llvm: fix missing return attributes

Closes #31636

src/codegen/llvm.zig

@@ -1259,11 +1259,6 @@ pub const Object = struct {
         defer if (deinit_wip) wip.deinit();
         wip.cursor = .{ .block = try wip.block(0, "Entry") };
 
-        if (ccAbiPromoteInt(fn_info.cc, zcu, Type.fromInterned(fn_info.return_type))) |s| switch (s) {
-            .signed => try attributes.addRetAttr(.signext, &o.builder),
-            .unsigned => try attributes.addRetAttr(.zeroext, &o.builder),
-        };
-
         // This is the list of args we will use that correspond directly to the AIR arg
         // instructions. Depending on the calling convention, this list is not necessarily
         // a bijection with the actual LLVM parameters of the function.
@@ -2812,7 +2807,11 @@ pub const Object = struct {
             const raw_llvm_ret_ty = try o.lowerType(.fromInterned(fn_info.return_type));
             try attributes.addParamAttr(it.llvm_index, .{ .sret = raw_llvm_ret_ty }, &o.builder);
             it.llvm_index += 1;
-        }
+        } else if (ccAbiPromoteInt(fn_info.cc, zcu, Type.fromInterned(fn_info.return_type))) |s| switch (s) {
+            .signed => try attributes.addRetAttr(.signext, &o.builder),
+            .unsigned => try attributes.addRetAttr(.zeroext, &o.builder),
+        };
+
         const err_return_tracing = fn_info.cc == .auto and zcu.comp.config.any_error_tracing;
         if (err_return_tracing) {
             try attributes.addParamAttr(it.llvm_index, .nonnull, &o.builder);

src/codegen/llvm/FuncGen.zig

@@ -611,14 +611,20 @@ fn airCall(self: *FuncGen, inst: Air.Inst.Index, modifier: std.builtin.CallModif
         .no_suspend, .always_inline, .compile_time => unreachable,
     }
 
-    const ret_ptr = if (!sret) null else blk: {
+    const ret_ptr = if (sret) ret_ptr: {
         const llvm_ret_ty = try o.lowerType(return_type);
         try attributes.addParamAttr(0, .{ .sret = llvm_ret_ty }, &o.builder);
 
         const alignment = return_type.abiAlignment(zcu).toLlvm();
         const ret_ptr = try self.buildAlloca(llvm_ret_ty, alignment);
         try llvm_args.append(ret_ptr);
-        break :blk ret_ptr;
+        break :ret_ptr ret_ptr;
+    } else ret_ptr: {
+        if (ccAbiPromoteInt(fn_info.cc, zcu, Type.fromInterned(fn_info.return_type))) |s| switch (s) {
+            .signed => try attributes.addRetAttr(.signext, &o.builder),
+            .unsigned => try attributes.addRetAttr(.zeroext, &o.builder),
+        };
+        break :ret_ptr null;
     };
 
     const err_return_tracing = fn_info.cc == .auto and zcu.comp.config.any_error_tracing;

test/behavior/call.zig

@@ -731,3 +731,33 @@ test "tail call function pointer" {
 
     S.foo(100);
 }
+
+test "tail call with potentially extended types" {
+    if (builtin.zig_backend == .stage2_wasm) return error.SkipZigTest; // TODO
+    if (builtin.zig_backend == .stage2_x86_64) return error.SkipZigTest; // TODO
+
+    if (builtin.zig_backend == .stage2_llvm) {
+        if (builtin.cpu.arch.isMIPS() or builtin.cpu.arch.isPowerPC() or builtin.cpu.arch.isWasm()) {
+            return error.SkipZigTest;
+        }
+    }
+
+    if (builtin.zig_backend == .stage2_c and builtin.os.tag == .windows) return error.SkipZigTest; // MSVC doesn't support always tail calls
+
+    const S = struct {
+        fn Test(comptime Return: type) type {
+            return struct {
+                fn callee(@"u8": u8, @"i8": i8, @"u16": u16, @"i16": i16) callconv(.c) Return {
+                    return @intCast(@as(i32, @"u8") + @as(i32, @"i8") + @as(i32, @"u16") + @as(i32, @"i16"));
+                }
+                fn caller(@"u8": u8, @"i8": i8, @"u16": u16, @"i16": i16) callconv(.c) Return {
+                    return @call(.always_tail, callee, .{ @"u8", @"i8", @"u16", @"i16" });
+                }
+            };
+        }
+    };
+    try std.testing.expect(S.Test(u8).caller(1, -2, 3, 4) == 6);
+    try std.testing.expect(S.Test(i8).caller(5, -6, 7, -8) == -2);
+    try std.testing.expect(S.Test(u16).caller(9, 10, 11, 12) == 42);
+    try std.testing.expect(S.Test(i16).caller(13, 14, 15, -16) == 26);
+}