diff --git a/lib/std/os/windows.zig b/lib/std/os/windows.zig
index 800a16f660..b949f92feb 100644
--- a/lib/std/os/windows.zig
+++ b/lib/std/os/windows.zig
@@ -2693,7 +2693,10 @@ pub const MEM_RESERVE_PLACEHOLDERS = 0x2;
 pub const MEM_DECOMMIT = 0x4000;
 pub const MEM_RELEASE = 0x8000;
 
-pub const PTHREAD_START_ROUTINE = fn (LPVOID) callconv(.C) DWORD;
+pub const PTHREAD_START_ROUTINE = switch (builtin.zig_backend) {
+    .stage1 => fn (LPVOID) callconv(.C) DWORD,
+    else => *const fn (LPVOID) callconv(.C) DWORD,
+};
 pub const LPTHREAD_START_ROUTINE = PTHREAD_START_ROUTINE;
 
 pub const WIN32_FIND_DATAW = extern struct {
@@ -2866,7 +2869,10 @@ pub const IMAGE_TLS_DIRECTORY = extern struct {
 pub const IMAGE_TLS_DIRECTORY64 = IMAGE_TLS_DIRECTORY;
 pub const IMAGE_TLS_DIRECTORY32 = IMAGE_TLS_DIRECTORY;
 
-pub const PIMAGE_TLS_CALLBACK = ?fn (PVOID, DWORD, PVOID) callconv(.C) void;
+pub const PIMAGE_TLS_CALLBACK = switch (builtin.zig_backend) {
+    .stage1 => ?fn (PVOID, DWORD, PVOID) callconv(.C) void,
+    else => ?*const fn (PVOID, DWORD, PVOID) callconv(.C) void,
+};
 
 pub const PROV_RSA_FULL = 1;
 
@@ -2892,7 +2898,10 @@ pub const FILE_ACTION_MODIFIED = 0x00000003;
 pub const FILE_ACTION_RENAMED_OLD_NAME = 0x00000004;
 pub const FILE_ACTION_RENAMED_NEW_NAME = 0x00000005;
 
-pub const LPOVERLAPPED_COMPLETION_ROUTINE = ?fn (DWORD, DWORD, *OVERLAPPED) callconv(.C) void;
+pub const LPOVERLAPPED_COMPLETION_ROUTINE = switch (builtin.zig_backend) {
+    .stage1 => ?fn (DWORD, DWORD, *OVERLAPPED) callconv(.C) void,
+    else => ?*const fn (DWORD, DWORD, *OVERLAPPED) callconv(.C) void,
+};
 
 pub const FILE_NOTIFY_CHANGE_CREATION = 64;
 pub const FILE_NOTIFY_CHANGE_SIZE = 8;
@@ -2945,7 +2954,10 @@ pub const RTL_CRITICAL_SECTION = extern struct {
 pub const CRITICAL_SECTION = RTL_CRITICAL_SECTION;
 pub const INIT_ONCE = RTL_RUN_ONCE;
 pub const INIT_ONCE_STATIC_INIT = RTL_RUN_ONCE_INIT;
-pub const INIT_ONCE_FN = fn (InitOnce: *INIT_ONCE, Parameter: ?*anyopaque, Context: ?*anyopaque) callconv(.C) BOOL;
+pub const INIT_ONCE_FN = switch (builtin.zig_backend) {
+    .stage1 => fn (InitOnce: *INIT_ONCE, Parameter: ?*anyopaque, Context: ?*anyopaque) callconv(.C) BOOL,
+    else => *const fn (InitOnce: *INIT_ONCE, Parameter: ?*anyopaque, Context: ?*anyopaque) callconv(.C) BOOL,
+};
 
 pub const RTL_RUN_ONCE = extern struct {
     Ptr: ?*anyopaque,
@@ -3230,7 +3242,10 @@ pub const EXCEPTION_POINTERS = extern struct {
     ContextRecord: *std.os.windows.CONTEXT,
 };
 
-pub const VECTORED_EXCEPTION_HANDLER = fn (ExceptionInfo: *EXCEPTION_POINTERS) callconv(WINAPI) c_long;
+pub const VECTORED_EXCEPTION_HANDLER = switch (builtin.zig_backend) {
+    .stage1 => fn (ExceptionInfo: *EXCEPTION_POINTERS) callconv(WINAPI) c_long,
+    else => *const fn (ExceptionInfo: *EXCEPTION_POINTERS) callconv(WINAPI) c_long,
+};
 
 pub const OBJECT_ATTRIBUTES = extern struct {
     Length: ULONG,
@@ -3506,7 +3521,10 @@ pub const RTL_DRIVE_LETTER_CURDIR = extern struct {
     DosPath: UNICODE_STRING,
 };
 
-pub const PPS_POST_PROCESS_INIT_ROUTINE = ?fn () callconv(.C) void;
+pub const PPS_POST_PROCESS_INIT_ROUTINE = switch (builtin.zig_backend) {
+    .stage1 => ?fn () callconv(.C) void,
+    else => ?*const fn () callconv(.C) void,
+};
 
 pub const FILE_BOTH_DIR_INFORMATION = extern struct {
     NextEntryOffset: ULONG,
@@ -3526,7 +3544,10 @@ pub const FILE_BOTH_DIR_INFORMATION = extern struct {
 };
 pub const FILE_BOTH_DIRECTORY_INFORMATION = FILE_BOTH_DIR_INFORMATION;
 
-pub const IO_APC_ROUTINE = fn (PVOID, *IO_STATUS_BLOCK, ULONG) callconv(.C) void;
+pub const IO_APC_ROUTINE = switch (builtin.zig_backend) {
+    .stage1 => fn (PVOID, *IO_STATUS_BLOCK, ULONG) callconv(.C) void,
+    else => *const fn (PVOID, *IO_STATUS_BLOCK, ULONG) callconv(.C) void,
+};
 
 pub const CURDIR = extern struct {
     DosPath: UNICODE_STRING,
@@ -3598,8 +3619,14 @@ pub const ENUM_PAGE_FILE_INFORMATION = extern struct {
     PeakUsage: SIZE_T,
 };
 
-pub const PENUM_PAGE_FILE_CALLBACKW = ?fn (?LPVOID, *ENUM_PAGE_FILE_INFORMATION, LPCWSTR) callconv(.C) BOOL;
-pub const PENUM_PAGE_FILE_CALLBACKA = ?fn (?LPVOID, *ENUM_PAGE_FILE_INFORMATION, LPCSTR) callconv(.C) BOOL;
+pub const PENUM_PAGE_FILE_CALLBACKW = switch (builtin.zig_backend) {
+    .stage1 => ?fn (?LPVOID, *ENUM_PAGE_FILE_INFORMATION, LPCWSTR) callconv(.C) BOOL,
+    else => ?*const fn (?LPVOID, *ENUM_PAGE_FILE_INFORMATION, LPCWSTR) callconv(.C) BOOL,
+};
+pub const PENUM_PAGE_FILE_CALLBACKA = switch (builtin.zig_backend) {
+    .stage1 => ?fn (?LPVOID, *ENUM_PAGE_FILE_INFORMATION, LPCSTR) callconv(.C) BOOL,
+    else => ?*const fn (?LPVOID, *ENUM_PAGE_FILE_INFORMATION, LPCSTR) callconv(.C) BOOL,
+};
 
 pub const PSAPI_WS_WATCH_INFORMATION_EX = extern struct {
     BasicInfo: PSAPI_WS_WATCH_INFORMATION,
@@ -3699,4 +3726,7 @@ pub const CTRL_CLOSE_EVENT: DWORD = 2;
 pub const CTRL_LOGOFF_EVENT: DWORD = 5;
 pub const CTRL_SHUTDOWN_EVENT: DWORD = 6;
 
-pub const HANDLER_ROUTINE = fn (dwCtrlType: DWORD) callconv(.C) BOOL;
+pub const HANDLER_ROUTINE = switch (builtin.zig_backend) {
+    .stage1 => fn (dwCtrlType: DWORD) callconv(.C) BOOL,
+    else => *const fn (dwCtrlType: DWORD) callconv(.C) BOOL,
+};
diff --git a/lib/std/pdb.zig b/lib/std/pdb.zig
index 46078d6252..a44296c920 100644
--- a/lib/std/pdb.zig
+++ b/lib/std/pdb.zig
@@ -14,7 +14,7 @@ const ArrayList = std.ArrayList;
 // documentation and/or contributors.
 
 // https://llvm.org/docs/PDB/DbiStream.html#stream-header
-pub const DbiStreamHeader = packed struct {
+pub const DbiStreamHeader = extern struct {
     VersionSignature: i32,
     VersionHeader: u32,
     Age: u32,
@@ -37,7 +37,7 @@ pub const DbiStreamHeader = packed struct {
     Padding: u32,
 };
 
-pub const SectionContribEntry = packed struct {
+pub const SectionContribEntry = extern struct {
     /// COFF Section index, 1-based
     Section: u16,
     Padding1: [2]u8,
@@ -50,7 +50,7 @@ pub const SectionContribEntry = packed struct {
     RelocCrc: u32,
 };
 
-pub const ModInfo = packed struct {
+pub const ModInfo = extern struct {
     Unused1: u32,
     SectionContr: SectionContribEntry,
     Flags: u16,
@@ -68,7 +68,7 @@ pub const ModInfo = packed struct {
     //ObjFileName: char[],
 };
 
-pub const SectionMapHeader = packed struct {
+pub const SectionMapHeader = extern struct {
     /// Number of segment descriptors
     Count: u16,
 
@@ -76,7 +76,7 @@ pub const SectionMapHeader = packed struct {
     LogCount: u16,
 };
 
-pub const SectionMapEntry = packed struct {
+pub const SectionMapEntry = extern struct {
     /// See the SectionMapEntryFlags enum below.
     Flags: u16,
 
@@ -310,7 +310,7 @@ pub const SymbolKind = enum(u16) {
 
 pub const TypeIndex = u32;
 
-pub const ProcSym = packed struct {
+pub const ProcSym = extern struct {
     Parent: u32,
     End: u32,
     Next: u32,
@@ -342,7 +342,7 @@ pub const SectionContrSubstreamVersion = enum(u32) {
     _,
 };
 
-pub const RecordPrefix = packed struct {
+pub const RecordPrefix = extern struct {
     /// Record length, starting from &RecordKind.
     RecordLen: u16,
 
@@ -354,7 +354,7 @@ pub const RecordPrefix = packed struct {
 /// The structure definition follows.
 /// LineBlockFragmentHeader Blocks[]
 /// Each `LineBlockFragmentHeader` as specified below.
-pub const LineFragmentHeader = packed struct {
+pub const LineFragmentHeader = extern struct {
     /// Code offset of line contribution.
     RelocOffset: u32,
 
@@ -376,7 +376,7 @@ pub const LineFlags = packed struct {
 /// header.  The structure definitions follow.
 /// LineNumberEntry   Lines[NumLines];
 /// ColumnNumberEntry Columns[NumLines];
-pub const LineBlockFragmentHeader = packed struct {
+pub const LineBlockFragmentHeader = extern struct {
     /// Offset of FileChecksum entry in File
     /// checksums buffer.  The checksum entry then
     /// contains another offset into the string
@@ -388,7 +388,7 @@ pub const LineBlockFragmentHeader = packed struct {
     BlockSize: u32,
 };
 
-pub const LineNumberEntry = packed struct {
+pub const LineNumberEntry = extern struct {
     /// Offset to start of code bytes for line number
     Offset: u32,
     Flags: u32,
@@ -404,13 +404,13 @@ pub const LineNumberEntry = packed struct {
     };
 };
 
-pub const ColumnNumberEntry = packed struct {
+pub const ColumnNumberEntry = extern struct {
     StartColumn: u16,
     EndColumn: u16,
 };
 
 /// Checksum bytes follow.
-pub const FileChecksumEntryHeader = packed struct {
+pub const FileChecksumEntryHeader = extern struct {
     /// Byte offset of filename in global string table.
     FileNameOffset: u32,
 
@@ -441,7 +441,7 @@ pub const DebugSubsectionKind = enum(u32) {
     CoffSymbolRVA = 0xfd,
 };
 
-pub const DebugSubsectionHeader = packed struct {
+pub const DebugSubsectionHeader = extern struct {
     /// codeview::DebugSubsectionKind enum
     Kind: DebugSubsectionKind,
 
@@ -449,7 +449,7 @@ pub const DebugSubsectionHeader = packed struct {
     Length: u32,
 };
 
-pub const PDBStringTableHeader = packed struct {
+pub const PDBStringTableHeader = extern struct {
     /// PDBStringTableSignature
     Signature: u32,
 
@@ -686,12 +686,12 @@ pub const Pdb = struct {
 
         var symbol_i: usize = 0;
         while (symbol_i != module.symbols.len) {
-            const prefix = @ptrCast(*RecordPrefix, &module.symbols[symbol_i]);
+            const prefix = @ptrCast(*align(1) RecordPrefix, &module.symbols[symbol_i]);
             if (prefix.RecordLen < 2)
                 return null;
             switch (prefix.RecordKind) {
                 .S_LPROC32, .S_GPROC32 => {
-                    const proc_sym = @ptrCast(*ProcSym, &module.symbols[symbol_i + @sizeOf(RecordPrefix)]);
+                    const proc_sym = @ptrCast(*align(1) ProcSym, &module.symbols[symbol_i + @sizeOf(RecordPrefix)]);
                     if (address >= proc_sym.CodeOffset and address < proc_sym.CodeOffset + proc_sym.CodeSize) {
                         return mem.sliceTo(@ptrCast([*:0]u8, proc_sym) + @sizeOf(ProcSym), 0);
                     }
@@ -712,7 +712,7 @@ pub const Pdb = struct {
         var skip_len: usize = undefined;
         const checksum_offset = module.checksum_offset orelse return error.MissingDebugInfo;
         while (sect_offset != subsect_info.len) : (sect_offset += skip_len) {
-            const subsect_hdr = @ptrCast(*DebugSubsectionHeader, &subsect_info[sect_offset]);
+            const subsect_hdr = @ptrCast(*align(1) DebugSubsectionHeader, &subsect_info[sect_offset]);
             skip_len = subsect_hdr.Length;
             sect_offset += @sizeOf(DebugSubsectionHeader);
 
@@ -720,7 +720,7 @@ pub const Pdb = struct {
                 .Lines => {
                     var line_index = sect_offset;
 
-                    const line_hdr = @ptrCast(*LineFragmentHeader, &subsect_info[line_index]);
+                    const line_hdr = @ptrCast(*align(1) LineFragmentHeader, &subsect_info[line_index]);
                     if (line_hdr.RelocSegment == 0)
                         return error.MissingDebugInfo;
                     line_index += @sizeOf(LineFragmentHeader);
@@ -734,7 +734,7 @@ pub const Pdb = struct {
                         const subsection_end_index = sect_offset + subsect_hdr.Length;
 
                         while (line_index < subsection_end_index) {
-                            const block_hdr = @ptrCast(*LineBlockFragmentHeader, &subsect_info[line_index]);
+                            const block_hdr = @ptrCast(*align(1) LineBlockFragmentHeader, &subsect_info[line_index]);
                             line_index += @sizeOf(LineBlockFragmentHeader);
                             const start_line_index = line_index;
 
@@ -746,7 +746,7 @@ pub const Pdb = struct {
                             // This is done with a simple linear search.
                             var line_i: u32 = 0;
                             while (line_i < block_hdr.NumLines) : (line_i += 1) {
-                                const line_num_entry = @ptrCast(*LineNumberEntry, &subsect_info[line_index]);
+                                const line_num_entry = @ptrCast(*align(1) LineNumberEntry, &subsect_info[line_index]);
                                 line_index += @sizeOf(LineNumberEntry);
 
                                 const vaddr_start = frag_vaddr_start + line_num_entry.Offset;
@@ -758,7 +758,7 @@ pub const Pdb = struct {
                             // line_i == 0 would mean that no matching LineNumberEntry was found.
                             if (line_i > 0) {
                                 const subsect_index = checksum_offset + block_hdr.NameIndex;
-                                const chksum_hdr = @ptrCast(*FileChecksumEntryHeader, &module.subsect_info[subsect_index]);
+                                const chksum_hdr = @ptrCast(*align(1) FileChecksumEntryHeader, &module.subsect_info[subsect_index]);
                                 const strtab_offset = @sizeOf(PDBStringTableHeader) + chksum_hdr.FileNameOffset;
                                 try self.string_table.?.seekTo(strtab_offset);
                                 const source_file_name = try self.string_table.?.reader().readUntilDelimiterAlloc(self.allocator, 0, 1024);
@@ -768,12 +768,12 @@ pub const Pdb = struct {
                                 const column = if (has_column) blk: {
                                     const start_col_index = start_line_index + @sizeOf(LineNumberEntry) * block_hdr.NumLines;
                                     const col_index = start_col_index + @sizeOf(ColumnNumberEntry) * line_entry_idx;
-                                    const col_num_entry = @ptrCast(*ColumnNumberEntry, &subsect_info[col_index]);
+                                    const col_num_entry = @ptrCast(*align(1) ColumnNumberEntry, &subsect_info[col_index]);
                                     break :blk col_num_entry.StartColumn;
                                 } else 0;
 
                                 const found_line_index = start_line_index + line_entry_idx * @sizeOf(LineNumberEntry);
-                                const line_num_entry = @ptrCast(*LineNumberEntry, &subsect_info[found_line_index]);
+                                const line_num_entry = @ptrCast(*align(1) LineNumberEntry, &subsect_info[found_line_index]);
                                 const flags = @ptrCast(*LineNumberEntry.Flags, &line_num_entry.Flags);
 
                                 return debug.LineInfo{
@@ -833,7 +833,7 @@ pub const Pdb = struct {
         var sect_offset: usize = 0;
         var skip_len: usize = undefined;
         while (sect_offset != mod.subsect_info.len) : (sect_offset += skip_len) {
-            const subsect_hdr = @ptrCast(*DebugSubsectionHeader, &mod.subsect_info[sect_offset]);
+            const subsect_hdr = @ptrCast(*align(1) DebugSubsectionHeader, &mod.subsect_info[sect_offset]);
             skip_len = subsect_hdr.Length;
             sect_offset += @sizeOf(DebugSubsectionHeader);
 
@@ -969,7 +969,7 @@ fn blockCountFromSize(size: u32, block_size: u32) u32 {
 }
 
 // https://llvm.org/docs/PDB/MsfFile.html#the-superblock
-const SuperBlock = packed struct {
+const SuperBlock = extern struct {
     /// The LLVM docs list a space between C / C++ but empirically this is not the case.
     const file_magic = "Microsoft C/C++ MSF 7.00\r\n\x1a\x44\x53\x00\x00\x00";
 
diff --git a/lib/std/special/test_runner.zig b/lib/std/special/test_runner.zig
index d71f2b9274..4b6c80ac66 100644
--- a/lib/std/special/test_runner.zig
+++ b/lib/std/special/test_runner.zig
@@ -28,7 +28,7 @@ pub fn main() void {
     {
         return main2() catch @panic("test failure");
     }
-    if (builtin.zig_backend == .stage1) processArgs();
+    processArgs();
     const test_fn_list = builtin.test_functions;
     var ok_count: usize = 0;
     var skip_count: usize = 0;
diff --git a/lib/std/start.zig b/lib/std/start.zig
index 5690337317..81c7942bc1 100644
--- a/lib/std/start.zig
+++ b/lib/std/start.zig
@@ -29,9 +29,7 @@ comptime {
         builtin.zig_backend == .stage2_aarch64 or
         builtin.zig_backend == .stage2_arm or
         builtin.zig_backend == .stage2_riscv64 or
-        builtin.zig_backend == .stage2_sparcv9 or
-        (builtin.zig_backend == .stage2_llvm and native_os != .linux and native_os != .macos) or
-        (builtin.zig_backend == .stage2_llvm and native_arch != .x86_64 and native_arch != .aarch64))
+        builtin.zig_backend == .stage2_sparcv9)
     {
         if (builtin.output_mode == .Exe) {
             if ((builtin.link_libc or builtin.object_format == .c) and @hasDecl(root, "main")) {
diff --git a/src/Air.zig b/src/Air.zig
index 0968d95180..2dae8454cf 100644
--- a/src/Air.zig
+++ b/src/Air.zig
@@ -815,6 +815,8 @@ pub const VectorCmp = struct {
 /// 1. `Inst.Ref` for every inputs_len
 /// 2. for every outputs_len
 ///    - constraint: memory at this position is reinterpreted as a null
+///      terminated string.
+///    - name: memory at this position is reinterpreted as a null
 ///      terminated string. pad to the next u32 after the null byte.
 /// 3. for every inputs_len
 ///    - constraint: memory at this position is reinterpreted as a null
diff --git a/src/Sema.zig b/src/Sema.zig
index 87cb7cc9d2..d9c2d1cc2a 100644
--- a/src/Sema.zig
+++ b/src/Sema.zig
@@ -4323,7 +4323,7 @@ pub fn analyzeExport(
     const exported_decl = mod.declPtr(exported_decl_index);
     // TODO run the same checks as we do for C ABI struct fields
     switch (exported_decl.ty.zigTypeTag()) {
-        .Fn, .Int, .Enum, .Struct, .Union, .Array, .Float => {},
+        .Fn, .Int, .Enum, .Struct, .Union, .Array, .Float, .Pointer, .Optional => {},
         else => return sema.fail(block, src, "unable to export type '{}'", .{
             exported_decl.ty.fmt(sema.mod),
         }),
@@ -10535,7 +10535,11 @@ fn zirAsm(
     var output_type_bits = extra.data.output_type_bits;
     var needed_capacity: usize = @typeInfo(Air.Asm).Struct.fields.len + outputs_len + inputs_len;
 
-    const Output = struct { constraint: []const u8, ty: Type };
+    const Output = struct {
+        constraint: []const u8,
+        name: []const u8,
+        ty: Type,
+    };
     const output: ?Output = if (outputs_len == 0) null else blk: {
         const output = sema.code.extraData(Zir.Inst.Asm.Output, extra_i);
         extra_i = output.end;
@@ -10548,10 +10552,12 @@ fn zirAsm(
         }
 
         const constraint = sema.code.nullTerminatedString(output.data.constraint);
-        needed_capacity += constraint.len / 4 + 1;
+        const name = sema.code.nullTerminatedString(output.data.name);
+        needed_capacity += (constraint.len + name.len + (2 + 3)) / 4;
 
         break :blk Output{
             .constraint = constraint,
+            .name = name,
             .ty = try sema.resolveType(block, ret_ty_src, output.data.operand),
         };
     };
@@ -10573,7 +10579,7 @@ fn zirAsm(
 
         const constraint = sema.code.nullTerminatedString(input.data.constraint);
         const name = sema.code.nullTerminatedString(input.data.name);
-        needed_capacity += (constraint.len + name.len + 1) / 4 + 1;
+        needed_capacity += (constraint.len + name.len + (2 + 3)) / 4;
         inputs[arg_i] = .{ .c = constraint, .n = name };
     }
 
@@ -10611,7 +10617,9 @@ fn zirAsm(
         const buffer = mem.sliceAsBytes(sema.air_extra.unusedCapacitySlice());
         mem.copy(u8, buffer, o.constraint);
         buffer[o.constraint.len] = 0;
-        sema.air_extra.items.len += o.constraint.len / 4 + 1;
+        mem.copy(u8, buffer[o.constraint.len + 1 ..], o.name);
+        buffer[o.constraint.len + 1 + o.name.len] = 0;
+        sema.air_extra.items.len += (o.constraint.len + o.name.len + (2 + 3)) / 4;
     }
     for (inputs) |input| {
         const buffer = mem.sliceAsBytes(sema.air_extra.unusedCapacitySlice());
@@ -10619,7 +10627,7 @@ fn zirAsm(
         buffer[input.c.len] = 0;
         mem.copy(u8, buffer[input.c.len + 1 ..], input.n);
         buffer[input.c.len + 1 + input.n.len] = 0;
-        sema.air_extra.items.len += (input.c.len + input.n.len + 1) / 4 + 1;
+        sema.air_extra.items.len += (input.c.len + input.n.len + (2 + 3)) / 4;
     }
     for (clobbers) |clobber| {
         const buffer = mem.sliceAsBytes(sema.air_extra.unusedCapacitySlice());
diff --git a/src/arch/aarch64/CodeGen.zig b/src/arch/aarch64/CodeGen.zig
index 5ed7b63db3..3b27633f69 100644
--- a/src/arch/aarch64/CodeGen.zig
+++ b/src/arch/aarch64/CodeGen.zig
@@ -3272,10 +3272,12 @@ fn airAsm(self: *Self, inst: Air.Inst.Index) !void {
             if (output != .none) {
                 return self.fail("TODO implement codegen for non-expr asm", .{});
             }
+            const extra_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
             const constraint = std.mem.sliceTo(std.mem.sliceAsBytes(self.air.extra[extra_i..]), 0);
+            const name = std.mem.sliceTo(extra_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
             // for the string, we still use the next u32 for the null terminator.
-            extra_i += constraint.len / 4 + 1;
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             break constraint;
         } else null;
@@ -3283,10 +3285,10 @@ fn airAsm(self: *Self, inst: Air.Inst.Index) !void {
         for (inputs) |input| {
             const input_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
             const constraint = std.mem.sliceTo(input_bytes, 0);
-            const input_name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
+            const name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
             // for the string, we still use the next u32 for the null terminator.
-            extra_i += (constraint.len + input_name.len + 1) / 4 + 1;
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             if (constraint.len < 3 or constraint[0] != '{' or constraint[constraint.len - 1] != '}') {
                 return self.fail("unrecognized asm input constraint: '{s}'", .{constraint});
diff --git a/src/arch/arm/CodeGen.zig b/src/arch/arm/CodeGen.zig
index 73f51f6481..87d51b0276 100644
--- a/src/arch/arm/CodeGen.zig
+++ b/src/arch/arm/CodeGen.zig
@@ -4078,10 +4078,12 @@ fn airAsm(self: *Self, inst: Air.Inst.Index) !void {
             if (output != .none) {
                 return self.fail("TODO implement codegen for non-expr asm", .{});
             }
+            const extra_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
             const constraint = std.mem.sliceTo(std.mem.sliceAsBytes(self.air.extra[extra_i..]), 0);
+            const name = std.mem.sliceTo(extra_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
             // for the string, we still use the next u32 for the null terminator.
-            extra_i += constraint.len / 4 + 1;
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             break constraint;
         } else null;
@@ -4089,10 +4091,10 @@ fn airAsm(self: *Self, inst: Air.Inst.Index) !void {
         for (inputs) |input| {
             const input_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
             const constraint = std.mem.sliceTo(input_bytes, 0);
-            const input_name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
+            const name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
             // for the string, we still use the next u32 for the null terminator.
-            extra_i += (constraint.len + input_name.len + 1) / 4 + 1;
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             if (constraint.len < 3 or constraint[0] != '{' or constraint[constraint.len - 1] != '}') {
                 return self.fail("unrecognized asm input constraint: '{s}'", .{constraint});
diff --git a/src/arch/riscv64/CodeGen.zig b/src/arch/riscv64/CodeGen.zig
index 61fddee207..96d30c31ce 100644
--- a/src/arch/riscv64/CodeGen.zig
+++ b/src/arch/riscv64/CodeGen.zig
@@ -2098,10 +2098,12 @@ fn airAsm(self: *Self, inst: Air.Inst.Index) !void {
             if (output != .none) {
                 return self.fail("TODO implement codegen for non-expr asm", .{});
             }
+            const extra_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
             const constraint = std.mem.sliceTo(std.mem.sliceAsBytes(self.air.extra[extra_i..]), 0);
+            const name = std.mem.sliceTo(extra_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
             // for the string, we still use the next u32 for the null terminator.
-            extra_i += constraint.len / 4 + 1;
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             break constraint;
         } else null;
@@ -2109,10 +2111,10 @@ fn airAsm(self: *Self, inst: Air.Inst.Index) !void {
         for (inputs) |input| {
             const input_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
             const constraint = std.mem.sliceTo(input_bytes, 0);
-            const input_name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
+            const name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
             // for the string, we still use the next u32 for the null terminator.
-            extra_i += (constraint.len + input_name.len + 1) / 4 + 1;
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             if (constraint.len < 3 or constraint[0] != '{' or constraint[constraint.len - 1] != '}') {
                 return self.fail("unrecognized asm input constraint: '{s}'", .{constraint});
diff --git a/src/arch/sparcv9/CodeGen.zig b/src/arch/sparcv9/CodeGen.zig
index bcd8cf8eeb..7d93916fc1 100644
--- a/src/arch/sparcv9/CodeGen.zig
+++ b/src/arch/sparcv9/CodeGen.zig
@@ -642,10 +642,12 @@ fn airAsm(self: *Self, inst: Air.Inst.Index) !void {
             if (output != .none) {
                 return self.fail("TODO implement codegen for non-expr asm", .{});
             }
+            const extra_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
             const constraint = std.mem.sliceTo(std.mem.sliceAsBytes(self.air.extra[extra_i..]), 0);
+            const name = std.mem.sliceTo(extra_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
             // for the string, we still use the next u32 for the null terminator.
-            extra_i += constraint.len / 4 + 1;
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             break constraint;
         } else null;
@@ -653,10 +655,10 @@ fn airAsm(self: *Self, inst: Air.Inst.Index) !void {
         for (inputs) |input| {
             const input_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
             const constraint = std.mem.sliceTo(input_bytes, 0);
-            const input_name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
+            const name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
             // for the string, we still use the next u32 for the null terminator.
-            extra_i += (constraint.len + input_name.len + 1) / 4 + 1;
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             if (constraint.len < 3 or constraint[0] != '{' or constraint[constraint.len - 1] != '}') {
                 return self.fail("unrecognized asm input constraint: '{s}'", .{constraint});
diff --git a/src/arch/x86_64/CodeGen.zig b/src/arch/x86_64/CodeGen.zig
index ed85874413..c8dab1c500 100644
--- a/src/arch/x86_64/CodeGen.zig
+++ b/src/arch/x86_64/CodeGen.zig
@@ -4738,10 +4738,12 @@ fn airAsm(self: *Self, inst: Air.Inst.Index) !void {
             if (output != .none) {
                 return self.fail("TODO implement codegen for non-expr asm", .{});
             }
+            const extra_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
             const constraint = std.mem.sliceTo(std.mem.sliceAsBytes(self.air.extra[extra_i..]), 0);
+            const name = std.mem.sliceTo(extra_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
             // for the string, we still use the next u32 for the null terminator.
-            extra_i += constraint.len / 4 + 1;
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             break constraint;
         } else null;
@@ -4749,10 +4751,10 @@ fn airAsm(self: *Self, inst: Air.Inst.Index) !void {
         for (inputs) |input| {
             const input_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
             const constraint = std.mem.sliceTo(input_bytes, 0);
-            const input_name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
+            const name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
             // for the string, we still use the next u32 for the null terminator.
-            extra_i += (constraint.len + input_name.len + 1) / 4 + 1;
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             if (constraint.len < 3 or constraint[0] != '{' or constraint[constraint.len - 1] != '}') {
                 return self.fail("unrecognized asm input constraint: '{s}'", .{constraint});
diff --git a/src/codegen/c.zig b/src/codegen/c.zig
index 2cd93d47fc..44b616c493 100644
--- a/src/codegen/c.zig
+++ b/src/codegen/c.zig
@@ -3018,10 +3018,12 @@ fn airAsm(f: *Function, inst: Air.Inst.Index) !CValue {
         if (output != .none) {
             return f.fail("TODO implement codegen for non-expr asm", .{});
         }
+        const extra_bytes = std.mem.sliceAsBytes(f.air.extra[extra_i..]);
         const constraint = std.mem.sliceTo(std.mem.sliceAsBytes(f.air.extra[extra_i..]), 0);
+        const name = std.mem.sliceTo(extra_bytes[constraint.len + 1 ..], 0);
         // This equation accounts for the fact that even if we have exactly 4 bytes
         // for the string, we still use the next u32 for the null terminator.
-        extra_i += constraint.len / 4 + 1;
+        extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
         break constraint;
     } else null;
@@ -3031,10 +3033,12 @@ fn airAsm(f: *Function, inst: Air.Inst.Index) !CValue {
 
     const inputs_extra_begin = extra_i;
     for (inputs) |input, i| {
-        const constraint = std.mem.sliceTo(std.mem.sliceAsBytes(f.air.extra[extra_i..]), 0);
+        const input_bytes = std.mem.sliceAsBytes(f.air.extra[extra_i..]);
+        const constraint = std.mem.sliceTo(input_bytes, 0);
+        const name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
         // This equation accounts for the fact that even if we have exactly 4 bytes
         // for the string, we still use the next u32 for the null terminator.
-        extra_i += constraint.len / 4 + 1;
+        extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
         if (constraint[0] == '{' and constraint[constraint.len - 1] == '}') {
             const reg = constraint[1 .. constraint.len - 1];
diff --git a/src/codegen/llvm.zig b/src/codegen/llvm.zig
index c9ea5bebac..6c74fc8223 100644
--- a/src/codegen/llvm.zig
+++ b/src/codegen/llvm.zig
@@ -340,6 +340,10 @@ pub const Object = struct {
 
         llvm_module.setModuleDataLayout(target_data);
 
+        if (options.pic) llvm_module.setModulePICLevel();
+        if (options.pie) llvm_module.setModulePIELevel();
+        if (code_model != .Default) llvm_module.setModuleCodeModel(code_model);
+
         return Object{
             .gpa = gpa,
             .module = options.module.?,
@@ -4577,6 +4581,10 @@ pub const FuncGen = struct {
         const operand_ty = self.air.typeOf(pl_op.operand);
         const name = self.air.nullTerminatedString(pl_op.payload);
 
+        if (needDbgVarWorkaround(self.dg, operand_ty)) {
+            return null;
+        }
+
         const di_local_var = dib.createAutoVariable(
             self.di_scope.?,
             name.ptr,
@@ -4638,14 +4646,19 @@ pub const FuncGen = struct {
         var llvm_param_i: usize = 0;
         var total_i: usize = 0;
 
+        var name_map: std.StringArrayHashMapUnmanaged(void) = .{};
+        try name_map.ensureUnusedCapacity(arena, outputs.len + inputs.len);
+
         for (outputs) |output| {
             if (output != .none) {
                 return self.todo("implement inline asm with non-returned output", .{});
             }
+            const extra_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
             const constraint = std.mem.sliceTo(std.mem.sliceAsBytes(self.air.extra[extra_i..]), 0);
+            const name = std.mem.sliceTo(extra_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
             // for the string, we still use the next u32 for the null terminator.
-            extra_i += constraint.len / 4 + 1;
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             try llvm_constraints.ensureUnusedCapacity(self.gpa, constraint.len + 1);
             if (total_i != 0) {
@@ -4654,17 +4667,17 @@ pub const FuncGen = struct {
             llvm_constraints.appendAssumeCapacity('=');
             llvm_constraints.appendSliceAssumeCapacity(constraint[1..]);
 
+            name_map.putAssumeCapacityNoClobber(name, {});
             total_i += 1;
         }
 
-        const input_start_extra_i = extra_i;
         for (inputs) |input| {
-            const input_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
-            const constraint = std.mem.sliceTo(input_bytes, 0);
-            const input_name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
+            const extra_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
+            const constraint = std.mem.sliceTo(extra_bytes, 0);
+            const name = std.mem.sliceTo(extra_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
             // for the string, we still use the next u32 for the null terminator.
-            extra_i += (constraint.len + input_name.len + 1) / 4 + 1;
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             const arg_llvm_value = try self.resolveInst(input);
 
@@ -4677,6 +4690,7 @@ pub const FuncGen = struct {
             }
             llvm_constraints.appendSliceAssumeCapacity(constraint);
 
+            name_map.putAssumeCapacityNoClobber(name, {});
             llvm_param_i += 1;
             total_i += 1;
         }
@@ -4739,20 +4753,11 @@ pub const FuncGen = struct {
                         const name = asm_source[name_start..i];
                         state = .start;
 
-                        extra_i = input_start_extra_i;
-                        for (inputs) |_, input_i| {
-                            const input_bytes = std.mem.sliceAsBytes(self.air.extra[extra_i..]);
-                            const constraint = std.mem.sliceTo(input_bytes, 0);
-                            const input_name = std.mem.sliceTo(input_bytes[constraint.len + 1 ..], 0);
-                            extra_i += (constraint.len + input_name.len + 1) / 4 + 1;
-
-                            if (std.mem.eql(u8, name, input_name)) {
-                                try rendered_template.writer().print("{d}", .{input_i});
-                                break;
-                            }
-                        } else {
-                            return self.todo("TODO validate asm in Sema", .{});
-                        }
+                        const index = name_map.getIndex(name) orelse {
+                            // we should validate the assembly in Sema; by now it is too late
+                            return self.todo("unknown input or output name: '{s}'", .{name});
+                        };
+                        try rendered_template.writer().print("{d}", .{index});
                     },
                     else => {},
                 },
@@ -6150,6 +6155,10 @@ pub const FuncGen = struct {
 
         const inst_ty = self.air.typeOfIndex(inst);
         if (self.dg.object.di_builder) |dib| {
+            if (needDbgVarWorkaround(self.dg, inst_ty)) {
+                return arg_val;
+            }
+
             const src_index = self.getSrcArgIndex(self.arg_index - 1);
             const func = self.dg.decl.getFunction().?;
             const lbrace_line = self.dg.module.declPtr(func.owner_decl).src_line + func.lbrace_line + 1;
@@ -8251,3 +8260,15 @@ const AnnotatedDITypePtr = enum(usize) {
 };
 
 const lt_errors_fn_name = "__zig_lt_errors_len";
+
+/// Without this workaround, LLVM crashes with "unknown codeview register H1"
+/// TODO use llvm-reduce and file upstream LLVM bug for this.
+fn needDbgVarWorkaround(dg: *DeclGen, ty: Type) bool {
+    if (ty.tag() == .f16) {
+        const target = dg.module.getTarget();
+        if (target.os.tag == .windows and target.cpu.arch == .aarch64) {
+            return true;
+        }
+    }
+    return false;
+}
diff --git a/src/codegen/llvm/bindings.zig b/src/codegen/llvm/bindings.zig
index b8dc3e1830..560b9544dd 100644
--- a/src/codegen/llvm/bindings.zig
+++ b/src/codegen/llvm/bindings.zig
@@ -310,6 +310,15 @@ pub const Module = opaque {
     pub const setModuleDataLayout = LLVMSetModuleDataLayout;
     extern fn LLVMSetModuleDataLayout(*const Module, *const TargetData) void;
 
+    pub const setModulePICLevel = ZigLLVMSetModulePICLevel;
+    extern fn ZigLLVMSetModulePICLevel(module: *const Module) void;
+
+    pub const setModulePIELevel = ZigLLVMSetModulePIELevel;
+    extern fn ZigLLVMSetModulePIELevel(module: *const Module) void;
+
+    pub const setModuleCodeModel = ZigLLVMSetModuleCodeModel;
+    extern fn ZigLLVMSetModuleCodeModel(module: *const Module, code_model: CodeModel) void;
+
     pub const addFunction = LLVMAddFunction;
     extern fn LLVMAddFunction(*const Module, Name: [*:0]const u8, FunctionTy: *const Type) *const Value;
 
diff --git a/src/print_air.zig b/src/print_air.zig
index 6e336e138b..c01d96ed7f 100644
--- a/src/print_air.zig
+++ b/src/print_air.zig
@@ -542,15 +542,19 @@ const Writer = struct {
         extra_i += inputs.len;
 
         for (outputs) |output| {
-            const constraint = w.air.nullTerminatedString(extra_i);
+            const extra_bytes = std.mem.sliceAsBytes(w.air.extra[extra_i..]);
+            const constraint = std.mem.sliceTo(extra_bytes, 0);
+            const name = std.mem.sliceTo(extra_bytes[constraint.len + 1 ..], 0);
+
             // This equation accounts for the fact that even if we have exactly 4 bytes
-            // for the string, we still use the next u32 for the null terminator.
-            extra_i += constraint.len / 4 + 1;
+            // for the strings and their null terminators, we still use the next u32
+            // for the null terminator.
+            extra_i += (constraint.len + name.len + (2 + 3)) / 4;
 
             if (output == .none) {
-                try s.print(", -> {s}", .{constraint});
+                try s.print(", [{s}] -> {s}", .{ name, constraint });
             } else {
-                try s.print(", out {s} = (", .{constraint});
+                try s.print(", [{s}] out {s} = (", .{ name, constraint });
                 try w.writeOperand(s, inst, op_index, output);
                 op_index += 1;
                 try s.writeByte(')');
@@ -558,12 +562,15 @@ const Writer = struct {
         }
 
         for (inputs) |input| {
-            const constraint = w.air.nullTerminatedString(extra_i);
+            const extra_bytes = std.mem.sliceAsBytes(w.air.extra[extra_i..]);
+            const constraint = std.mem.sliceTo(extra_bytes, 0);
+            const name = std.mem.sliceTo(extra_bytes[constraint.len + 1 ..], 0);
             // This equation accounts for the fact that even if we have exactly 4 bytes
-            // for the string, we still use the next u32 for the null terminator.
-            extra_i += constraint.len / 4 + 1;
+            // for the strings and their null terminators, we still use the next u32
+            // for the null terminator.
+            extra_i += (constraint.len + name.len + 1) / 4 + 1;
 
-            try s.print(", in {s} = (", .{constraint});
+            try s.print(", [{s}] in {s} = (", .{ name, constraint });
             try w.writeOperand(s, inst, op_index, input);
             op_index += 1;
             try s.writeByte(')');
@@ -572,7 +579,8 @@ const Writer = struct {
         {
             var clobber_i: u32 = 0;
             while (clobber_i < clobbers_len) : (clobber_i += 1) {
-                const clobber = w.air.nullTerminatedString(extra_i);
+                const extra_bytes = std.mem.sliceAsBytes(w.air.extra[extra_i..]);
+                const clobber = std.mem.sliceTo(extra_bytes, 0);
                 // This equation accounts for the fact that even if we have exactly 4 bytes
                 // for the string, we still use the next u32 for the null terminator.
                 extra_i += clobber.len / 4 + 1;
diff --git a/test/behavior/struct.zig b/test/behavior/struct.zig
index 89f6e20aa0..b05126db62 100644
--- a/test/behavior/struct.zig
+++ b/test/behavior/struct.zig
@@ -421,21 +421,10 @@ test "packed struct 24bits" {
     if (builtin.cpu.arch == .arm) return error.SkipZigTest; // TODO
 
     comptime {
-        // TODO Remove if and leave only the else branch when it is also fixed in stage2
-        if (builtin.zig_backend == .stage2_llvm or builtin.zig_backend == .stage2_x86 or
-            builtin.zig_backend == .stage2_riscv64)
-        {
-            // Stage 2 still expects the wrong values
-            try expect(@sizeOf(Foo24Bits) == 4);
-            if (@sizeOf(usize) == 4) {
-                try expect(@sizeOf(Foo96Bits) == 12);
-            } else {
-                try expect(@sizeOf(Foo96Bits) == 16);
-            }
-        } else {
-            // Stage1 is now fixed and is expected to return right values
-            try expectEqual(@sizeOf(Foo24Bits), 3);
-            try expectEqual(@sizeOf(Foo96Bits), 12);
+        // stage1 gets the wrong answer for sizeof
+        if (builtin.zig_backend != .stage1) {
+            std.debug.assert(@sizeOf(Foo24Bits) == @sizeOf(u24));
+            std.debug.assert(@sizeOf(Foo96Bits) == @sizeOf(u96));
         }
     }
 
diff --git a/test/incremental/aarch64-macos/hello_world_with_updates.0.zig b/test/incremental/aarch64-macos/hello_world_with_updates.0.zig
index 11a379a9a9..300baac1a4 100644
--- a/test/incremental/aarch64-macos/hello_world_with_updates.0.zig
+++ b/test/incremental/aarch64-macos/hello_world_with_updates.0.zig
@@ -2,4 +2,4 @@
 // output_mode=Exe
 // target=aarch64-macos
 //
-// :109:9: error: struct 'tmp.tmp' has no member named 'main'
+// :107:9: error: struct 'tmp.tmp' has no member named 'main'
diff --git a/test/incremental/x86_64-linux/hello_world_with_updates.0.zig b/test/incremental/x86_64-linux/hello_world_with_updates.0.zig
index 960fae5e64..47b5e9db7a 100644
--- a/test/incremental/x86_64-linux/hello_world_with_updates.0.zig
+++ b/test/incremental/x86_64-linux/hello_world_with_updates.0.zig
@@ -2,4 +2,4 @@
 // output_mode=Exe
 // target=x86_64-linux
 //
-// :109:9: error: struct 'tmp.tmp' has no member named 'main'
+// :107:9: error: struct 'tmp.tmp' has no member named 'main'
diff --git a/test/incremental/x86_64-macos/hello_world_with_updates.0.zig b/test/incremental/x86_64-macos/hello_world_with_updates.0.zig
index 34440c4603..dd43748e6e 100644
--- a/test/incremental/x86_64-macos/hello_world_with_updates.0.zig
+++ b/test/incremental/x86_64-macos/hello_world_with_updates.0.zig
@@ -2,4 +2,4 @@
 // output_mode=Exe
 // target=x86_64-macos
 //
-// :109:9: error: struct 'tmp.tmp' has no member named 'main'
+// :107:9: error: struct 'tmp.tmp' has no member named 'main'