods<'a, 'b>> FnBuilder<'a, 'b, Builder> {
             | ir::RValue::Ref(_, _, _)
             | ir::RValue::Cast(_, _, _)
             | ir::RValue::Discriminant(_) => true,
-            ir::RValue::Aggregate(_, _) => {
+            ir::RValue::Repeat(_, _) | ir::RValue::Aggregate(_, _) => {
                 // check if the type is a ZST, and if so this satisfies the
                 // case that the rvalue creates an operand...
                 let ty = self.ty_of_rvalue(rvalue);