Support for bitfields

[bcardosolopes]

Involves adding new CIR operations and CodeGen support.

[felipealmeida]

Hello @bcardosolopes ,

I'm moving all discussion to here. About your comment: "It might also be worth checking out how polygeist does some of that lowering, could provide interesting insights.". Unfortunately I haven't found absolutely anything about bitfields in Polygeist. I looked online and shallowly in the source code. I'll take a deeper look soon, but it may very well be that we are the first, specially because bitfields are not very well seen and hardly does what people expect it to do (standard-wise), but it still useful where implementation defined behavior is OK (which is my case).

[bcardosolopes]

Ok, looks like we have some interesting land to explore, thanks for taking a look @felipealmeida - we much very want to see support for bitfields too.

[felipealmeida]

Hello, I'm using this example:

struct field {
  char c;
  unsigned i0 : 8;
  unsigned i1 : 8;
  unsigned i2 : 8;
  unsigned i3 : 8;
  char d;
};


void foo (struct field* f, char a) {

}

Generating this (as a test):

#loc2 = loc(fused["test.c":12:11, "test.c":12:25])
#loc3 = loc(fused["test.c":12:28, "test.c":12:33])
module {
  cir.func @foo(%arg0: !cir.ptr<!cir.struct<"struct.field", i8, !cir.bitfield<i32, "offsets"<0, 8, 16, 24>, "sizes"<8, 8, 8, 8> >, i8>> loc(fused["test.c":12:11, "test.c":12:25]), %arg1: i8 loc(fused["test.c":12:28, "test.c":12:33])) {
    %0 = cir.alloca !cir.ptr<!cir.struct<"struct.field", i8, !cir.bitfield<i32, "offsets"<0, 8, 16, 24>, "sizes"<8, 8, 8, 8> >, i8>>, cir.ptr <!cir.ptr<!cir.struct<"struct.field", i8, !cir.bitfield<i32, "offsets"<0, 8, 16, 24>, "sizes"<8, 8, 8, 8> >, i8>>>, ["f", paraminit] {alignment = 8 : i64} loc(#loc2)
    %1 = cir.alloca i8, cir.ptr <i8>, ["a", paraminit] {alignment = 1 : i64} loc(#loc3)
    cir.store %arg0, %0 : !cir.ptr<!cir.struct<"struct.field", i8, !cir.bitfield<i32, "offsets"<0, 8, 16, 24>, "sizes"<8, 8, 8, 8> >, i8>>, cir.ptr <!cir.ptr<!cir.struct<"struct.field", i8, !cir.bitfield<i32, "offsets"<0, 8, 16, 24>, "sizes"<8, 8, 8, 8> >, i8>>> loc(#loc4)
    cir.store %arg1, %1 : i8, cir.ptr <i8> loc(#loc4)
    cir.return loc(#loc5)
  } loc(#loc1)
} loc(#loc0)
#loc0 = loc(unknown)
#loc1 = loc(fused["test.c":12:1, "test.c":14:1])
#loc4 = loc("test.c":12:36)
#loc5 = loc("test.c":14:1)

I don't think the current bitfield definition is good, probably not even legal, maybe cir.bitfield<i32, [0, 8], [8, 8], [16, 8], [24, 8]>. Let me know if anyone has any criticism.

Kind regards,

[gitoleg]

Hey,
as far as I understand there is no progress on this issue, so can you give an advice on implementation? The part I think about is mlir::LValue that is similar to clang::LValue, but lack the union with BitFieldInfo, and looks like it's not easy to implement it, since the value is not stored by pointer. So... Any ideas? Thank you!

[gitoleg]

Is it a right place to discuss CIR design questions?