[llvm-commits] [ARM] Scheduler FixedFrame object vs byval formal parameters

[Stepan Dyatkovskiy]

Hello.
Please find the patch in attachment. It adds two new fields to the 
InputArg structure: Function's argument index and InputArg's part offset 
in bytes relative to the start position of Function's argument. E.g.: If 
function's argument is 128 bit width and it was splitted onto 32bit 
regs, then we got 4 InputArg structs with same arg index, but different 
offset values.

I also modified ARMISelLowering, to fix wrong scheduling issue. The fix 
is similar to r164553, but it uses new InputArg fields.

-Stepan.

Manman Ren wrote:
>
> On Oct 5, 2012, at 2:59 AM, Stepan Dyatkovskiy <stpworld at narod.ru> wrote:
>
>> Thanks. The idea looks fine. But r164553 assumes that size of Function's arguments set is equal to size of InputArgs set:
>> "for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo, ++FuncArg) {"
>>
>> But seems that it not always right. In SelectionDAGISel::LowerArguments we can see that Function argument be a set of multiple VTs, in that case multiple InputArg flags will be created for single argument. Try test your commit for complex params (structures or arrays).
> I agree with you here. A function argument can be expanded to a set of VTs, and each VT can be further partitioned with TLI.getNumRegisters.
>
>>
>> Perhaps, we can provide InputArg type with some info about Function's argument and offset?
> I am okay with this approach, perhaps an index of the original function argument.
>
> Thanks,
> Manman
>
>>
>> -Stepan.
>>
>> Sergei Larin wrote:
>>> Stepan,
>>>
>>>    ...also see r164553.
>>>
>>
>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: byval-scheduler-fix-2.patch
Type: text/x-patch
Size: 5906 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20121008/15c98360/attachment.bin>