Merge pull request #9960 from Snektron/bit-not

Some not and vector stuff

4 files changed, 210 insertions, 31 deletions

diff --git a/src/Sema.zig b/src/Sema.zig
index cfc541453a..2211e69fb0 100644
--- a/src/Sema.zig
+++ b/src/Sema.zig
@@ -6629,8 +6629,42 @@ fn zirBitNot(sema: *Sema, block: *Block, inst: Zir.Inst.Index) CompileError!Air.
     const tracy = trace(@src());
     defer tracy.end();
 
-    _ = inst;
-    return sema.fail(block, sema.src, "TODO implement zirBitNot", .{});
+    const inst_data = sema.code.instructions.items(.data)[inst].un_node;
+    const src = inst_data.src();
+    const operand_src = src; // TODO put this on the operand, not the '~'
+
+    const operand = sema.resolveInst(inst_data.operand);
+    const operand_type = sema.typeOf(operand);
+    const scalar_type = operand_type.scalarType();
+
+    if (scalar_type.zigTypeTag() != .Int) {
+        return sema.fail(block, src, "unable to perform binary not operation on type '{}'", .{operand_type});
+    }
+
+    if (try sema.resolveMaybeUndefVal(block, operand_src, operand)) |val| {
+        const target = sema.mod.getTarget();
+        if (val.isUndef()) {
+            return sema.addConstUndef(scalar_type);
+        } else if (operand_type.zigTypeTag() == .Vector) {
+            const vec_len = operand_type.arrayLen();
+            var elem_val_buf: Value.ElemValueBuffer = undefined;
+            const elems = try sema.arena.alloc(Value, vec_len);
+            for (elems) |*elem, i| {
+                const elem_val = val.elemValueBuffer(i, &elem_val_buf);
+                elem.* = try elem_val.bitwiseNot(scalar_type, sema.arena, target);
+            }
+            return sema.addConstant(
+                operand_type,
+                try Value.Tag.array.create(sema.arena, elems),
+            );
+        } else {
+            const result_val = try val.bitwiseNot(scalar_type, sema.arena, target);
+            return sema.addConstant(scalar_type, result_val);
+        }
+    }
+
+    try sema.requireRuntimeBlock(block, src);
+    return block.addTyOp(.not, operand_type, operand);
 }
 
 fn zirArrayCat(sema: *Sema, block: *Block, inst: Zir.Inst.Index) CompileError!Air.Inst.Ref {
@@ -8239,12 +8273,13 @@ fn zirBoolNot(sema: *Sema, block: *Block, inst: Zir.Inst.Index) CompileError!Air
 
     const bool_type = Type.initTag(.bool);
     const operand = try sema.coerce(block, bool_type, uncasted_operand, operand_src);
-    if (try sema.resolveDefinedValue(block, operand_src, operand)) |val| {
-        if (val.toBool()) {
-            return Air.Inst.Ref.bool_false;
-        } else {
-            return Air.Inst.Ref.bool_true;
-        }
+    if (try sema.resolveMaybeUndefVal(block, operand_src, operand)) |val| {
+        return if (val.isUndef())
+            sema.addConstUndef(bool_type)
+        else if (val.toBool())
+            Air.Inst.Ref.bool_false
+        else
+            Air.Inst.Ref.bool_true;
     }
     try sema.requireRuntimeBlock(block, src);
     return block.addTyOp(.not, bool_type, operand);
@@ -11640,7 +11675,11 @@ fn coerce(
             else => {},
         },
         .Array => switch (inst_ty.zigTypeTag()) {
-            .Vector => return sema.coerceVectorToArray(block, dest_ty, dest_ty_src, inst, inst_src),
+            .Vector => return sema.coerceVectorInMemory(block, dest_ty, dest_ty_src, inst, inst_src),
+            else => {},
+        },
+        .Vector => switch (inst_ty.zigTypeTag()) {
+            .Array => return sema.coerceVectorInMemory(block, dest_ty, dest_ty_src, inst, inst_src),
             else => {},
         },
         else => {},
@@ -12224,46 +12263,49 @@ fn coerceEnumToUnion(
     return sema.failWithOwnedErrorMsg(msg);
 }
 
-fn coerceVectorToArray(
+// Coerces vectors/arrays which have the same in-memory layout. This can be used for
+// both coercing from and to vectors.
+fn coerceVectorInMemory(
     sema: *Sema,
     block: *Block,
-    array_ty: Type,
-    array_ty_src: LazySrcLoc,
-    vector: Air.Inst.Ref,
-    vector_src: LazySrcLoc,
+    dest_ty: Type,
+    dest_ty_src: LazySrcLoc,
+    inst: Air.Inst.Ref,
+    inst_src: LazySrcLoc,
 ) !Air.Inst.Ref {
-    const vector_ty = sema.typeOf(vector);
-    const array_len = array_ty.arrayLen();
-    const vector_len = vector_ty.arrayLen();
-    if (array_len != vector_len) {
+    const inst_ty = sema.typeOf(inst);
+    const inst_len = inst_ty.arrayLen();
+    const dest_len = dest_ty.arrayLen();
+
+    if (dest_len != inst_len) {
         const msg = msg: {
-            const msg = try sema.errMsg(block, vector_src, "expected {}, found {}", .{
-                array_ty, vector_ty,
+            const msg = try sema.errMsg(block, inst_src, "expected {}, found {}", .{
+                dest_ty, inst_ty,
             });
             errdefer msg.destroy(sema.gpa);
-            try sema.errNote(block, array_ty_src, msg, "array has length {d}", .{array_len});
-            try sema.errNote(block, vector_src, msg, "vector has length {d}", .{vector_len});
+            try sema.errNote(block, dest_ty_src, msg, "destination has length {d}", .{dest_len});
+            try sema.errNote(block, inst_src, msg, "source has length {d}", .{inst_len});
             break :msg msg;
         };
         return sema.failWithOwnedErrorMsg(msg);
     }
 
     const target = sema.mod.getTarget();
-    const array_elem_ty = array_ty.childType();
-    const vector_elem_ty = vector_ty.childType();
-    const in_memory_result = coerceInMemoryAllowed(array_elem_ty, vector_elem_ty, false, target);
+    const dest_elem_ty = dest_ty.childType();
+    const inst_elem_ty = inst_ty.childType();
+    const in_memory_result = coerceInMemoryAllowed(dest_elem_ty, inst_elem_ty, false, target);
     if (in_memory_result != .ok) {
         // TODO recursive error notes for coerceInMemoryAllowed failure
-        return sema.fail(block, vector_src, "expected {}, found {}", .{ array_ty, vector_ty });
+        return sema.fail(block, inst_src, "expected {}, found {}", .{ dest_ty, inst_ty });
     }
 
-    if (try sema.resolveMaybeUndefVal(block, vector_src, vector)) |vector_val| {
+    if (try sema.resolveMaybeUndefVal(block, inst_src, inst)) |inst_val| {
         // These types share the same comptime value representation.
-        return sema.addConstant(array_ty, vector_val);
+        return sema.addConstant(dest_ty, inst_val);
     }
 
-    try sema.requireRuntimeBlock(block, vector_src);
-    return block.addTyOp(.bitcast, array_ty, vector);
+    try sema.requireRuntimeBlock(block, inst_src);
+    return block.addTyOp(.bitcast, dest_ty, inst);
 }
 
 fn coerceCompatibleErrorSets(
diff --git a/src/codegen/llvm.zig b/src/codegen/llvm.zig
index 68b0f6b635..61361b6dcc 100644
--- a/src/codegen/llvm.zig
+++ b/src/codegen/llvm.zig
@@ -1299,6 +1299,66 @@ pub const DeclGen = struct {
                 }
                 return llvm_union_ty.constNamedStruct(&fields, fields.len);
             },
+            .Vector => switch (tv.val.tag()) {
+                .bytes => {
+                    // Note, sentinel is not stored even if the type has a sentinel.
+                    const bytes = tv.val.castTag(.bytes).?.data;
+                    const vector_len = tv.ty.arrayLen();
+                    assert(vector_len == bytes.len or vector_len + 1 == bytes.len);
+
+                    const elem_ty = tv.ty.elemType();
+                    const llvm_elems = try self.gpa.alloc(*const llvm.Value, vector_len);
+                    defer self.gpa.free(llvm_elems);
+                    for (llvm_elems) |*elem, i| {
+                        var byte_payload: Value.Payload.U64 = .{
+                            .base = .{ .tag = .int_u64 },
+                            .data = bytes[i],
+                        };
+
+                        elem.* = try self.genTypedValue(.{
+                            .ty = elem_ty,
+                            .val = Value.initPayload(&byte_payload.base),
+                        });
+                    }
+                    return llvm.constVector(
+                        llvm_elems.ptr,
+                        @intCast(c_uint, llvm_elems.len),
+                    );
+                },
+                .array => {
+                    // Note, sentinel is not stored even if the type has a sentinel.
+                    // The value includes the sentinel in those cases.
+                    const elem_vals = tv.val.castTag(.array).?.data;
+                    const vector_len = tv.ty.arrayLen();
+                    assert(vector_len == elem_vals.len or vector_len + 1 == elem_vals.len);
+                    const elem_ty = tv.ty.elemType();
+                    const llvm_elems = try self.gpa.alloc(*const llvm.Value, vector_len);
+                    defer self.gpa.free(llvm_elems);
+                    for (llvm_elems) |*elem, i| {
+                        elem.* = try self.genTypedValue(.{ .ty = elem_ty, .val = elem_vals[i] });
+                    }
+                    return llvm.constVector(
+                        llvm_elems.ptr,
+                        @intCast(c_uint, llvm_elems.len),
+                    );
+                },
+                .repeated => {
+                    // Note, sentinel is not stored even if the type has a sentinel.
+                    const val = tv.val.castTag(.repeated).?.data;
+                    const elem_ty = tv.ty.elemType();
+                    const len = tv.ty.arrayLen();
+                    const llvm_elems = try self.gpa.alloc(*const llvm.Value, len);
+                    defer self.gpa.free(llvm_elems);
+                    for (llvm_elems) |*elem| {
+                        elem.* = try self.genTypedValue(.{ .ty = elem_ty, .val = val });
+                    }
+                    return llvm.constVector(
+                        llvm_elems.ptr,
+                        @intCast(c_uint, llvm_elems.len),
+                    );
+                },
+                else => unreachable,
+            },
 
             .ComptimeInt => unreachable,
             .ComptimeFloat => unreachable,
@@ -1313,7 +1373,6 @@ pub const DeclGen = struct {
 
             .Frame,
             .AnyFrame,
-            .Vector,
             => return self.todo("implement const of type '{}'", .{tv.ty}),
         }
     }
@@ -2992,6 +3051,43 @@ pub const FuncGen = struct {
                 }
             }
             return array_ptr;
+        } else if (operand_ty.zigTypeTag() == .Array and inst_ty.zigTypeTag() == .Vector) {
+            const target = self.dg.module.getTarget();
+            const elem_ty = operand_ty.childType();
+            const llvm_vector_ty = try self.dg.llvmType(inst_ty);
+            if (!isByRef(operand_ty)) {
+                return self.dg.todo("implement bitcast non-ref array to vector", .{});
+            }
+
+            const bitcast_ok = elem_ty.bitSize(target) == elem_ty.abiSize(target) * 8;
+            if (bitcast_ok) {
+                const llvm_vector_ptr_ty = llvm_vector_ty.pointerType(0);
+                const casted_ptr = self.builder.buildBitCast(operand, llvm_vector_ptr_ty, "");
+                const vector = self.builder.buildLoad(casted_ptr, "");
+                // The array is aligned to the element's alignment, while the vector might have a completely
+                // different alignment. This means we need to enforce the alignment of this load.
+                vector.setAlignment(elem_ty.abiAlignment(target));
+                return vector;
+            } else {
+                // If the ABI size of the element type is not evenly divisible by size in bits;
+                // a simple bitcast will not work, and we fall back to extractelement.
+                const llvm_usize = try self.dg.llvmType(Type.usize);
+                const llvm_u32 = self.context.intType(32);
+                const zero = llvm_usize.constNull();
+                const vector_len = operand_ty.arrayLen();
+                var vector = llvm_vector_ty.getUndef();
+                var i: u64 = 0;
+                while (i < vector_len) : (i += 1) {
+                    const index_usize = llvm_usize.constInt(i, .False);
+                    const index_u32 = llvm_u32.constInt(i, .False);
+                    const indexes: [2]*const llvm.Value = .{ zero, index_usize };
+                    const elem_ptr = self.builder.buildInBoundsGEP(operand, &indexes, indexes.len, "");
+                    const elem = self.builder.buildLoad(elem_ptr, "");
+                    vector = self.builder.buildInsertElement(vector, elem, index_u32, "");
+                }
+
+                return vector;
+            }
         }
 
         return self.builder.buildBitCast(operand, llvm_dest_ty, "");
diff --git a/src/codegen/llvm/bindings.zig b/src/codegen/llvm/bindings.zig
index 29a46a81ee..1b7806d22a 100644
--- a/src/codegen/llvm/bindings.zig
+++ b/src/codegen/llvm/bindings.zig
@@ -313,6 +313,12 @@ pub const VerifierFailureAction = enum(c_int) {
 pub const constNeg = LLVMConstNeg;
 extern fn LLVMConstNeg(ConstantVal: *const Value) *const Value;
 
+pub const constVector = LLVMConstVector;
+extern fn LLVMConstVector(
+    ScalarConstantVals: [*]*const Value,
+    Size: c_uint,
+) *const Value;
+
 pub const getEnumAttributeKindForName = LLVMGetEnumAttributeKindForName;
 extern fn LLVMGetEnumAttributeKindForName(Name: [*]const u8, SLen: usize) c_uint;
 
@@ -567,6 +573,15 @@ pub const Builder = opaque {
         Name: [*:0]const u8,
     ) *const Value;
 
+    pub const buildInsertElement = LLVMBuildInsertElement;
+    extern fn LLVMBuildInsertElement(
+        *const Builder,
+        VecVal: *const Value,
+        EltVal: *const Value,
+        Index: *const Value,
+        Name: [*:0]const u8,
+    ) *const Value;
+
     pub const buildPtrToInt = LLVMBuildPtrToInt;
     extern fn LLVMBuildPtrToInt(
         *const Builder,
diff --git a/src/value.zig b/src/value.zig
index 127c6c27b4..030ee8f783 100644
--- a/src/value.zig
+++ b/src/value.zig
@@ -2081,6 +2081,32 @@ pub const Value = extern union {
         };
     }
 
+    /// operands must be integers; handles undefined.
+    pub fn bitwiseNot(val: Value, ty: Type, arena: *Allocator, target: Target) !Value {
+        if (val.isUndef()) return Value.initTag(.undef);
+
+        const info = ty.intInfo(target);
+
+        // TODO is this a performance issue? maybe we should try the operation without
+        // resorting to BigInt first.
+        var val_space: Value.BigIntSpace = undefined;
+        const val_bigint = val.toBigInt(&val_space);
+        const limbs = try arena.alloc(
+            std.math.big.Limb,
+            std.math.big.int.calcTwosCompLimbCount(info.bits),
+        );
+
+        var result_bigint = BigIntMutable{ .limbs = limbs, .positive = undefined, .len = undefined };
+        result_bigint.bitNotWrap(val_bigint, info.signedness, info.bits);
+        const result_limbs = result_bigint.limbs[0..result_bigint.len];
+
+        if (result_bigint.positive) {
+            return Value.Tag.int_big_positive.create(arena, result_limbs);
+        } else {
+            return Value.Tag.int_big_negative.create(arena, result_limbs);
+        }
+    }
+
     /// operands must be integers; handles undefined. 
     pub fn bitwiseAnd(lhs: Value, rhs: Value, arena: *Allocator) !Value {
         if (lhs.isUndef() or rhs.isUndef()) return Value.initTag(.undef);