[LLVMdev] Assert on Large Zeroinitializer Store

[Eli Friedman]

On Wed, Aug 24, 2011 at 12:07 PM, John Criswell <criswell at illinois.edu> wrote:
> On 8/24/11 1:51 PM, Eli Friedman wrote:
>>
>> On Wed, Aug 24, 2011 at 11:41 AM, John Criswell<criswell at illinois.edu>
>>  wrote:
>>>
>>> Dear All,
>>>
>>> I currently have one of my transforms creating the following store
>>> instruction:
>>>
>>> store [65536 x i8] zeroinitializer, [65536 x i8]* %buf.i, align 16
>>>
>>> ... which causes the SelectionDAG code to assert out:
>>>
>>> Assertion failed: (ResNo<  NumValues&&  "Illegal result number!"),
>>> function getValueType, file
>>> /Users/criswell/src/llvm/include/llvm/CodeGen/SelectionDAGNodes.h, line
>>> 589.
>>>
>>> Should LLVM be able to code generate the above store instruction?  If
>>> not, how large a value can a store instruction store to memory without
>>> triggering problems in the code generator?
>>
>> I don't know the exact limit, but if you're storing an array or struct
>> with more that 10 elements with a store instruction, you're doing
>> something wrong.
>
> Okay.
>
> I can certainly use memset (I'm just zero'ing out memory by storing a
> zeroinitializer).  Using stores just looked a little cleaner to me.
>
> Is there a reason why storing a zeroinitializer of such large size is
> considered "wrong?"  It seems like it wouldn't generate any performance hit
> if the code generator took care of it properly.

In practice, the way we do code generation for constructs involving
first-class structs/arrays assumes that it is reasonable to scalarize
the type in question.

-Eli