[llvm] r280783 - Remove unnecessary call to getAllocatableRegClass
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Thu Oct 27 16:41:36 PDT 2016
> On Oct 27, 2016, at 15:39, Justin Bogner <mail at justinbogner.com> wrote:
> Quentin Colombet via llvm-commits <llvm-commits at lists.llvm.org> writes:
>>> On Sep 30, 2016, at 12:07 PM, Krzysztof Parzyszek via llvm-commits
>>> <llvm-commits at lists.llvm.org> wrote:
>>> On 9/30/2016 1:46 PM, Justin Bogner via llvm-commits wrote:
>>>> The assert I'm seeing is in the case where `OpRC != nullptr` but
>>>> `ConstrainedRC == nullptr`. Quentin's suggestion would call
>>>> getAllocatableClass(DstRC) in this case, whereas in the committed code
>>>> we just go ahead and do MRI->createVirtualRegister(OpRC). The call to
>>>> createVirtualRegister asserts because OpRC isn't allocatable.
>>> This doesn't sound right. This implies that a register was created
>>> for the operand that is not compatible with the operand's
>>> class. Sounds like the problem is somewhere else.
>> That’s exactly what this code is supposed to fix IIRC. Thus, the
>> current patch does not sound right :).
> Ping. Matt, have you had a chance to look at this?
The problem is I want to use the output from constrainRegClass as the virtual register class instead of getAllocatableClass. Instead of the assert in the case where constrainRegClass fails, would it work to try using getAllocatableClass instead? i.e. getAllocatableClass is the fallback if constrainRegClass fails and then the same create vreg and copy
More information about the llvm-commits