llvm: fix aarch64 c abi HFA detection

Aggregate types do not count as Homogeneous Aggregates if they have
padding gaps between fields or at the end due to field alignments.

src/codegen/aarch64/abi.zig

@@ -1,5 +1,6 @@
 const assert = @import("std").debug.assert;
 const std = @import("std");
+const InternPool = @import("../../InternPool.zig");
 const Type = @import("../../Type.zig");
 const Zcu = @import("../../Zcu.zig");
 
@@ -15,12 +16,10 @@ pub const Class = union(enum) {
 pub fn classifyType(ty: Type, zcu: *Zcu) Class {
     assert(ty.hasRuntimeBits(zcu));
 
-    var maybe_float_bits: ?u16 = null;
     switch (ty.zigTypeTag(zcu)) {
         .@"struct" => {
             if (ty.containerLayout(zcu) == .@"packed") return .byval;
-            const float_count = countFloats(ty, zcu, &maybe_float_bits);
-            if (float_count <= sret_float_count) return .{ .float_array = float_count };
+            if (countFloats(ty, zcu)) |float| return .{ .float_array = float.count };
 
             const bit_size = ty.bitSize(zcu);
             if (bit_size > 128) return .memory;
@@ -29,8 +28,7 @@ pub fn classifyType(ty: Type, zcu: *Zcu) Class {
         },
         .@"union" => {
             if (ty.containerLayout(zcu) == .@"packed") return .byval;
-            const float_count = countFloats(ty, zcu, &maybe_float_bits);
-            if (float_count <= sret_float_count) return .{ .float_array = float_count };
+            if (countFloats(ty, zcu)) |float| return .{ .float_array = float.count };
 
             const bit_size = ty.bitSize(zcu);
             if (bit_size > 128) return .memory;
@@ -70,46 +68,52 @@ pub fn classifyType(ty: Type, zcu: *Zcu) Class {
     }
 }
 
-const sret_float_count = 4;
-fn countFloats(ty: Type, zcu: *Zcu, maybe_float_bits: *?u16) u8 {
+const CountFloatsResult = struct {
+    ty: Type,
+    count: std.math.IntFittingRange(0, max_count),
+
+    const none: CountFloatsResult = .{ .ty = .void, .count = 0 };
+
+    const max_count = 4;
+};
+fn countFloats(ty: Type, zcu: *Zcu) ?CountFloatsResult {
     const ip = &zcu.intern_pool;
-    const target = zcu.getTarget();
-    const invalid = std.math.maxInt(u8);
+    if (!ty.hasRuntimeBits(zcu)) return .none;
     switch (ty.zigTypeTag(zcu)) {
         .@"union" => {
-            const union_obj = zcu.typeToUnion(ty).?;
-            var max_count: u8 = 0;
-            for (union_obj.field_types.get(ip)) |field_ty| {
-                const field_count = countFloats(Type.fromInterned(field_ty), zcu, maybe_float_bits);
-                if (field_count == invalid) return invalid;
-                if (field_count > max_count) max_count = field_count;
-                if (max_count > sret_float_count) return invalid;
+            const loaded_union = zcu.typeToUnion(ty).?;
+            var result: CountFloatsResult = .none;
+            for (loaded_union.field_types.get(ip)) |field_ty| {
+                const float = countFloats(Type.fromInterned(field_ty), zcu) orelse return null;
+                if (result.ty.toIntern() == .void_type) {
+                    result.ty = float.ty;
+                } else if (result.ty.bitSize(zcu) != float.ty.bitSize(zcu)) return null;
+                result.count = @max(result.count, float.count);
             }
-            return max_count;
+            if (ty.abiSize(zcu) != result.ty.abiSize(zcu) * result.count) return null;
+            return result;
         },
         .@"struct" => {
-            const fields_len = ty.structFieldCount(zcu);
-            var count: u8 = 0;
-            var i: u32 = 0;
-            while (i < fields_len) : (i += 1) {
-                const field_ty = ty.fieldType(i, zcu);
-                const field_count = countFloats(field_ty, zcu, maybe_float_bits);
-                if (field_count == invalid) return invalid;
-                count += field_count;
-                if (count > sret_float_count) return invalid;
+            var result: CountFloatsResult = .none;
+            var field_it: InternPool.LoadedStructType.RuntimeOrderIterator = if (zcu.typeToStruct(ty)) |loaded_struct|
+                loaded_struct.iterateRuntimeOrder(ip)
+            else
+                .{ .runtime_order = null, .fields_len = ty.structFieldCount(zcu), .next_index = 0 };
+            while (field_it.next()) |field_index| {
+                if (ty.structFieldOffset(field_index, zcu) != result.ty.abiSize(zcu) * result.count) return null;
+                const field_ty = ty.fieldType(field_index, zcu);
+                const float = countFloats(field_ty, zcu) orelse return null;
+                if (result.ty.toIntern() == .void_type) {
+                    result.ty = float.ty;
+                } else if (result.ty.bitSize(zcu) != float.ty.bitSize(zcu)) return null;
+                if (float.count > CountFloatsResult.max_count - result.count) return null;
+                result.count += float.count;
             }
-            return count;
+            if (ty.abiSize(zcu) != result.ty.abiSize(zcu) * result.count) return null;
+            return result;
         },
-        .float => {
-            const float_bits = maybe_float_bits.* orelse {
-                maybe_float_bits.* = ty.floatBits(target);
-                return 1;
-            };
-            if (ty.floatBits(target) == float_bits) return 1;
-            return invalid;
-        },
-        .void => return 0,
-        else => return invalid,
+        .float => return .{ .ty = ty, .count = 1 },
+        else => return null,
     }
 }
 
@@ -117,17 +121,19 @@ pub fn getFloatArrayType(ty: Type, zcu: *Zcu) ?Type {
     const ip = &zcu.intern_pool;
     switch (ty.zigTypeTag(zcu)) {
         .@"union" => {
-            const union_obj = zcu.typeToUnion(ty).?;
-            for (union_obj.field_types.get(ip)) |field_ty| {
+            const loaded_union = zcu.typeToUnion(ty).?;
+            for (loaded_union.field_types.get(ip)) |field_ty| {
                 if (getFloatArrayType(Type.fromInterned(field_ty), zcu)) |some| return some;
             }
             return null;
         },
         .@"struct" => {
-            const fields_len = ty.structFieldCount(zcu);
-            var i: u32 = 0;
-            while (i < fields_len) : (i += 1) {
-                const field_ty = ty.fieldType(i, zcu);
+            var field_it: InternPool.LoadedStructType.RuntimeOrderIterator = if (zcu.typeToStruct(ty)) |loaded_struct|
+                loaded_struct.iterateRuntimeOrder(ip)
+            else
+                .{ .runtime_order = null, .fields_len = ty.structFieldCount(zcu), .next_index = 0 };
+            while (field_it.next()) |field_index| {
+                const field_ty = ty.fieldType(field_index, zcu);
                 if (getFloatArrayType(field_ty, zcu)) |some| return some;
             }
             return null;

src/codegen/llvm/FuncGen.zig

@@ -6707,17 +6707,18 @@ const ParamTypeIterator = struct {
                     .double_integer => return Lowering{ .i64_array = 2 },
                     .fields => {
                         it.types_len = 0;
-                        var offset: u64 = 0;
-                        for (0..ty.structFieldCount(zcu)) |field_index| {
+                        var field_it: InternPool.LoadedStructType.RuntimeOrderIterator = if (zcu.typeToStruct(ty)) |loaded_struct|
+                            loaded_struct.iterateRuntimeOrder(&zcu.intern_pool)
+                        else
+                            .{ .runtime_order = null, .fields_len = ty.structFieldCount(zcu), .next_index = 0 };
+                        while (field_it.next()) |field_index| {
                             const field_ty = ty.fieldType(field_index, zcu);
                             if (!field_ty.hasRuntimeBits(zcu)) continue;
-                            offset = field_ty.abiAlignment(zcu).forward(offset);
                             it.types_buffer[it.types_len] = try it.object.lowerType(field_ty);
-                            it.offsets_buffer[it.types_len] = offset;
+                            it.offsets_buffer[it.types_len] = ty.structFieldOffset(field_index, zcu);
                             it.types_len += 1;
-                            offset += field_ty.abiSize(zcu);
                         }
-                        it.offsets_buffer[it.types_len] = offset;
+                        it.offsets_buffer[it.types_len] = ty.abiSize(zcu);
                         it.llvm_index += it.types_len - 1;
                         return .multiple_llvm_types;
                     },

test/c_abi/cfuncs.c

@@ -1,5 +1,6 @@
 #include <complex.h>
 #include <inttypes.h>
+#include <stdalign.h>
 #include <stdbool.h>
 #include <stdlib.h>
 #include <string.h>
@@ -209,7 +210,7 @@ struct Struct_u8 zig_ret_struct_u8(void);
 void zig_struct_u8(struct Struct_u8, size_t);
 
 struct Struct_u8 c_ret_struct_u8(void) {
-    return (struct Struct_u8){ 4 };
+    return (struct Struct_u8){ .a = 4 };
 }
 
 void c_struct_u8(struct Struct_u8 s, size_t i) {
@@ -226,7 +227,7 @@ struct Struct_u16 zig_ret_struct_u16(void);
 void zig_struct_u16(struct Struct_u16, size_t);
 
 struct Struct_u16 c_ret_struct_u16(void) {
-    return (struct Struct_u16){ 10 };
+    return (struct Struct_u16){ .a = 10 };
 }
 
 void c_struct_u16(struct Struct_u16 s, size_t i) {
@@ -243,7 +244,7 @@ struct Struct_u32 zig_ret_struct_u32(void);
 void zig_struct_u32(struct Struct_u32, size_t);
 
 struct Struct_u32 c_ret_struct_u32(void) {
-    return (struct Struct_u32){ 16 };
+    return (struct Struct_u32){ .a = 16 };
 }
 
 void c_struct_u32(struct Struct_u32 s, size_t i) {
@@ -260,7 +261,7 @@ struct Struct_u64 zig_ret_struct_u64(void);
 void zig_struct_u64(struct Struct_u64, size_t);
 
 struct Struct_u64 c_ret_struct_u64(void) {
-    return (struct Struct_u64){ 22 };
+    return (struct Struct_u64){ .a = 22 };
 }
 
 void c_struct_u64(struct Struct_u64 s, size_t i) {
@@ -286,7 +287,7 @@ void zig_struct_u64_u64_7(size_t, size_t, size_t, size_t, size_t, size_t, size_t
 void zig_struct_u64_u64_8(size_t, size_t, size_t, size_t, size_t, size_t, size_t, size_t, struct Struct_u64_u64, size_t);
 
 struct Struct_u64_u64 c_ret_struct_u64_u64(void) {
-    return (struct Struct_u64_u64){ 21, 22 };
+    return (struct Struct_u64_u64){ .a = 21, .b = 22 };
 }
 
 void c_struct_u64_u64_0(struct Struct_u64_u64 s, size_t i) {
@@ -344,7 +345,7 @@ struct Struct_f32 zig_ret_struct_f32(void);
 void zig_struct_f32(struct Struct_f32);
 
 struct Struct_f32 c_ret_struct_f32(void) {
-    return (struct Struct_f32){ 2.5f };
+    return (struct Struct_f32){ .a = 2.5f };
 }
 
 void c_struct_f32(struct Struct_f32 s) {
@@ -360,13 +361,49 @@ struct Struct_f64 zig_ret_struct_f64(void);
 void zig_struct_f64(struct Struct_f64);
 
 struct Struct_f64 c_ret_struct_f64(void) {
-    return (struct Struct_f64){ 2.5 };
+    return (struct Struct_f64){ .a = 2.5 };
 }
 
 void c_struct_f64(struct Struct_f64 s) {
     assert_or_panic(s.a == 2.5);
 }
 
+struct Struct_f32a8 {
+    alignas(8) float a;
+};
+
+struct Struct_f32a8 zig_ret_struct_f32a8(void);
+
+void zig_struct_f32a8(struct Struct_f32a8, float);
+
+struct Struct_f32a8 c_ret_struct_f32a8(void) {
+    return (struct Struct_f32a8){ .a = 4.125f };
+}
+
+void c_struct_f32a8(struct Struct_f32a8 s, float f) {
+    assert_or_panic(s.a == 5.375f);
+    assert_or_panic(f == 6.5f);
+}
+
+struct Struct_f32a8_f32a8 {
+    alignas(8) float a;
+    alignas(8) float b;
+};
+
+struct Struct_f32a8_f32a8 zig_ret_struct_f32a8_f32a8(void);
+
+void zig_struct_f32a8_f32a8(struct Struct_f32a8_f32a8, float);
+
+struct Struct_f32a8_f32a8 c_ret_struct_f32a8_f32a8(void) {
+    return (struct Struct_f32a8_f32a8){ .a = 6.625f, .b = 7.875f };
+}
+
+void c_struct_f32a8_f32a8(struct Struct_f32a8_f32a8 s, float f) {
+    assert_or_panic(s.a == 8.0625f);
+    assert_or_panic(s.b == 9.1875f);
+    assert_or_panic(f == 10.5f);
+}
+
 struct Struct_f32f32_f32 {
     struct {
         float b, c;
@@ -379,7 +416,7 @@ struct Struct_f32f32_f32 zig_ret_struct_f32f32_f32(void);
 void zig_struct_f32f32_f32(struct Struct_f32f32_f32);
 
 struct Struct_f32f32_f32 c_ret_struct_f32f32_f32(void) {
-    return (struct Struct_f32f32_f32){ { 1.0f, 2.0f }, 3.0f };
+    return (struct Struct_f32f32_f32){ .a = { .b = 1.0f, .c = 2.0f }, .d = 3.0f };
 }
 
 void c_struct_f32f32_f32(struct Struct_f32f32_f32 s) {
@@ -400,7 +437,7 @@ struct Struct_f32_f32f32 zig_ret_struct_f32_f32f32(void);
 void zig_struct_f32_f32f32(struct Struct_f32_f32f32);
 
 struct Struct_f32_f32f32 c_ret_struct_f32_f32f32(void) {
-    return (struct Struct_f32_f32f32){ 1.0f, { 2.0f, 3.0f } };
+    return (struct Struct_f32_f32f32){ .a = 1.0f, .b = { .c = 2.0f, .d = 3.0f } };
 }
 
 void c_struct_f32_f32f32(struct Struct_f32_f32f32 s) {
@@ -2870,7 +2907,7 @@ void run_c_tests(void) {
     {
         struct Struct_f32 s = zig_ret_struct_f32();
         assert_or_panic(s.a == 2.5f);
-        zig_struct_f32((struct Struct_f32){ 2.5f });
+        zig_struct_f32((struct Struct_f32){ .a = 2.5f });
     }
 #endif
 
@@ -2879,17 +2916,62 @@ void run_c_tests(void) {
     {
         struct Struct_f64 s = zig_ret_struct_f64();
         assert_or_panic(s.a == 2.5);
-        zig_struct_f64((struct Struct_f64){ 2.5 });
+        zig_struct_f64((struct Struct_f64){ .a = 2.5 });
     }
 #endif
+#endif
 
+#if !defined(__arm__)
+#if !defined(__loongarch__)
+#if !defined(__mips64__)
+#if !defined(__powerpc__)
+#if !defined(ZIG_RISCV32)
+#if !defined(__s390x__)
+#if !defined(__i386__)
+    {
+        struct Struct_f32a8 s = zig_ret_struct_f32a8();
+        assert_or_panic(s.a == 1.25f);
+        zig_struct_f32a8((struct Struct_f32a8){ .a = 2.75f }, 3.5f);
+    }
+#endif
+#endif
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#if !defined(__arm__)
+#if !defined(__loongarch__)
+#if !defined(__mips64__)
+#if !defined(__powerpc__)
+#if !defined(__riscv)
+#if !defined(__s390x__)
+#if !defined(__i386__)
+#if !defined(__x86_64__)
+    {
+        struct Struct_f32a8_f32a8 s = zig_ret_struct_f32a8_f32a8();
+        assert_or_panic(s.a == 1.25f);
+        assert_or_panic(s.b == 2.75f);
+        zig_struct_f32a8_f32a8((struct Struct_f32a8_f32a8){ .a = 3.125f, .b = 4.375f }, 5.5f);
+    }
+#endif
+#endif
+#endif
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#if !(defined(__arm__) && defined(__SOFTFP__))
 #if !defined(__loongarch__) && !defined(__mips64__)
     {
         struct Struct_f32f32_f32 s = zig_ret_struct_f32f32_f32();
         assert_or_panic(s.a.b == 1.0f);
         assert_or_panic(s.a.c == 2.0f);
         assert_or_panic(s.d == 3.0f);
-        zig_struct_f32f32_f32((struct Struct_f32f32_f32){ { 1.0f, 2.0f }, 3.0f });
+        zig_struct_f32f32_f32((struct Struct_f32f32_f32){ .a = { .b = 1.0f, .c = 2.0f }, .d = 3.0f });
     }
 
     {
@@ -2897,7 +2979,7 @@ void run_c_tests(void) {
         assert_or_panic(s.a == 1.0f);
         assert_or_panic(s.b.c == 2.0f);
         assert_or_panic(s.b.d == 3.0f);
-        zig_struct_f32_f32f32((struct Struct_f32_f32f32){ 1.0f, { 2.0f, 3.0f } });
+        zig_struct_f32_f32f32((struct Struct_f32_f32f32){ .a = 1.0f, .b = { .c = 2.0f, .d = 3.0f } });
     }
 #endif
 #endif

test/c_abi/main.zig

@@ -444,7 +444,7 @@ extern fn c_struct_u64_u64_6(usize, usize, usize, usize, usize, usize, Struct_u6
 extern fn c_struct_u64_u64_7(usize, usize, usize, usize, usize, usize, usize, Struct_u64_u64, usize) void;
 extern fn c_struct_u64_u64_8(usize, usize, usize, usize, usize, usize, usize, usize, Struct_u64_u64, usize) void;
 
-test "C ABI struct u64 u64" {
+test "C ABI struct u64, u64" {
     if (builtin.cpu.arch.isMIPS64()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPowerPC32()) return error.SkipZigTest;
     if (builtin.cpu.arch == .hexagon) return error.SkipZigTest;
@@ -518,6 +518,80 @@ test "C ABI struct f64" {
     c_struct_f64(.{ .a = 2.5 });
 }
 
+const Struct_f32a8 = extern struct {
+    a: f32 align(8),
+};
+
+export fn zig_ret_struct_f32a8() Struct_f32a8 {
+    return .{ .a = 1.25 };
+}
+
+export fn zig_struct_f32a8(s: Struct_f32a8, f: f32) void {
+    expect(s.a == 2.75) catch @panic("test failure");
+    expect(f == 3.5) catch @panic("test failure");
+}
+
+extern fn c_ret_struct_f32a8() Struct_f32a8;
+
+extern fn c_struct_f32a8(Struct_f32a8, f32) void;
+
+test "C ABI struct f32 align(8)" {
+    if (builtin.cpu.arch.isArm()) return error.SkipZigTest;
+    if (builtin.cpu.arch.isLoongArch()) return error.SkipZigTest;
+    if (builtin.cpu.arch.isMIPS64()) return error.SkipZigTest;
+    if (builtin.cpu.arch.isPowerPC()) return error.SkipZigTest;
+    if (builtin.cpu.arch == .riscv32) return error.SkipZigTest;
+    if (builtin.cpu.arch == .s390x) return error.SkipZigTest;
+    if (builtin.cpu.arch == .x86) return error.SkipZigTest;
+
+    const s = c_ret_struct_f32a8();
+    try expect(s.a == 4.125);
+    c_struct_f32a8(.{ .a = 5.375 }, 6.5);
+}
+
+const Struct_f32a8_f32a8 = extern struct {
+    a: f32 align(8),
+    b: f32 align(8),
+};
+
+comptime {
+    skip: {
+        if (builtin.zig_backend == .stage2_x86_64) break :skip;
+
+        _ = struct {
+            export fn zig_ret_struct_f32a8_f32a8() Struct_f32a8_f32a8 {
+                return .{ .a = 1.25, .b = 2.75 };
+            }
+
+            export fn zig_struct_f32a8_f32a8(s: Struct_f32a8_f32a8, f: f32) void {
+                expect(s.a == 3.125) catch @panic("test failure");
+                expect(s.b == 4.375) catch @panic("test failure");
+                expect(f == 5.5) catch @panic("test failure");
+            }
+        };
+    }
+}
+
+extern fn c_ret_struct_f32a8_f32a8() Struct_f32a8_f32a8;
+
+extern fn c_struct_f32a8_f32a8(Struct_f32a8_f32a8, f32) void;
+
+test "C ABI struct f32 align(8), f32 align(8)" {
+    if (builtin.cpu.arch.isArm()) return error.SkipZigTest;
+    if (builtin.cpu.arch.isLoongArch()) return error.SkipZigTest;
+    if (builtin.cpu.arch.isMIPS64()) return error.SkipZigTest;
+    if (builtin.cpu.arch.isPowerPC()) return error.SkipZigTest;
+    if (builtin.cpu.arch.isRISCV()) return error.SkipZigTest;
+    if (builtin.cpu.arch == .s390x) return error.SkipZigTest;
+    if (builtin.cpu.arch == .x86) return error.SkipZigTest;
+    if (builtin.cpu.arch == .x86_64) return error.SkipZigTest;
+
+    const s = c_ret_struct_f32a8_f32a8();
+    try expect(s.a == 6.625);
+    try expect(s.b == 7.875);
+    c_struct_f32a8_f32a8(.{ .a = 8.0625, .b = 9.1875 }, 10.5);
+}
+
 const Struct_f32f32_f32 = extern struct {
     a: extern struct { b: f32, c: f32 },
     d: f32,
@@ -537,7 +611,7 @@ extern fn c_ret_struct_f32f32_f32() Struct_f32f32_f32;
 
 extern fn c_struct_f32f32_f32(Struct_f32f32_f32) void;
 
-test "C ABI struct {f32,f32} f32" {
+test "C ABI struct {f32, f32}, f32" {
     if (builtin.cpu.arch.isMIPS64()) return error.SkipZigTest;
     if (builtin.cpu.arch.isPowerPC32()) return error.SkipZigTest;
     if (builtin.cpu.arch.isArm() and builtin.abi.float() == .soft) return error.SkipZigTest;