[llvm-commits] PATCH: CONCAT_VECTORS expansion through the stack generated incorrect stack offsets, causing broken code to be generated

[Eli Friedman]

On Thu, May 3, 2012 at 6:41 AM, Duncan Sands <baldrick at free.fr> wrote:
> Hi,
>
> On 03/05/12 15:24, Heikki Kultala wrote:
>> The expansion of CONCAT_VECTORS goes into
>> SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node).
>>
>> The routine calculated the offsets of the stores from the size of the element
>> size of the result vector. This works when the routine is used to expand
>> BUILD_VECTOR of scalars, but it works incorrectly for CONCAT_VECTORS where one
>> operand has bigger size than single element of the result vector.
>>
>> This patch fixes the store offsets to be calculated by the size of the input
>> operands instead of the size of the elements in the result vectors.
>
> can you please add a testcase.  Also, are you saying that the result element
> type of CONCAT_VECTORS doesn't have to agree with the input element type, eg
> that <4 x i32> concat_vectors <2 x i64>, <2 x i64> is possible?  If so, this
> isn't documented in ISDOpcodes.h so can you please document it there.
>
>> I hope this patch makes it into llvm 3.1 as this bug causes broken code to be
>> generated on our code(tce).
>
> It's probably too late, but you can always ask Bill (CC'd).

Yes, it's rather late... especially given it doesn't affect x86.

-Eli