stage2 llvm: pointer access with excess alignment, causing segfault

[topolarity]

Zig Version

0.10.0-dev.4472+a4eb221b9

Steps to Reproduce

Build Zig and save the LLVM IR:

$ /build/zig2 build -p stage3_debug -Dstrip=true -Denable-llvm=true --verbose-llvm-ir 2>llvm_ir.txt

If you examine Autodocs.walkInstruction in the generated LLVM IR, you'll see this:

; Function Attrs: nounwind sspstrong uwtable
define internal fastcc void @Autodoc.walkInstruction(ptr noalias nonnull sret({ %Autodoc.DocData.WalkResult, i16, [14 x i8] }) %0, ptr nonnull align 8 %1, ptr nonnull align 8 %2, ptr nonnull align 8 %3, ptr nonnull readonly align 8 %4, i64 %5, i1 %6) unnamed_addr #0 {
Entry:
  ...
  %891 = alloca %Autodoc.DocData.Type, align 8
  ...
  %2849 = getelementptr inbounds { %Autodoc.DocData.Type.Type__struct_34832, i5, [15 x i8] }, ptr %891, i32 0, i32 0
  call void @llvm.memcpy.p0.p0.i64(ptr align 16 %2849, ptr align 16 %892, i64 592, i1 false)
  ...
}

Zig claims that %2849 has alignment 16, but it actually has alignment 8.

Expected Behavior

Alignment should be consistent between alloca and memcpy.

Actual Behavior

Alignments do not match.

[topolarity]

This LLVM IR is generated from this part of Autodocs.walkInstruction, at line 1581:

const type_slot_index = self.types.items.len;
try self.types.append(self.arena, .{
    .Pointer = .{
        .size = ptr.size,
        .child = elem_type_ref.expr,
        .has_align = ptr.flags.has_align,
        .@"align" = @"align",
        .has_addrspace = ptr.flags.has_addrspace,
        .address_space = address_space,
        .has_sentinel = ptr.flags.has_sentinel,
        .sentinel = sentinel,
        .is_mutable = ptr.flags.is_mutable,
        .is_volatile = ptr.flags.is_volatile,
        .has_bit_range = ptr.flags.has_bit_range,
        .bit_start = bit_start,
        .host_size = host_size,
    },
});

[topolarity]

In release modes, this memcpy sometimes gets lowered to a movaps instruction, which requires 16-byte alignment and causes a segfault otherwise.

That segfault is preventing #13074 from passing CI.

[topolarity]

Reduction:

const Type = union(enum) {
    Pointer: struct {
        child: u32 align(16),
    },
};

fn foo(x: Type) void {
    _ = x;
}

test {
    var child: u32 = 1;
    foo(.{ .Pointer = .{ .child = child } });
}

Generates an under-aligned alloca:

define internal fastcc i16 @test22.test_0(ptr nonnull %0) unnamed_addr #0 !dbg !4390 {
Entry:
  %1 = alloca %test22.Type, align 8
  ...