x86_64: fix C abi for unions

Closes #19721

9 files changed, 386 insertions, 385 deletions

diff --git a/src/InternPool.zig b/src/InternPool.zig
index e62bdc8499..a63c81f2ba 100644
--- a/src/InternPool.zig
+++ b/src/InternPool.zig
@@ -1921,7 +1921,7 @@ pub const LoadedUnionType = struct {
         return self.flagsPtr(ip).layout;
     }
 
-    pub fn fieldAlign(self: LoadedUnionType, ip: *const InternPool, field_index: u32) Alignment {
+    pub fn fieldAlign(self: LoadedUnionType, ip: *const InternPool, field_index: usize) Alignment {
         if (self.field_aligns.len == 0) return .none;
         return self.field_aligns.get(ip)[field_index];
     }
@@ -2087,41 +2087,41 @@ pub const LoadedStructType = struct {
 
     /// Returns the already-existing field with the same name, if any.
     pub fn addFieldName(
-        self: @This(),
+        self: LoadedStructType,
         ip: *InternPool,
         name: NullTerminatedString,
     ) ?u32 {
         return ip.addFieldName(self.names_map.unwrap().?, self.field_names.start, name);
     }
 
-    pub fn fieldAlign(s: @This(), ip: *const InternPool, i: usize) Alignment {
+    pub fn fieldAlign(s: LoadedStructType, ip: *const InternPool, i: usize) Alignment {
         if (s.field_aligns.len == 0) return .none;
         return s.field_aligns.get(ip)[i];
     }
 
-    pub fn fieldInit(s: @This(), ip: *const InternPool, i: usize) Index {
+    pub fn fieldInit(s: LoadedStructType, ip: *const InternPool, i: usize) Index {
         if (s.field_inits.len == 0) return .none;
         assert(s.haveFieldInits(ip));
         return s.field_inits.get(ip)[i];
     }
 
     /// Returns `none` in the case the struct is a tuple.
-    pub fn fieldName(s: @This(), ip: *const InternPool, i: usize) OptionalNullTerminatedString {
+    pub fn fieldName(s: LoadedStructType, ip: *const InternPool, i: usize) OptionalNullTerminatedString {
         if (s.field_names.len == 0) return .none;
         return s.field_names.get(ip)[i].toOptional();
     }
 
-    pub fn fieldIsComptime(s: @This(), ip: *const InternPool, i: usize) bool {
+    pub fn fieldIsComptime(s: LoadedStructType, ip: *const InternPool, i: usize) bool {
         return s.comptime_bits.getBit(ip, i);
     }
 
-    pub fn setFieldComptime(s: @This(), ip: *InternPool, i: usize) void {
+    pub fn setFieldComptime(s: LoadedStructType, ip: *InternPool, i: usize) void {
         s.comptime_bits.setBit(ip, i);
     }
 
     /// Reads the non-opv flag calculated during AstGen. Used to short-circuit more
     /// complicated logic.
-    pub fn knownNonOpv(s: @This(), ip: *InternPool) bool {
+    pub fn knownNonOpv(s: LoadedStructType, ip: *InternPool) bool {
         return switch (s.layout) {
             .@"packed" => false,
             .auto, .@"extern" => s.flagsPtr(ip).known_non_opv,
@@ -2130,7 +2130,7 @@ pub const LoadedStructType = struct {
 
     /// The returned pointer expires with any addition to the `InternPool`.
     /// Asserts the struct is not packed.
-    pub fn flagsPtr(self: @This(), ip: *const InternPool) *Tag.TypeStruct.Flags {
+    pub fn flagsPtr(self: LoadedStructType, ip: *const InternPool) *Tag.TypeStruct.Flags {
         assert(self.layout != .@"packed");
         const flags_field_index = std.meta.fieldIndex(Tag.TypeStruct, "flags").?;
         return @ptrCast(&ip.extra.items[self.extra_index + flags_field_index]);
@@ -2138,13 +2138,13 @@ pub const LoadedStructType = struct {
 
     /// The returned pointer expires with any addition to the `InternPool`.
     /// Asserts that the struct is packed.
-    pub fn packedFlagsPtr(self: @This(), ip: *const InternPool) *Tag.TypeStructPacked.Flags {
+    pub fn packedFlagsPtr(self: LoadedStructType, ip: *const InternPool) *Tag.TypeStructPacked.Flags {
         assert(self.layout == .@"packed");
         const flags_field_index = std.meta.fieldIndex(Tag.TypeStructPacked, "flags").?;
         return @ptrCast(&ip.extra.items[self.extra_index + flags_field_index]);
     }
 
-    pub fn assumeRuntimeBitsIfFieldTypesWip(s: @This(), ip: *InternPool) bool {
+    pub fn assumeRuntimeBitsIfFieldTypesWip(s: LoadedStructType, ip: *InternPool) bool {
         if (s.layout == .@"packed") return false;
         const flags_ptr = s.flagsPtr(ip);
         if (flags_ptr.field_types_wip) {
@@ -2154,7 +2154,7 @@ pub const LoadedStructType = struct {
         return false;
     }
 
-    pub fn setTypesWip(s: @This(), ip: *InternPool) bool {
+    pub fn setTypesWip(s: LoadedStructType, ip: *InternPool) bool {
         if (s.layout == .@"packed") return false;
         const flags_ptr = s.flagsPtr(ip);
         if (flags_ptr.field_types_wip) return true;
@@ -2162,12 +2162,12 @@ pub const LoadedStructType = struct {
         return false;
     }
 
-    pub fn clearTypesWip(s: @This(), ip: *InternPool) void {
+    pub fn clearTypesWip(s: LoadedStructType, ip: *InternPool) void {
         if (s.layout == .@"packed") return;
         s.flagsPtr(ip).field_types_wip = false;
     }
 
-    pub fn setLayoutWip(s: @This(), ip: *InternPool) bool {
+    pub fn setLayoutWip(s: LoadedStructType, ip: *InternPool) bool {
         if (s.layout == .@"packed") return false;
         const flags_ptr = s.flagsPtr(ip);
         if (flags_ptr.layout_wip) return true;
@@ -2175,12 +2175,12 @@ pub const LoadedStructType = struct {
         return false;
     }
 
-    pub fn clearLayoutWip(s: @This(), ip: *InternPool) void {
+    pub fn clearLayoutWip(s: LoadedStructType, ip: *InternPool) void {
         if (s.layout == .@"packed") return;
         s.flagsPtr(ip).layout_wip = false;
     }
 
-    pub fn setAlignmentWip(s: @This(), ip: *InternPool) bool {
+    pub fn setAlignmentWip(s: LoadedStructType, ip: *InternPool) bool {
         if (s.layout == .@"packed") return false;
         const flags_ptr = s.flagsPtr(ip);
         if (flags_ptr.alignment_wip) return true;
@@ -2188,12 +2188,12 @@ pub const LoadedStructType = struct {
         return false;
     }
 
-    pub fn clearAlignmentWip(s: @This(), ip: *InternPool) void {
+    pub fn clearAlignmentWip(s: LoadedStructType, ip: *InternPool) void {
         if (s.layout == .@"packed") return;
         s.flagsPtr(ip).alignment_wip = false;
     }
 
-    pub fn setInitsWip(s: @This(), ip: *InternPool) bool {
+    pub fn setInitsWip(s: LoadedStructType, ip: *InternPool) bool {
         switch (s.layout) {
             .@"packed" => {
                 const flag = &s.packedFlagsPtr(ip).field_inits_wip;
@@ -2210,14 +2210,14 @@ pub const LoadedStructType = struct {
         }
     }
 
-    pub fn clearInitsWip(s: @This(), ip: *InternPool) void {
+    pub fn clearInitsWip(s: LoadedStructType, ip: *InternPool) void {
         switch (s.layout) {
             .@"packed" => s.packedFlagsPtr(ip).field_inits_wip = false,
             .auto, .@"extern" => s.flagsPtr(ip).field_inits_wip = false,
         }
     }
 
-    pub fn setFullyResolved(s: @This(), ip: *InternPool) bool {
+    pub fn setFullyResolved(s: LoadedStructType, ip: *InternPool) bool {
         if (s.layout == .@"packed") return true;
         const flags_ptr = s.flagsPtr(ip);
         if (flags_ptr.fully_resolved) return true;
@@ -2225,13 +2225,13 @@ pub const LoadedStructType = struct {
         return false;
     }
 
-    pub fn clearFullyResolved(s: @This(), ip: *InternPool) void {
+    pub fn clearFullyResolved(s: LoadedStructType, ip: *InternPool) void {
         s.flagsPtr(ip).fully_resolved = false;
     }
 
     /// The returned pointer expires with any addition to the `InternPool`.
     /// Asserts the struct is not packed.
-    pub fn size(self: @This(), ip: *InternPool) *u32 {
+    pub fn size(self: LoadedStructType, ip: *InternPool) *u32 {
         assert(self.layout != .@"packed");
         const size_field_index = std.meta.fieldIndex(Tag.TypeStruct, "size").?;
         return @ptrCast(&ip.extra.items[self.extra_index + size_field_index]);
@@ -2241,50 +2241,50 @@ pub const LoadedStructType = struct {
     /// this type or the user specifies it, it is stored here. This will be
     /// set to `none` until the layout is resolved.
     /// Asserts the struct is packed.
-    pub fn backingIntType(s: @This(), ip: *const InternPool) *Index {
+    pub fn backingIntType(s: LoadedStructType, ip: *const InternPool) *Index {
         assert(s.layout == .@"packed");
         const field_index = std.meta.fieldIndex(Tag.TypeStructPacked, "backing_int_ty").?;
         return @ptrCast(&ip.extra.items[s.extra_index + field_index]);
     }
 
     /// Asserts the struct is not packed.
-    pub fn setZirIndex(s: @This(), ip: *InternPool, new_zir_index: TrackedInst.Index.Optional) void {
+    pub fn setZirIndex(s: LoadedStructType, ip: *InternPool, new_zir_index: TrackedInst.Index.Optional) void {
         assert(s.layout != .@"packed");
         const field_index = std.meta.fieldIndex(Tag.TypeStruct, "zir_index").?;
         ip.extra.items[s.extra_index + field_index] = @intFromEnum(new_zir_index);
     }
 
-    pub fn haveFieldTypes(s: @This(), ip: *const InternPool) bool {
+    pub fn haveFieldTypes(s: LoadedStructType, ip: *const InternPool) bool {
         const types = s.field_types.get(ip);
         return types.len == 0 or types[0] != .none;
     }
 
-    pub fn haveFieldInits(s: @This(), ip: *const InternPool) bool {
+    pub fn haveFieldInits(s: LoadedStructType, ip: *const InternPool) bool {
         return switch (s.layout) {
             .@"packed" => s.packedFlagsPtr(ip).inits_resolved,
             .auto, .@"extern" => s.flagsPtr(ip).inits_resolved,
         };
     }
 
-    pub fn setHaveFieldInits(s: @This(), ip: *InternPool) void {
+    pub fn setHaveFieldInits(s: LoadedStructType, ip: *InternPool) void {
         switch (s.layout) {
             .@"packed" => s.packedFlagsPtr(ip).inits_resolved = true,
             .auto, .@"extern" => s.flagsPtr(ip).inits_resolved = true,
         }
     }
 
-    pub fn haveLayout(s: @This(), ip: *InternPool) bool {
+    pub fn haveLayout(s: LoadedStructType, ip: *InternPool) bool {
         return switch (s.layout) {
             .@"packed" => s.backingIntType(ip).* != .none,
             .auto, .@"extern" => s.flagsPtr(ip).layout_resolved,
         };
     }
 
-    pub fn isTuple(s: @This(), ip: *InternPool) bool {
+    pub fn isTuple(s: LoadedStructType, ip: *InternPool) bool {
         return s.layout != .@"packed" and s.flagsPtr(ip).is_tuple;
     }
 
-    pub fn hasReorderedFields(s: @This()) bool {
+    pub fn hasReorderedFields(s: LoadedStructType) bool {
         return s.layout == .auto;
     }
 
@@ -2318,7 +2318,7 @@ pub const LoadedStructType = struct {
     /// Iterates over non-comptime fields in the order they are laid out in memory at runtime.
     /// May or may not include zero-bit fields.
     /// Asserts the struct is not packed.
-    pub fn iterateRuntimeOrder(s: @This(), ip: *InternPool) RuntimeOrderIterator {
+    pub fn iterateRuntimeOrder(s: LoadedStructType, ip: *InternPool) RuntimeOrderIterator {
         assert(s.layout != .@"packed");
         return .{
             .ip = ip,
@@ -2358,7 +2358,7 @@ pub const LoadedStructType = struct {
         }
     };
 
-    pub fn iterateRuntimeOrderReverse(s: @This(), ip: *InternPool) ReverseRuntimeOrderIterator {
+    pub fn iterateRuntimeOrderReverse(s: LoadedStructType, ip: *InternPool) ReverseRuntimeOrderIterator {
         assert(s.layout != .@"packed");
         return .{
             .ip = ip,
diff --git a/src/Module.zig b/src/Module.zig
index 320627a129..004ee89e91 100644
--- a/src/Module.zig
+++ b/src/Module.zig
@@ -6140,18 +6140,18 @@ pub const UnionLayout = struct {
     padding: u32,
 };
 
-pub fn getUnionLayout(mod: *Module, u: InternPool.LoadedUnionType) UnionLayout {
+pub fn getUnionLayout(mod: *Module, loaded_union: InternPool.LoadedUnionType) UnionLayout {
     const ip = &mod.intern_pool;
-    assert(u.haveLayout(ip));
+    assert(loaded_union.haveLayout(ip));
     var most_aligned_field: u32 = undefined;
     var most_aligned_field_size: u64 = undefined;
     var biggest_field: u32 = undefined;
     var payload_size: u64 = 0;
     var payload_align: Alignment = .@"1";
-    for (u.field_types.get(ip), 0..) |field_ty, i| {
+    for (loaded_union.field_types.get(ip), 0..) |field_ty, field_index| {
         if (!Type.fromInterned(field_ty).hasRuntimeBitsIgnoreComptime(mod)) continue;
 
-        const explicit_align = u.fieldAlign(ip, @intCast(i));
+        const explicit_align = loaded_union.fieldAlign(ip, field_index);
         const field_align = if (explicit_align != .none)
             explicit_align
         else
@@ -6159,16 +6159,16 @@ pub fn getUnionLayout(mod: *Module, u: InternPool.LoadedUnionType) UnionLayout {
         const field_size = Type.fromInterned(field_ty).abiSize(mod);
         if (field_size > payload_size) {
             payload_size = field_size;
-            biggest_field = @intCast(i);
+            biggest_field = @intCast(field_index);
         }
         if (field_align.compare(.gte, payload_align)) {
             payload_align = field_align;
-            most_aligned_field = @intCast(i);
+            most_aligned_field = @intCast(field_index);
             most_aligned_field_size = field_size;
         }
     }
-    const have_tag = u.flagsPtr(ip).runtime_tag.hasTag();
-    if (!have_tag or !Type.fromInterned(u.enum_tag_ty).hasRuntimeBits(mod)) {
+    const have_tag = loaded_union.flagsPtr(ip).runtime_tag.hasTag();
+    if (!have_tag or !Type.fromInterned(loaded_union.enum_tag_ty).hasRuntimeBits(mod)) {
         return .{
             .abi_size = payload_align.forward(payload_size),
             .abi_align = payload_align,
@@ -6183,10 +6183,10 @@ pub fn getUnionLayout(mod: *Module, u: InternPool.LoadedUnionType) UnionLayout {
         };
     }
 
-    const tag_size = Type.fromInterned(u.enum_tag_ty).abiSize(mod);
-    const tag_align = Type.fromInterned(u.enum_tag_ty).abiAlignment(mod).max(.@"1");
+    const tag_size = Type.fromInterned(loaded_union.enum_tag_ty).abiSize(mod);
+    const tag_align = Type.fromInterned(loaded_union.enum_tag_ty).abiAlignment(mod).max(.@"1");
     return .{
-        .abi_size = u.size(ip).*,
+        .abi_size = loaded_union.size(ip).*,
         .abi_align = tag_align.max(payload_align),
         .most_aligned_field = most_aligned_field,
         .most_aligned_field_size = most_aligned_field_size,
@@ -6195,24 +6195,24 @@ pub fn getUnionLayout(mod: *Module, u: InternPool.LoadedUnionType) UnionLayout {
         .payload_align = payload_align,
         .tag_align = tag_align,
         .tag_size = tag_size,
-        .padding = u.padding(ip).*,
+        .padding = loaded_union.padding(ip).*,
     };
 }
 
-pub fn unionAbiSize(mod: *Module, u: InternPool.LoadedUnionType) u64 {
-    return mod.getUnionLayout(u).abi_size;
+pub fn unionAbiSize(mod: *Module, loaded_union: InternPool.LoadedUnionType) u64 {
+    return mod.getUnionLayout(loaded_union).abi_size;
 }
 
 /// Returns 0 if the union is represented with 0 bits at runtime.
-pub fn unionAbiAlignment(mod: *Module, u: InternPool.LoadedUnionType) Alignment {
+pub fn unionAbiAlignment(mod: *Module, loaded_union: InternPool.LoadedUnionType) Alignment {
     const ip = &mod.intern_pool;
-    const have_tag = u.flagsPtr(ip).runtime_tag.hasTag();
+    const have_tag = loaded_union.flagsPtr(ip).runtime_tag.hasTag();
     var max_align: Alignment = .none;
-    if (have_tag) max_align = Type.fromInterned(u.enum_tag_ty).abiAlignment(mod);
-    for (u.field_types.get(ip), 0..) |field_ty, field_index| {
+    if (have_tag) max_align = Type.fromInterned(loaded_union.enum_tag_ty).abiAlignment(mod);
+    for (loaded_union.field_types.get(ip), 0..) |field_ty, field_index| {
         if (!Type.fromInterned(field_ty).hasRuntimeBits(mod)) continue;
 
-        const field_align = mod.unionFieldNormalAlignment(u, @intCast(field_index));
+        const field_align = mod.unionFieldNormalAlignment(loaded_union, @intCast(field_index));
         max_align = max_align.max(field_align);
     }
     return max_align;
@@ -6221,20 +6221,20 @@ pub fn unionAbiAlignment(mod: *Module, u: InternPool.LoadedUnionType) Alignment
 /// Returns the field alignment, assuming the union is not packed.
 /// Keep implementation in sync with `Sema.unionFieldAlignment`.
 /// Prefer to call that function instead of this one during Sema.
-pub fn unionFieldNormalAlignment(mod: *Module, u: InternPool.LoadedUnionType, field_index: u32) Alignment {
+pub fn unionFieldNormalAlignment(mod: *Module, loaded_union: InternPool.LoadedUnionType, field_index: u32) Alignment {
     const ip = &mod.intern_pool;
-    const field_align = u.fieldAlign(ip, field_index);
+    const field_align = loaded_union.fieldAlign(ip, field_index);
     if (field_align != .none) return field_align;
-    const field_ty = Type.fromInterned(u.field_types.get(ip)[field_index]);
+    const field_ty = Type.fromInterned(loaded_union.field_types.get(ip)[field_index]);
     return field_ty.abiAlignment(mod);
 }
 
 /// Returns the index of the active field, given the current tag value
-pub fn unionTagFieldIndex(mod: *Module, u: InternPool.LoadedUnionType, enum_tag: Value) ?u32 {
+pub fn unionTagFieldIndex(mod: *Module, loaded_union: InternPool.LoadedUnionType, enum_tag: Value) ?u32 {
     const ip = &mod.intern_pool;
     if (enum_tag.toIntern() == .none) return null;
-    assert(ip.typeOf(enum_tag.toIntern()) == u.enum_tag_ty);
-    return u.loadTagType(ip).tagValueIndex(ip, enum_tag.toIntern());
+    assert(ip.typeOf(enum_tag.toIntern()) == loaded_union.enum_tag_ty);
+    return loaded_union.loadTagType(ip).tagValueIndex(ip, enum_tag.toIntern());
 }
 
 /// Returns the field alignment of a non-packed struct in byte units.
diff --git a/src/Sema.zig b/src/Sema.zig
index d80cf81888..dd45a2db2c 100644
--- a/src/Sema.zig
+++ b/src/Sema.zig
@@ -35405,7 +35405,7 @@ pub fn resolveUnionAlignment(
         const field_ty = Type.fromInterned(union_type.field_types.get(ip)[field_index]);
         if (!(try sema.typeHasRuntimeBits(field_ty))) continue;
 
-        const explicit_align = union_type.fieldAlign(ip, @intCast(field_index));
+        const explicit_align = union_type.fieldAlign(ip, field_index);
         const field_align = if (explicit_align != .none)
             explicit_align
         else
@@ -35465,7 +35465,7 @@ fn resolveUnionLayout(sema: *Sema, ty: Type) CompileError!void {
             else => return err,
         });
 
-        const explicit_align = union_type.fieldAlign(ip, @intCast(field_index));
+        const explicit_align = union_type.fieldAlign(ip, field_index);
         const field_align = if (explicit_align != .none)
             explicit_align
         else
diff --git a/src/arch/x86_64/CodeGen.zig b/src/arch/x86_64/CodeGen.zig
index 0681bc975d..7faf3d6d50 100644
--- a/src/arch/x86_64/CodeGen.zig
+++ b/src/arch/x86_64/CodeGen.zig
@@ -14316,7 +14316,7 @@ fn moveStrategy(self: *Self, ty: Type, class: Register.Class, aligned: bool) !Mo
         .mmx => {},
         .sse => switch (ty.zigTypeTag(mod)) {
             else => {
-                const classes = mem.sliceTo(&abi.classifySystemV(ty, mod, .other), .none);
+                const classes = mem.sliceTo(&abi.classifySystemV(ty, mod, self.target.*, .other), .none);
                 assert(std.mem.indexOfNone(abi.Class, classes, &.{
                     .integer, .sse, .memory, .float, .float_combine,
                 }) == null);
@@ -18450,7 +18450,7 @@ fn airVaArg(self: *Self, inst: Air.Inst.Index) !void {
             const overflow_arg_area: MCValue = .{ .indirect = .{ .reg = ptr_arg_list_reg, .off = 8 } };
             const reg_save_area: MCValue = .{ .indirect = .{ .reg = ptr_arg_list_reg, .off = 16 } };
 
-            const classes = mem.sliceTo(&abi.classifySystemV(promote_ty, mod, .arg), .none);
+            const classes = mem.sliceTo(&abi.classifySystemV(promote_ty, mod, self.target.*, .arg), .none);
             switch (classes[0]) {
                 .integer => {
                     assert(classes.len == 1);
@@ -18800,7 +18800,7 @@ fn resolveCallingConventionValues(
                 var ret_tracking_i: usize = 0;
 
                 const classes = switch (resolved_cc) {
-                    .SysV => mem.sliceTo(&abi.classifySystemV(ret_ty, mod, .ret), .none),
+                    .SysV => mem.sliceTo(&abi.classifySystemV(ret_ty, mod, self.target.*, .ret), .none),
                     .Win64 => &.{abi.classifyWindows(ret_ty, mod)},
                     else => unreachable,
                 };
@@ -18875,7 +18875,7 @@ fn resolveCallingConventionValues(
                 var arg_mcv_i: usize = 0;
 
                 const classes = switch (resolved_cc) {
-                    .SysV => mem.sliceTo(&abi.classifySystemV(ty, mod, .arg), .none),
+                    .SysV => mem.sliceTo(&abi.classifySystemV(ty, mod, self.target.*, .arg), .none),
                     .Win64 => &.{abi.classifyWindows(ty, mod)},
                     else => unreachable,
                 };
@@ -19090,7 +19090,7 @@ fn memSize(self: *Self, ty: Type) Memory.Size {
 
 fn splitType(self: *Self, ty: Type) ![2]Type {
     const mod = self.bin_file.comp.module.?;
-    const classes = mem.sliceTo(&abi.classifySystemV(ty, mod, .other), .none);
+    const classes = mem.sliceTo(&abi.classifySystemV(ty, mod, self.target.*, .other), .none);
     var parts: [2]Type = undefined;
     if (classes.len == 2) for (&parts, classes, 0..) |*part, class, part_i| {
         part.* = switch (class) {
diff --git a/src/arch/x86_64/abi.zig b/src/arch/x86_64/abi.zig
index c3ab91af8d..8c81ad9ed2 100644
--- a/src/arch/x86_64/abi.zig
+++ b/src/arch/x86_64/abi.zig
@@ -11,6 +11,37 @@ pub const Class = enum {
     float,
     float_combine,
     integer_per_element,
+
+    fn isX87(class: Class) bool {
+        return switch (class) {
+            .x87, .x87up, .complex_x87 => true,
+            else => false,
+        };
+    }
+
+    /// Combine a field class with the prev one.
+    fn combineSystemV(prev_class: Class, next_class: Class) Class {
+        // "If both classes are equal, this is the resulting class."
+        if (prev_class == next_class)
+            return if (prev_class == .float) .float_combine else prev_class;
+
+        // "If one of the classes is NO_CLASS, the resulting class
+        // is the other class."
+        if (prev_class == .none) return next_class;
+
+        // "If one of the classes is MEMORY, the result is the MEMORY class."
+        if (prev_class == .memory or next_class == .memory) return .memory;
+
+        // "If one of the classes is INTEGER, the result is the INTEGER."
+        if (prev_class == .integer or next_class == .integer) return .integer;
+
+        // "If one of the classes is X87, X87UP, COMPLEX_X87 class,
+        // MEMORY is used as class."
+        if (prev_class.isX87() or next_class.isX87()) return .memory;
+
+        // "Otherwise class SSE is used."
+        return .sse;
+    }
 };
 
 pub fn classifyWindows(ty: Type, zcu: *Zcu) Class {
@@ -69,9 +100,7 @@ pub const Context = enum { ret, arg, field, other };
 
 /// There are a maximum of 8 possible return slots. Returned values are in
 /// the beginning of the array; unused slots are filled with .none.
-pub fn classifySystemV(ty: Type, zcu: *Zcu, ctx: Context) [8]Class {
-    const ip = &zcu.intern_pool;
-    const target = zcu.getTarget();
+pub fn classifySystemV(ty: Type, zcu: *Zcu, target: std.Target, ctx: Context) [8]Class {
     const memory_class = [_]Class{
         .memory, .none, .none, .none,
         .none,   .none, .none, .none,
@@ -231,121 +260,30 @@ pub fn classifySystemV(ty: Type, zcu: *Zcu, ctx: Context) [8]Class {
             }
             return memory_class;
         },
-        .Struct => {
+        .Struct, .Union => {
             // "If the size of an object is larger than eight eightbytes, or
             // it contains unaligned fields, it has class MEMORY"
             // "If the size of the aggregate exceeds a single eightbyte, each is classified
             // separately.".
-            const loaded_struct = ip.loadStructType(ty.toIntern());
             const ty_size = ty.abiSize(zcu);
-            if (loaded_struct.layout == .@"packed") {
-                assert(ty_size <= 16);
-                result[0] = .integer;
-                if (ty_size > 8) result[1] = .integer;
-                return result;
-            }
-            if (ty_size > 64)
-                return memory_class;
-
-            var byte_offset: u64 = 0;
-            classifySystemVStruct(&result, &byte_offset, loaded_struct, zcu);
-
-            // Post-merger cleanup
-
-            // "If one of the classes is MEMORY, the whole argument is passed in memory"
-            // "If X87UP is not preceded by X87, the whole argument is passed in memory."
-            var found_sseup = false;
-            for (result, 0..) |item, i| switch (item) {
-                .memory => return memory_class,
-                .x87up => if (i == 0 or result[i - 1] != .x87) return memory_class,
-                .sseup => found_sseup = true,
-                else => continue,
-            };
-            // "If the size of the aggregate exceeds two eightbytes and the first eight-
-            // byte isn’t SSE or any other eightbyte isn’t SSEUP, the whole argument
-            // is passed in memory."
-            if (ty_size > 16 and (result[0] != .sse or !found_sseup)) return memory_class;
-
-            // "If SSEUP is not preceded by SSE or SSEUP, it is converted to SSE."
-            for (&result, 0..) |*item, i| {
-                if (item.* == .sseup) switch (result[i - 1]) {
-                    .sse, .sseup => continue,
-                    else => item.* = .sse,
-                };
-            }
-            return result;
-        },
-        .Union => {
-            // "If the size of an object is larger than eight eightbytes, or
-            // it contains unaligned fields, it has class MEMORY"
-            // "If the size of the aggregate exceeds a single eightbyte, each is classified
-            // separately.".
-            const union_obj = zcu.typeToUnion(ty).?;
-            const ty_size = zcu.unionAbiSize(union_obj);
-            if (union_obj.getLayout(ip) == .@"packed") {
-                assert(ty_size <= 16);
-                result[0] = .integer;
-                if (ty_size > 8) result[1] = .integer;
-                return result;
+            switch (ty.containerLayout(zcu)) {
+                .auto, .@"extern" => {},
+                .@"packed" => {
+                    assert(ty_size <= 16);
+                    result[0] = .integer;
+                    if (ty_size > 8) result[1] = .integer;
+                    return result;
+                },
             }
             if (ty_size > 64)
                 return memory_class;
 
-            for (union_obj.field_types.get(ip), 0..) |field_ty, field_index| {
-                const field_align = union_obj.fieldAlign(ip, @intCast(field_index));
-                if (field_align != .none and
-                    field_align.compare(.lt, Type.fromInterned(field_ty).abiAlignment(zcu)))
-                {
-                    return memory_class;
-                }
-                // Combine this field with the previous one.
-                const field_class = classifySystemV(Type.fromInterned(field_ty), zcu, .field);
-                for (&result, 0..) |*result_item, i| {
-                    const field_item = field_class[i];
-                    // "If both classes are equal, this is the resulting class."
-                    if (result_item.* == field_item) {
-                        continue;
-                    }
-
-                    // "If one of the classes is NO_CLASS, the resulting class
-                    // is the other class."
-                    if (result_item.* == .none) {
-                        result_item.* = field_item;
-                        continue;
-                    }
-                    if (field_item == .none) {
-                        continue;
-                    }
-
-                    // "If one of the classes is MEMORY, the result is the MEMORY class."
-                    if (result_item.* == .memory or field_item == .memory) {
-                        result_item.* = .memory;
-                        continue;
-                    }
-
-                    // "If one of the classes is INTEGER, the result is the INTEGER."
-                    if (result_item.* == .integer or field_item == .integer) {
-                        result_item.* = .integer;
-                        continue;
-                    }
-
-                    // "If one of the classes is X87, X87UP, COMPLEX_X87 class,
-                    // MEMORY is used as class."
-                    if (result_item.* == .x87 or
-                        result_item.* == .x87up or
-                        result_item.* == .complex_x87 or
-                        field_item == .x87 or
-                        field_item == .x87up or
-                        field_item == .complex_x87)
-                    {
-                        result_item.* = .memory;
-                        continue;
-                    }
-
-                    // "Otherwise class SSE is used."
-                    result_item.* = .sse;
-                }
-            }
+            _ = if (zcu.typeToStruct(ty)) |loaded_struct|
+                classifySystemVStruct(&result, 0, loaded_struct, zcu, target)
+            else if (zcu.typeToUnion(ty)) |loaded_union|
+                classifySystemVUnion(&result, 0, loaded_union, zcu, target)
+            else
+                unreachable;
 
             // Post-merger cleanup
 
@@ -391,78 +329,85 @@ pub fn classifySystemV(ty: Type, zcu: *Zcu, ctx: Context) [8]Class {
 
 fn classifySystemVStruct(
     result: *[8]Class,
-    byte_offset: *u64,
+    starting_byte_offset: u64,
     loaded_struct: InternPool.LoadedStructType,
     zcu: *Zcu,
-) void {
+    target: std.Target,
+) u64 {
     const ip = &zcu.intern_pool;
+    var byte_offset = starting_byte_offset;
     var field_it = loaded_struct.iterateRuntimeOrder(ip);
     while (field_it.next()) |field_index| {
         const field_ty = Type.fromInterned(loaded_struct.field_types.get(ip)[field_index]);
         const field_align = loaded_struct.fieldAlign(ip, field_index);
-        byte_offset.* = std.mem.alignForward(
+        byte_offset = std.mem.alignForward(
             u64,
-            byte_offset.*,
+            byte_offset,
             field_align.toByteUnits() orelse field_ty.abiAlignment(zcu).toByteUnits().?,
         );
         if (zcu.typeToStruct(field_ty)) |field_loaded_struct| {
-            if (field_loaded_struct.layout != .@"packed") {
-                classifySystemVStruct(result, byte_offset, field_loaded_struct, zcu);
-                continue;
-            }
-        }
-        const field_class = std.mem.sliceTo(&classifySystemV(field_ty, zcu, .field), .none);
-        const field_size = field_ty.abiSize(zcu);
-        combine: {
-            // Combine this field with the previous one.
-            const result_class = &result[@intCast(byte_offset.* / 8)];
-            // "If both classes are equal, this is the resulting class."
-            if (result_class.* == field_class[0]) {
-                if (result_class.* == .float) {
-                    result_class.* = .float_combine;
-                }
-                break :combine;
-            }
-
-            // "If one of the classes is NO_CLASS, the resulting class
-            // is the other class."
-            if (result_class.* == .none) {
-                result_class.* = field_class[0];
-                break :combine;
+            switch (field_loaded_struct.layout) {
+                .auto, .@"extern" => {
+                    byte_offset = classifySystemVStruct(result, byte_offset, field_loaded_struct, zcu, target);
+                    continue;
+                },
+                .@"packed" => {},
             }
-            assert(field_class[0] != .none);
-
-            // "If one of the classes is MEMORY, the result is the MEMORY class."
-            if (result_class.* == .memory or field_class[0] == .memory) {
-                result_class.* = .memory;
-                break :combine;
+        } else if (zcu.typeToUnion(field_ty)) |field_loaded_union| {
+            switch (field_loaded_union.getLayout(ip)) {
+                .auto, .@"extern" => {
+                    byte_offset = classifySystemVUnion(result, byte_offset, field_loaded_union, zcu, target);
+                    continue;
+                },
+                .@"packed" => {},
             }
+        }
+        const field_classes = std.mem.sliceTo(&classifySystemV(field_ty, zcu, target, .field), .none);
+        for (result[@intCast(byte_offset / 8)..][0..field_classes.len], field_classes) |*result_class, field_class|
+            result_class.* = result_class.combineSystemV(field_class);
+        byte_offset += field_ty.abiSize(zcu);
+    }
+    const final_byte_offset = starting_byte_offset + loaded_struct.size(ip).*;
+    std.debug.assert(final_byte_offset == std.mem.alignForward(
+        u64,
+        byte_offset,
+        loaded_struct.flagsPtr(ip).alignment.toByteUnits().?,
+    ));
+    return final_byte_offset;
+}
 
-            // "If one of the classes is INTEGER, the result is the INTEGER."
-            if (result_class.* == .integer or field_class[0] == .integer) {
-                result_class.* = .integer;
-                break :combine;
+fn classifySystemVUnion(
+    result: *[8]Class,
+    starting_byte_offset: u64,
+    loaded_union: InternPool.LoadedUnionType,
+    zcu: *Zcu,
+    target: std.Target,
+) u64 {
+    const ip = &zcu.intern_pool;
+    for (0..loaded_union.field_types.len) |field_index| {
+        const field_ty = Type.fromInterned(loaded_union.field_types.get(ip)[field_index]);
+        if (zcu.typeToStruct(field_ty)) |field_loaded_struct| {
+            switch (field_loaded_struct.layout) {
+                .auto, .@"extern" => {
+                    _ = classifySystemVStruct(result, starting_byte_offset, field_loaded_struct, zcu, target);
+                    continue;
+                },
+                .@"packed" => {},
             }
-
-            // "If one of the classes is X87, X87UP, COMPLEX_X87 class,
-            // MEMORY is used as class."
-            if (result_class.* == .x87 or
-                result_class.* == .x87up or
-                result_class.* == .complex_x87 or
-                field_class[0] == .x87 or
-                field_class[0] == .x87up or
-                field_class[0] == .complex_x87)
-            {
-                result_class.* = .memory;
-                break :combine;
+        } else if (zcu.typeToUnion(field_ty)) |field_loaded_union| {
+            switch (field_loaded_union.getLayout(ip)) {
+                .auto, .@"extern" => {
+                    _ = classifySystemVUnion(result, starting_byte_offset, field_loaded_union, zcu, target);
+                    continue;
+                },
+                .@"packed" => {},
             }
-
-            // "Otherwise class SSE is used."
-            result_class.* = .sse;
         }
-        @memcpy(result[@intCast(byte_offset.* / 8 + 1)..][0 .. field_class.len - 1], field_class[1..]);
-        byte_offset.* += field_size;
+        const field_classes = std.mem.sliceTo(&classifySystemV(field_ty, zcu, target, .field), .none);
+        for (result[@intCast(starting_byte_offset / 8)..][0..field_classes.len], field_classes) |*result_class, field_class|
+            result_class.* = result_class.combineSystemV(field_class);
     }
+    return starting_byte_offset + loaded_union.size(ip).*;
 }
 
 pub const SysV = struct {
diff --git a/src/codegen/c/Type.zig b/src/codegen/c/Type.zig
index 357677522c..5068073f32 100644
--- a/src/codegen/c/Type.zig
+++ b/src/codegen/c/Type.zig
@@ -1858,7 +1858,7 @@ pub const Pool = struct {
                                     loaded_tag.names.get(ip)[field_index].toSlice(ip),
                                 );
                                 const field_alignas = AlignAs.fromAlignment(.{
-                                    .@"align" = loaded_union.fieldAlign(ip, @intCast(field_index)),
+                                    .@"align" = loaded_union.fieldAlign(ip, field_index),
                                     .abi = field_type.abiAlignment(zcu),
                                 });
                                 pool.addHashedExtraAssumeCapacityTo(scratch, &hasher, Field, .{
diff --git a/src/codegen/llvm.zig b/src/codegen/llvm.zig
index f8fce85c80..a3c7fb0b86 100644
--- a/src/codegen/llvm.zig
+++ b/src/codegen/llvm.zig
@@ -1384,7 +1384,7 @@ pub const Object = struct {
         const namespace = zcu.namespacePtr(decl.src_namespace);
         const owner_mod = namespace.file_scope.mod;
         const fn_info = zcu.typeToFunc(decl.typeOf(zcu)).?;
-        const target = zcu.getTarget();
+        const target = owner_mod.resolved_target.result;
         const ip = &zcu.intern_pool;
 
         var dg: DeclGen = .{
@@ -1456,7 +1456,7 @@ pub const Object = struct {
         var llvm_arg_i: u32 = 0;
 
         // This gets the LLVM values from the function and stores them in `dg.args`.
-        const sret = firstParamSRet(fn_info, zcu);
+        const sret = firstParamSRet(fn_info, zcu, target);
         const ret_ptr: Builder.Value = if (sret) param: {
             const param = wip.arg(llvm_arg_i);
             llvm_arg_i += 1;
@@ -2755,7 +2755,7 @@ pub const Object = struct {
 
                 // Return type goes first.
                 if (Type.fromInterned(fn_info.return_type).hasRuntimeBitsIgnoreComptime(mod)) {
-                    const sret = firstParamSRet(fn_info, mod);
+                    const sret = firstParamSRet(fn_info, mod, target);
                     const ret_ty = if (sret) Type.void else Type.fromInterned(fn_info.return_type);
                     debug_param_types.appendAssumeCapacity(try o.lowerDebugType(ret_ty));
 
@@ -2881,7 +2881,7 @@ pub const Object = struct {
         assert(decl.has_tv);
         const fn_info = zcu.typeToFunc(zig_fn_type).?;
         const target = owner_mod.resolved_target.result;
-        const sret = firstParamSRet(fn_info, zcu);
+        const sret = firstParamSRet(fn_info, zcu, target);
 
         const is_extern = decl.isExtern(zcu);
         const function_index = try o.builder.addFunction(
@@ -3604,7 +3604,7 @@ pub const Object = struct {
         var llvm_params = std.ArrayListUnmanaged(Builder.Type){};
         defer llvm_params.deinit(o.gpa);
 
-        if (firstParamSRet(fn_info, mod)) {
+        if (firstParamSRet(fn_info, mod, target)) {
             try llvm_params.append(o.gpa, .ptr);
         }
 
@@ -5130,7 +5130,7 @@ pub const FuncGen = struct {
         const return_type = Type.fromInterned(fn_info.return_type);
         const llvm_fn = try self.resolveInst(pl_op.operand);
         const target = mod.getTarget();
-        const sret = firstParamSRet(fn_info, mod);
+        const sret = firstParamSRet(fn_info, mod, target);
 
         var llvm_args = std.ArrayList(Builder.Value).init(self.gpa);
         defer llvm_args.deinit();
@@ -10865,38 +10865,38 @@ fn toLlvmGlobalAddressSpace(wanted_address_space: std.builtin.AddressSpace, targ
     };
 }
 
-fn firstParamSRet(fn_info: InternPool.Key.FuncType, mod: *Module) bool {
+fn firstParamSRet(fn_info: InternPool.Key.FuncType, zcu: *Zcu, target: std.Target) bool {
     const return_type = Type.fromInterned(fn_info.return_type);
-    if (!return_type.hasRuntimeBitsIgnoreComptime(mod)) return false;
+    if (!return_type.hasRuntimeBitsIgnoreComptime(zcu)) return false;
 
-    const target = mod.getTarget();
-    switch (fn_info.cc) {
-        .Unspecified, .Inline => return isByRef(return_type, mod),
+    return switch (fn_info.cc) {
+        .Unspecified, .Inline => isByRef(return_type, zcu),
         .C => switch (target.cpu.arch) {
-            .mips, .mipsel => return false,
+            .mips, .mipsel => false,
+            .x86 => isByRef(return_type, zcu),
             .x86_64 => switch (target.os.tag) {
-                .windows => return x86_64_abi.classifyWindows(return_type, mod) == .memory,
-                else => return firstParamSRetSystemV(return_type, mod),
+                .windows => x86_64_abi.classifyWindows(return_type, zcu) == .memory,
+                else => firstParamSRetSystemV(return_type, zcu, target),
             },
-            .wasm32 => return wasm_c_abi.classifyType(return_type, mod)[0] == .indirect,
-            .aarch64, .aarch64_be => return aarch64_c_abi.classifyType(return_type, mod) == .memory,
-            .arm, .armeb => switch (arm_c_abi.classifyType(return_type, mod, .ret)) {
-                .memory, .i64_array => return true,
-                .i32_array => |size| return size != 1,
-                .byval => return false,
+            .wasm32 => wasm_c_abi.classifyType(return_type, zcu)[0] == .indirect,
+            .aarch64, .aarch64_be => aarch64_c_abi.classifyType(return_type, zcu) == .memory,
+            .arm, .armeb => switch (arm_c_abi.classifyType(return_type, zcu, .ret)) {
+                .memory, .i64_array => true,
+                .i32_array => |size| size != 1,
+                .byval => false,
             },
-            .riscv32, .riscv64 => return riscv_c_abi.classifyType(return_type, mod) == .memory,
-            else => return false, // TODO investigate C ABI for other architectures
+            .riscv32, .riscv64 => riscv_c_abi.classifyType(return_type, zcu) == .memory,
+            else => false, // TODO investigate C ABI for other architectures
         },
-        .SysV => return firstParamSRetSystemV(return_type, mod),
-        .Win64 => return x86_64_abi.classifyWindows(return_type, mod) == .memory,
-        .Stdcall => return !isScalar(mod, return_type),
-        else => return false,
-    }
+        .SysV => firstParamSRetSystemV(return_type, zcu, target),
+        .Win64 => x86_64_abi.classifyWindows(return_type, zcu) == .memory,
+        .Stdcall => !isScalar(zcu, return_type),
+        else => false,
+    };
 }
 
-fn firstParamSRetSystemV(ty: Type, mod: *Module) bool {
-    const class = x86_64_abi.classifySystemV(ty, mod, .ret);
+fn firstParamSRetSystemV(ty: Type, zcu: *Zcu, target: std.Target) bool {
+    const class = x86_64_abi.classifySystemV(ty, zcu, target, .ret);
     if (class[0] == .memory) return true;
     if (class[0] == .x87 and class[2] != .none) return true;
     return false;
@@ -10922,6 +10922,7 @@ fn lowerFnRetTy(o: *Object, fn_info: InternPool.Key.FuncType) Allocator.Error!Bu
         .C => {
             switch (target.cpu.arch) {
                 .mips, .mipsel => return o.lowerType(return_type),
+                .x86 => return if (isByRef(return_type, mod)) .void else o.lowerType(return_type),
                 .x86_64 => switch (target.os.tag) {
                     .windows => return lowerWin64FnRetTy(o, fn_info),
                     else => return lowerSystemVFnRetTy(o, fn_info),
@@ -11014,7 +11015,8 @@ fn lowerSystemVFnRetTy(o: *Object, fn_info: InternPool.Key.FuncType) Allocator.E
     if (isScalar(mod, return_type)) {
         return o.lowerType(return_type);
     }
-    const classes = x86_64_abi.classifySystemV(return_type, mod, .ret);
+    const target = mod.getTarget();
+    const classes = x86_64_abi.classifySystemV(return_type, mod, target, .ret);
     if (classes[0] == .memory) return .void;
     var types_index: u32 = 0;
     var types_buffer: [8]Builder.Type = undefined;
@@ -11098,8 +11100,8 @@ const ParamTypeIterator = struct {
 
     pub fn next(it: *ParamTypeIterator) Allocator.Error!?Lowering {
         if (it.zig_index >= it.fn_info.param_types.len) return null;
-        const mod = it.object.module;
-        const ip = &mod.intern_pool;
+        const zcu = it.object.module;
+        const ip = &zcu.intern_pool;
         const ty = it.fn_info.param_types.get(ip)[it.zig_index];
         it.byval_attr = false;
         return nextInner(it, Type.fromInterned(ty));
@@ -11107,8 +11109,8 @@ const ParamTypeIterator = struct {
 
     /// `airCall` uses this instead of `next` so that it can take into account variadic functions.
     pub fn nextCall(it: *ParamTypeIterator, fg: *FuncGen, args: []const Air.Inst.Ref) Allocator.Error!?Lowering {
-        const mod = it.object.module;
-        const ip = &mod.intern_pool;
+        const zcu = it.object.module;
+        const ip = &zcu.intern_pool;
         if (it.zig_index >= it.fn_info.param_types.len) {
             if (it.zig_index >= args.len) {
                 return null;
@@ -11121,10 +11123,10 @@ const ParamTypeIterator = struct {
     }
 
     fn nextInner(it: *ParamTypeIterator, ty: Type) Allocator.Error!?Lowering {
-        const mod = it.object.module;
-        const target = mod.getTarget();
+        const zcu = it.object.module;
+        const target = zcu.getTarget();
 
-        if (!ty.hasRuntimeBitsIgnoreComptime(mod)) {
+        if (!ty.hasRuntimeBitsIgnoreComptime(zcu)) {
             it.zig_index += 1;
             return .no_bits;
         }
@@ -11132,12 +11134,12 @@ const ParamTypeIterator = struct {
             .Unspecified, .Inline => {
                 it.zig_index += 1;
                 it.llvm_index += 1;
-                if (ty.isSlice(mod) or
-                    (ty.zigTypeTag(mod) == .Optional and ty.optionalChild(mod).isSlice(mod) and !ty.ptrAllowsZero(mod)))
+                if (ty.isSlice(zcu) or
+                    (ty.zigTypeTag(zcu) == .Optional and ty.optionalChild(zcu).isSlice(zcu) and !ty.ptrAllowsZero(zcu)))
                 {
                     it.llvm_index += 1;
                     return .slice;
-                } else if (isByRef(ty, mod)) {
+                } else if (isByRef(ty, zcu)) {
                     return .byref;
                 } else {
                     return .byval;
@@ -11146,87 +11148,85 @@ const ParamTypeIterator = struct {
             .Async => {
                 @panic("TODO implement async function lowering in the LLVM backend");
             },
-            .C => {
-                switch (target.cpu.arch) {
-                    .mips, .mipsel => {
-                        it.zig_index += 1;
-                        it.llvm_index += 1;
+            .C => switch (target.cpu.arch) {
+                .mips, .mipsel => {
+                    it.zig_index += 1;
+                    it.llvm_index += 1;
+                    return .byval;
+                },
+                .x86_64 => switch (target.os.tag) {
+                    .windows => return it.nextWin64(ty),
+                    else => return it.nextSystemV(ty),
+                },
+                .wasm32 => {
+                    it.zig_index += 1;
+                    it.llvm_index += 1;
+                    if (isScalar(zcu, ty)) {
                         return .byval;
-                    },
-                    .x86_64 => switch (target.os.tag) {
-                        .windows => return it.nextWin64(ty),
-                        else => return it.nextSystemV(ty),
-                    },
-                    .wasm32 => {
-                        it.zig_index += 1;
-                        it.llvm_index += 1;
-                        if (isScalar(mod, ty)) {
-                            return .byval;
-                        }
-                        const classes = wasm_c_abi.classifyType(ty, mod);
-                        if (classes[0] == .indirect) {
+                    }
+                    const classes = wasm_c_abi.classifyType(ty, zcu);
+                    if (classes[0] == .indirect) {
+                        return .byref;
+                    }
+                    return .abi_sized_int;
+                },
+                .aarch64, .aarch64_be => {
+                    it.zig_index += 1;
+                    it.llvm_index += 1;
+                    switch (aarch64_c_abi.classifyType(ty, zcu)) {
+                        .memory => return .byref_mut,
+                        .float_array => |len| return Lowering{ .float_array = len },
+                        .byval => return .byval,
+                        .integer => {
+                            it.types_len = 1;
+                            it.types_buffer[0] = .i64;
+                            return .multiple_llvm_types;
+                        },
+                        .double_integer => return Lowering{ .i64_array = 2 },
+                    }
+                },
+                .arm, .armeb => {
+                    it.zig_index += 1;
+                    it.llvm_index += 1;
+                    switch (arm_c_abi.classifyType(ty, zcu, .arg)) {
+                        .memory => {
+                            it.byval_attr = true;
                             return .byref;
-                        }
-                        return .abi_sized_int;
-                    },
-                    .aarch64, .aarch64_be => {
-                        it.zig_index += 1;
-                        it.llvm_index += 1;
-                        switch (aarch64_c_abi.classifyType(ty, mod)) {
-                            .memory => return .byref_mut,
-                            .float_array => |len| return Lowering{ .float_array = len },
-                            .byval => return .byval,
-                            .integer => {
-                                it.types_len = 1;
-                                it.types_buffer[0] = .i64;
-                                return .multiple_llvm_types;
-                            },
-                            .double_integer => return Lowering{ .i64_array = 2 },
-                        }
-                    },
-                    .arm, .armeb => {
-                        it.zig_index += 1;
-                        it.llvm_index += 1;
-                        switch (arm_c_abi.classifyType(ty, mod, .arg)) {
-                            .memory => {
-                                it.byval_attr = true;
-                                return .byref;
-                            },
-                            .byval => return .byval,
-                            .i32_array => |size| return Lowering{ .i32_array = size },
-                            .i64_array => |size| return Lowering{ .i64_array = size },
-                        }
-                    },
-                    .riscv32, .riscv64 => {
-                        it.zig_index += 1;
-                        it.llvm_index += 1;
-                        if (ty.toIntern() == .f16_type and
-                            !std.Target.riscv.featureSetHas(target.cpu.features, .d)) return .as_u16;
-                        switch (riscv_c_abi.classifyType(ty, mod)) {
-                            .memory => return .byref_mut,
-                            .byval => return .byval,
-                            .integer => return .abi_sized_int,
-                            .double_integer => return Lowering{ .i64_array = 2 },
-                            .fields => {
-                                it.types_len = 0;
-                                for (0..ty.structFieldCount(mod)) |field_index| {
-                                    const field_ty = ty.structFieldType(field_index, mod);
-                                    if (!field_ty.hasRuntimeBitsIgnoreComptime(mod)) continue;
-                                    it.types_buffer[it.types_len] = try it.object.lowerType(field_ty);
-                                    it.types_len += 1;
-                                }
-                                it.llvm_index += it.types_len - 1;
-                                return .multiple_llvm_types;
-                            },
-                        }
-                    },
-                    // TODO investigate C ABI for other architectures
-                    else => {
-                        it.zig_index += 1;
-                        it.llvm_index += 1;
-                        return .byval;
-                    },
-                }
+                        },
+                        .byval => return .byval,
+                        .i32_array => |size| return Lowering{ .i32_array = size },
+                        .i64_array => |size| return Lowering{ .i64_array = size },
+                    }
+                },
+                .riscv32, .riscv64 => {
+                    it.zig_index += 1;
+                    it.llvm_index += 1;
+                    if (ty.toIntern() == .f16_type and
+                        !std.Target.riscv.featureSetHas(target.cpu.features, .d)) return .as_u16;
+                    switch (riscv_c_abi.classifyType(ty, zcu)) {
+                        .memory => return .byref_mut,
+                        .byval => return .byval,
+                        .integer => return .abi_sized_int,
+                        .double_integer => return Lowering{ .i64_array = 2 },
+                        .fields => {
+                            it.types_len = 0;
+                            for (0..ty.structFieldCount(zcu)) |field_index| {
+                                const field_ty = ty.structFieldType(field_index, zcu);
+                                if (!field_ty.hasRuntimeBitsIgnoreComptime(zcu)) continue;
+                                it.types_buffer[it.types_len] = try it.object.lowerType(field_ty);
+                                it.types_len += 1;
+                            }
+                            it.llvm_index += it.types_len - 1;
+                            return .multiple_llvm_types;
+                        },
+                    }
+                },
+                // TODO investigate C ABI for other architectures
+                else => {
+                    it.zig_index += 1;
+                    it.llvm_index += 1;
+                    return .byval;
+                },
             },
             .Win64 => return it.nextWin64(ty),
             .SysV => return it.nextSystemV(ty),
@@ -11234,7 +11234,7 @@ const ParamTypeIterator = struct {
                 it.zig_index += 1;
                 it.llvm_index += 1;
 
-                if (isScalar(mod, ty)) {
+                if (isScalar(zcu, ty)) {
                     return .byval;
                 } else {
                     it.byval_attr = true;
@@ -11250,10 +11250,10 @@ const ParamTypeIterator = struct {
     }
 
     fn nextWin64(it: *ParamTypeIterator, ty: Type) ?Lowering {
-        const mod = it.object.module;
-        switch (x86_64_abi.classifyWindows(ty, mod)) {
+        const zcu = it.object.module;
+        switch (x86_64_abi.classifyWindows(ty, zcu)) {
             .integer => {
-                if (isScalar(mod, ty)) {
+                if (isScalar(zcu, ty)) {
                     it.zig_index += 1;
                     it.llvm_index += 1;
                     return .byval;
@@ -11283,16 +11283,17 @@ const ParamTypeIterator = struct {
     }
 
     fn nextSystemV(it: *ParamTypeIterator, ty: Type) Allocator.Error!?Lowering {
-        const mod = it.object.module;
-        const ip = &mod.intern_pool;
-        const classes = x86_64_abi.classifySystemV(ty, mod, .arg);
+        const zcu = it.object.module;
+        const ip = &zcu.intern_pool;
+        const target = zcu.getTarget();
+        const classes = x86_64_abi.classifySystemV(ty, zcu, target, .arg);
         if (classes[0] == .memory) {
             it.zig_index += 1;
             it.llvm_index += 1;
             it.byval_attr = true;
             return .byref;
         }
-        if (isScalar(mod, ty)) {
+        if (isScalar(zcu, ty)) {
             it.zig_index += 1;
             it.llvm_index += 1;
             return .byval;
diff --git a/test/c_abi/cfuncs.c b/test/c_abi/cfuncs.c
index c6477a3d9d..e948c067ac 100644
--- a/test/c_abi/cfuncs.c
+++ b/test/c_abi/cfuncs.c
@@ -269,6 +269,33 @@ void c_struct_f32_f32f32(struct Struct_f32_f32f32 s) {
     assert_or_panic(s.b.d == 3.0f);
 }
 
+struct Struct_u32_Union_u32_u32u32 {
+    uint32_t a;
+    union {
+        struct {
+            uint32_t d, e;
+        } c;
+    } b;
+};
+
+struct Struct_u32_Union_u32_u32u32 zig_ret_struct_u32_union_u32_u32u32(void);
+
+void zig_struct_u32_union_u32_u32u32(struct Struct_u32_Union_u32_u32u32);
+
+struct Struct_u32_Union_u32_u32u32 c_ret_struct_u32_union_u32_u32u32(void) {
+    struct Struct_u32_Union_u32_u32u32 s;
+    s.a = 1;
+    s.b.c.d = 2;
+    s.b.c.e = 3;
+    return s;
+}
+
+void c_struct_u32_union_u32_u32u32(struct Struct_u32_Union_u32_u32u32 s) {
+    assert_or_panic(s.a == 1);
+    assert_or_panic(s.b.c.d == 2);
+    assert_or_panic(s.b.c.e == 3);
+}
+
 struct BigStruct {
     uint64_t a;
     uint64_t b;
@@ -2664,6 +2691,16 @@ void run_c_tests(void) {
     }
 #endif
 
+#if !defined(__powerpc__)
+    {
+        struct Struct_u32_Union_u32_u32u32 s = zig_ret_struct_u32_union_u32_u32u32();
+        assert_or_panic(s.a == 1);
+        assert_or_panic(s.b.c.d == 2);
+        assert_or_panic(s.b.c.e == 3);
+        zig_struct_u32_union_u32_u32u32(s);
+    }
+#endif
+
     {
         struct BigStruct s = {1, 2, 3, 4, 5};
         zig_big_struct(s);
@@ -2678,7 +2715,7 @@ void run_c_tests(void) {
     }
 #endif
 
-#if !defined __i386__ && !defined __arm__ && !defined __aarch64__ && \
+#if !defined __arm__ && !defined __aarch64__ && \
     !defined __mips__ && !defined __powerpc__ && !defined ZIG_RISCV64
     {
         struct MedStructInts s = {1, 2, 3};
diff --git a/test/c_abi/main.zig b/test/c_abi/main.zig
index fb5879338b..028156f35d 100644
--- a/test/c_abi/main.zig
+++ b/test/c_abi/main.zig
@@ -10,11 +10,11 @@ const builtin = @import("builtin");
 const print = std.debug.print;
 const expect = std.testing.expect;
 const expectEqual = std.testing.expectEqual;
-const has_i128 = builtin.cpu.arch != .x86 and !builtin.cpu.arch.isARM() and
+const have_i128 = builtin.cpu.arch != .x86 and !builtin.cpu.arch.isARM() and
     !builtin.cpu.arch.isMIPS() and !builtin.cpu.arch.isPPC();
 
-const has_f128 = builtin.cpu.arch.isX86() and !builtin.os.tag.isDarwin();
-const has_f80 = builtin.cpu.arch.isX86();
+const have_f128 = builtin.cpu.arch.isX86() and !builtin.os.tag.isDarwin();
+const have_f80 = builtin.cpu.arch.isX86();
 
 extern fn run_c_tests() void;
 
@@ -53,13 +53,13 @@ test "C ABI integers" {
     c_u16(0xfffe);
     c_u32(0xfffffffd);
     c_u64(0xfffffffffffffffc);
-    if (has_i128) c_struct_u128(.{ .value = 0xfffffffffffffffc });
+    if (have_i128) c_struct_u128(.{ .value = 0xfffffffffffffffc });
 
     c_i8(-1);
     c_i16(-2);
     c_i32(-3);
     c_i64(-4);
-    if (has_i128) c_struct_i128(.{ .value = -6 });
+    if (have_i128) c_struct_i128(.{ .value = -6 });
     c_five_integers(12, 34, 56, 78, 90);
 }
 
@@ -186,7 +186,6 @@ const complex_abi_compatible = builtin.cpu.arch != .x86 and !builtin.cpu.arch.is
 
 test "C ABI complex float" {
     if (!complex_abi_compatible) return error.SkipZigTest;
-    if (builtin.zig_backend == .stage2_llvm and builtin.cpu.arch == .x86_64) return error.SkipZigTest; // See https://github.com/ziglang/zig/issues/8465
 
     const a = ComplexFloat{ .real = 1.25, .imag = 2.6 };
     const b = ComplexFloat{ .real = 11.3, .imag = -1.5 };
@@ -401,6 +400,42 @@ test "C ABI struct f32 {f32,f32}" {
     c_struct_f32_f32f32(.{ .a = 1.0, .b = .{ .c = 2.0, .d = 3.0 } });
 }
 
+const Struct_u32_Union_u32_u32u32 = extern struct {
+    a: u32,
+    b: extern union {
+        c: extern struct {
+            d: u32,
+            e: u32,
+        },
+    },
+};
+
+export fn zig_ret_struct_u32_union_u32_u32u32() Struct_u32_Union_u32_u32u32 {
+    return .{ .a = 1, .b = .{ .c = .{ .d = 2, .e = 3 } } };
+}
+
+export fn zig_struct_u32_union_u32_u32u32(s: Struct_u32_Union_u32_u32u32) void {
+    expect(s.a == 1) catch @panic("test failure");
+    expect(s.b.c.d == 2) catch @panic("test failure");
+    expect(s.b.c.e == 3) catch @panic("test failure");
+}
+
+extern fn c_ret_struct_u32_union_u32_u32u32() Struct_u32_Union_u32_u32u32;
+
+extern fn c_struct_u32_union_u32_u32u32(Struct_u32_Union_u32_u32u32) void;
+
+test "C ABI struct{u32,union{u32,struct{u32,u32}}}" {
+    if (builtin.cpu.arch.isMIPS()) return error.SkipZigTest;
+    if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
+    if (builtin.cpu.arch.isPPC64()) return error.SkipZigTest;
+
+    const s = c_ret_struct_u32_union_u32_u32u32();
+    try expect(s.a == 1);
+    try expect(s.b.c.d == 2);
+    try expect(s.b.c.e == 3);
+    c_struct_u32_union_u32_u32u32(.{ .a = 1, .b = .{ .c = .{ .d = 2, .e = 3 } } });
+}
+
 const BigStruct = extern struct {
     a: u64,
     b: u64,
@@ -470,7 +505,6 @@ extern fn c_med_struct_mixed(MedStructMixed) void;
 extern fn c_ret_med_struct_mixed() MedStructMixed;
 
 test "C ABI medium struct of ints and floats" {
-    if (builtin.cpu.arch == .x86) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC64()) return error.SkipZigTest;
@@ -538,7 +572,6 @@ extern fn c_med_struct_ints(MedStructInts) void;
 extern fn c_ret_med_struct_ints() MedStructInts;
 
 test "C ABI medium struct of ints" {
-    if (builtin.cpu.arch == .x86) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC64()) return error.SkipZigTest;
@@ -600,7 +633,7 @@ export fn zig_big_packed_struct(x: BigPackedStruct) void {
 }
 
 test "C ABI big packed struct" {
-    if (!has_i128) return error.SkipZigTest;
+    if (!have_i128) return error.SkipZigTest;
 
     const s = BigPackedStruct{ .a = 1, .b = 2 };
     c_big_packed_struct(s);
@@ -943,7 +976,6 @@ extern fn c_float_array_struct(FloatArrayStruct) void;
 extern fn c_ret_float_array_struct() FloatArrayStruct;
 
 test "Float array like struct" {
-    if (builtin.cpu.arch == .x86 and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
 
@@ -5318,7 +5350,6 @@ extern fn c_ptr_size_float_struct(Vector2) void;
 extern fn c_ret_ptr_size_float_struct() Vector2;
 
 test "C ABI pointer sized float struct" {
-    if (builtin.cpu.arch == .x86) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS()) return error.SkipZigTest;
     if (builtin.cpu.arch.isRISCV()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
@@ -5348,7 +5379,6 @@ test "DC: Zig passes to C" {
     try expectOk(c_assert_DC(.{ .v1 = -0.25, .v2 = 15 }));
 }
 test "DC: Zig returns to C" {
-    if (builtin.cpu.arch == .x86 and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS() and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch.isRISCV()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
@@ -5363,7 +5393,6 @@ test "DC: C passes to Zig" {
     try expectOk(c_send_DC());
 }
 test "DC: C returns to Zig" {
-    if (builtin.cpu.arch == .x86 and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS() and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch.isRISCV()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
@@ -5397,7 +5426,6 @@ test "CFF: Zig passes to C" {
     try expectOk(c_assert_CFF(.{ .v1 = 39, .v2 = 0.875, .v3 = 1.0 }));
 }
 test "CFF: Zig returns to C" {
-    if (builtin.cpu.arch == .x86 and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC64()) return error.SkipZigTest;
@@ -5414,7 +5442,6 @@ test "CFF: C passes to Zig" {
     try expectOk(c_send_CFF());
 }
 test "CFF: C returns to Zig" {
-    if (builtin.cpu.arch == .x86 and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch == .aarch64 and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch.isRISCV() and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS()) return error.SkipZigTest;
@@ -5442,28 +5469,24 @@ pub export fn zig_ret_CFF() CFF {
 const PD = extern struct { v1: ?*anyopaque, v2: f64 };
 
 test "PD: Zig passes to C" {
-    if (builtin.target.cpu.arch == .x86) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC64()) return error.SkipZigTest;
     try expectOk(c_assert_PD(.{ .v1 = null, .v2 = 0.5 }));
 }
 test "PD: Zig returns to C" {
-    if (builtin.target.cpu.arch == .x86) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS() and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC64()) return error.SkipZigTest;
     try expectOk(c_assert_ret_PD());
 }
 test "PD: C passes to Zig" {
-    if (builtin.target.cpu.arch == .x86) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC64()) return error.SkipZigTest;
     try expectOk(c_send_PD());
 }
 test "PD: C returns to Zig" {
-    if (builtin.target.cpu.arch == .x86) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS() and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC64()) return error.SkipZigTest;
@@ -5519,7 +5542,6 @@ const ByVal = extern struct {
 
 extern fn c_func_ptr_byval(*anyopaque, *anyopaque, ByVal, c_ulong, *anyopaque, c_ulong) void;
 test "C function that takes byval struct called via function pointer" {
-    if (builtin.cpu.arch == .x86 and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch.isMIPS() and builtin.mode != .Debug) return error.SkipZigTest;
     if (builtin.cpu.arch.isPPC()) return error.SkipZigTest;
 
@@ -5551,7 +5573,6 @@ const f16_struct = extern struct {
 };
 extern fn c_f16_struct(f16_struct) f16_struct;
 test "f16 struct" {
-    if (builtin.target.cpu.arch == .x86) return error.SkipZigTest;
     if (builtin.target.cpu.arch.isMIPS()) return error.SkipZigTest;
     if (builtin.target.cpu.arch.isPPC()) return error.SkipZigTest;
     if (builtin.target.cpu.arch.isPPC()) return error.SkipZigTest;
@@ -5563,7 +5584,7 @@ test "f16 struct" {
 
 extern fn c_f80(f80) f80;
 test "f80 bare" {
-    if (!has_f80) return error.SkipZigTest;
+    if (!have_f80) return error.SkipZigTest;
 
     const a = c_f80(12.34);
     try expect(@as(f64, @floatCast(a)) == 56.78);
@@ -5574,9 +5595,7 @@ const f80_struct = extern struct {
 };
 extern fn c_f80_struct(f80_struct) f80_struct;
 test "f80 struct" {
-    if (!has_f80) return error.SkipZigTest;
-    if (builtin.target.cpu.arch == .x86) return error.SkipZigTest;
-    if (builtin.zig_backend == .stage2_llvm and builtin.mode != .Debug) return error.SkipZigTest;
+    if (!have_f80) return error.SkipZigTest;
 
     const a = c_f80_struct(.{ .a = 12.34 });
     try expect(@as(f64, @floatCast(a.a)) == 56.78);
@@ -5588,8 +5607,7 @@ const f80_extra_struct = extern struct {
 };
 extern fn c_f80_extra_struct(f80_extra_struct) f80_extra_struct;
 test "f80 extra struct" {
-    if (!has_f80) return error.SkipZigTest;
-    if (builtin.target.cpu.arch == .x86) return error.SkipZigTest;
+    if (!have_f80) return error.SkipZigTest;
 
     const a = c_f80_extra_struct(.{ .a = 12.34, .b = 42 });
     try expect(@as(f64, @floatCast(a.a)) == 56.78);
@@ -5598,7 +5616,7 @@ test "f80 extra struct" {
 
 extern fn c_f128(f128) f128;
 test "f128 bare" {
-    if (!has_f128) return error.SkipZigTest;
+    if (!have_f128) return error.SkipZigTest;
 
     const a = c_f128(12.34);
     try expect(@as(f64, @floatCast(a)) == 56.78);
@@ -5609,7 +5627,7 @@ const f128_struct = extern struct {
 };
 extern fn c_f128_struct(f128_struct) f128_struct;
 test "f128 struct" {
-    if (!has_f128) return error.SkipZigTest;
+    if (!have_f128) return error.SkipZigTest;
 
     const a = c_f128_struct(.{ .a = 12.34 });
     try expect(@as(f64, @floatCast(a.a)) == 56.78);