IRGen failure with conditional operator which produces an lvalue bitfield

[zygoloid]

Bugzilla Link	13809
Version	unspecified
OS	Linux
CC	@DougGregor

Extended Description

Clang dies on this:

struct S { int x : 5, y : 9; };
void f(bool c, S &a) { (c ? a.x : a.y) = 4; }

... saying:

:2:31: error: cannot compile this conditional operator yet
void f(bool c, S &a, S &b) { (c ? a.x : a.y) = 4; }
^~~~~~~~~~~~~

We're also getting Sema wrong for this case. We accept:

int &g(bool c, S &a, S &b) { return (c ? a.x : a.y); }

... but we should not, because it binds a reference to a bitfield.

[zygoloid]

*** Bug llvm/llvm-bugzilla-archive#16553 has been marked as a duplicate of this bug. ***

[llvmbot]

As a note, I also encounter this bug in recent versions (8.x, 9.x, 10.x) of clang.

[zygoloid]

The latter half of this bug is fixed; we properly reject binding a non-const reference to a bitfield conditional lvalue, and create a temporary when binding a const reference to a bitfield conditional lvalue.

We still can't generate code for the former half.

[llvmbot]

Is it a relatively easy modification to add support for lvalue bitfield conditionals, or would it take some major changes? I'd be willing to try to write a patch if it's the former.

[zygoloid]

It's somewhere in between. Most of the changes are mechanical consequences of adding the new representation, but it's not a trivial amount of work.

It would mean extending Clang's CodeGen layer to be able to represent a new kind of LValue (for a bitfield where the position within the object is a runtime value rather than a compile-time constant -- maybe tracking a mask and a shift for the lvalue, or maybe replacing the current bitfield representation with the more generalized form) and updating all the code that produces and consumes LValues to handle the new kind of LValue.

[llvmbot]

Thanks, Richard. It sounds like the necessary work is a bit beyond my scope for now.

[zygoloid]

mentioned in issue llvm/llvm-bugzilla-archive#16553

[llvmbot]

@llvm/issue-subscribers-clang-codegen

Author: None (ec04fc15-fa35-46f2-80e1-5d271f2ef708)

| | | | --- | --- | | Bugzilla Link | [13809](https://llvm.org/bz13809) | | Version | unspecified | | OS | Linux | | CC | @DougGregor |

Extended Description

Clang dies on this:

struct S { int x : 5, y : 9; };
void f(bool c, S &a) { (c ? a.x : a.y) = 4; }

... saying:

<stdin>:2:31: error: cannot compile this conditional operator yet
void f(bool c, S &a, S &b) { (c ? a.x : a.y) = 4; }
^~~~~~~~~~~~~

We're also getting Sema wrong for this case. We accept:

int &g(bool c, S &a, S &b) { return (c ? a.x : a.y); }

... but we should not, because it binds a reference to a bitfield.

[shafik]

Second example is now fixed but the first is still an issue: https://godbolt.org/z/jz1h3s7oq

@zygoloid