Packed struct operations are not emitted as operations on their backing integer

[Validark]

Zig Version

0.14.0-dev.2097+d30e28754

Steps to Reproduce and Observed Behavior

This code: Godbolt link

const Pack = packed struct
{
    field1: bool,
    field2: bool,
    field3: bool,
};

export fn init(pack: *Pack) void {
    pack.field1 = true;
    pack.field2 = false;
    pack.field3 = true;
}

export fn init2(pack: *Pack) void {
    pack.* = .{
        .field1 = true,
        .field2 = false,
        .field3 = true,
    };
}

Results in this assembly for x86-64:

init:
        movzx   eax, byte ptr [rdi]
        and     al, -8
        or      al, 5
        mov     byte ptr [rdi], al
        ret

init2:
        mov     byte ptr [rdi], 5
        ret

Running this command: zig build-obj ./src/llvm_code.zig -O ReleaseFast -target x86_64-linux -mcpu znver4 --verbose-llvm-ir -fstrip >llvm_code.ll 2>&1 sheds a little light on the situation:

define dso_local void @init(ptr align 1 nonnull %0) #0 {
1:
  %2 = alloca [8 x i8], align 8
  %3 = alloca [8 x i8], align 8
  %4 = alloca [8 x i8], align 8
  store ptr %0, ptr %4, align 8
  %5 = load ptr, ptr %4, align 8
  %6 = load i8, ptr %5, align 1
  %7 = zext i1 true to i8
  %8 = shl i8 %7, 0
  %9 = xor i8 %8, -1
  %10 = and i8 %6, %9
  %11 = zext i1 true to i8
  %12 = shl i8 %11, 0
  %13 = or i8 %12, %10
  store i8 %13, ptr %5, align 1
  store ptr %0, ptr %3, align 8
  %14 = load ptr, ptr %3, align 8
  %15 = load i8, ptr %14, align 1
  %16 = zext i1 true to i8
  %17 = shl i8 %16, 1
  %18 = xor i8 %17, -1
  %19 = and i8 %15, %18
  %20 = zext i1 false to i8
  %21 = shl i8 %20, 1
  %22 = or i8 %21, %19
  store i8 %22, ptr %14, align 1
  store ptr %0, ptr %2, align 8
  %23 = load ptr, ptr %2, align 8
  %24 = load i8, ptr %23, align 1
  %25 = zext i1 true to i8
  %26 = shl i8 %25, 2
  %27 = xor i8 %26, -1
  %28 = and i8 %24, %27
  %29 = zext i1 true to i8
  %30 = shl i8 %29, 2
  %31 = or i8 %30, %28
  store i8 %31, ptr %23, align 1
  ret void
}

; Function Attrs: nounwind uwtable nosanitize_coverage skipprofile
define dso_local void @init2(ptr align 1 nonnull %0) #0 {
1:
  store i3 -3, ptr %0, align 1
  ret void
}

This Godbolt link shows that if we change the i8's to i3's in the LLVM IR of @init we can get the optimization we want out of LLVM. I did not verify that this makes 100% sense to do, but potentially this issue can be solved by using operations on i3's rather than i8's.

Expected Behavior

Both functions should compile down to mov byte ptr [rdi], 5

[Validark]

Potentially relevant considerations:

#14200
#17391

"When loading a value of a type like i20 with a size that is not an integral number of bytes, the result is undefined if the value was not originally written using a store of the same type."

[scheibo]

FWIW this is what I suspect is the bulk of the regression reported in #17768