Misaligned memory access causing segfault with f128 on i386

[jean-dao]

Zig Version

0.10.0-dev.657+f5471299d

Steps to Reproduce

foo.zig:

pub fn main() void {
    var buf: f128 = 0;
    var foo: f64 = @floatCast(f64, buf + 0);
    _ = foo;
}

Run zig run foo.zig -target i386-linux-none -mcpu pentium4 (will segfault).

Expected Behavior

No segfault.

Actual Behavior

Running zig run foo.zig -target i386-linux-none -mcpu pentium4 will segfault:

Segmentation fault at address 0x0
???:?:?: 0x42d426 in ??? (???)
foo.zig:3:40: 0x4258c3 in main (foo)
    var foo: f64 = @floatCast(f64, buf + 0);
                                       ^
/home/jean/src/zig/stage1/lib/zig/std/start.zig:551:22: 0x4203ea in std.start.callMain (foo)
            root.main();
                     ^
/home/jean/src/zig/stage1/lib/zig/std/start.zig:495:12: 0x405412 in std.start.callMainWithArgs (foo)
    return @call(.{ .modifier = .always_inline }, callMain, .{});
           ^
/home/jean/src/zig/stage1/lib/zig/std/start.zig:409:17: 0x404682 in std.start.posixCallMainAndExit (foo)
    std.os.exit(@call(.{ .modifier = .always_inline }, callMainWithArgs, .{ argc, argv, envp }));
                ^
/home/jean/src/zig/stage1/lib/zig/std/start.zig:322:5: 0x40452d in std.start._start (foo)
    @call(.{ .modifier = .never_inline }, posixCallMainAndExit, .{});
    ^
Aborted (core dumped)

Running the program from gdb, I get more info:

Program received signal SIGSEGV, Segmentation fault.
0x0042d426 in __addtf3 ()
(ins)(gdb) disas
Dump of assembler code for function __addtf3:
   0x0042d3f0 <+0>:     push   %ebp
   0x0042d3f1 <+1>:     mov    %esp,%ebp
   0x0042d3f3 <+3>:     push   %esi
   0x0042d3f4 <+4>:     sub    $0x14,%esp
   0x0042d3f7 <+7>:     mov    0x8(%ebp),%esi
   0x0042d3fa <+10>:    mov    0xc(%ebp),%edx
   0x0042d3fd <+13>:    sub    $0x4,%esp
   0x0042d400 <+16>:    lea    -0x18(%ebp),%ecx
   0x0042d403 <+19>:    push   0x28(%ebp)
   0x0042d406 <+22>:    push   0x24(%ebp)
   0x0042d409 <+25>:    push   0x20(%ebp)
   0x0042d40c <+28>:    push   0x1c(%ebp)
   0x0042d40f <+31>:    push   0x18(%ebp)
   0x0042d412 <+34>:    push   0x14(%ebp)
   0x0042d415 <+37>:    push   0x10(%ebp)
   0x0042d418 <+40>:    call   0x42d440 <compiler_rt.addXf3.addXf3.48>
   0x0042d41d <+45>:    add    $0x20,%esp
   0x0042d420 <+48>:    movaps -0x18(%ebp),%xmm0
   0x0042d424 <+52>:    mov    %esi,%eax
=> 0x0042d426 <+54>:    movaps %xmm0,(%esi)
   0x0042d429 <+57>:    add    $0x14,%esp
   0x0042d42c <+60>:    pop    %esi
   0x0042d42d <+61>:    pop    %ebp
   0x0042d42e <+62>:    ret    $0x4
End of assembler dump.
(ins)(gdb) i r esi
esi            0xffffd338          -11464

From my findings, movaps works only with align(16) addresses, which is not the case here. adding an align(16) to the foo variable in the example fixes the issue.