On Jul 6, 2007, at 4:00 AM, Rafael Espindola wrote:
>> The patch looks good. Some questions/comments:
>>
Sorry, just now catching up on this thread. :(
>> +++ include/llvm/CodeGen/SelectionDAGNodes.h (working copy)
>>
>> + SturtByVal = 1<<4, ///< Struct passed by value
>>
>> This looks like a typo "sturt"
> Changed to "ByVal" based on Duncan's comment.
Ok, sounds good.
>> +++ utils/TableGen/CallingConvEmitter.cpp (working copy)
>> @@ -67,12 +67,14 @@
>> O << IndentStr << "if (";
>>
>> if (Action->isSubClassOf("CCIfType")) {
>> + O << "!(ArgFlags & ISD::ParamFlags::StructByVal) && (";
>>
>>
>> I don't think CCIfType should automatically mask out byval
>> arguments. Instead, Plz add a new "notbyval" predicate or something
>> so that this logic is explicit in the .td files.
>
> The reason I did this is that if CCIfType matches "byval" arguments,
> the current backends will generate invalid code (handle it like a
> normal pointer). I have just removed the checks for now, since nothing
> is generating "byval"s.
>
> If a add a notbyval predicated, should I add it to all the current
> uses of CCIfType?
I'd suggest that we just let the backends upgrade as they add support
for byval. By default, they will codegen as if the thing is passed
like a pointer. Target authors should add support for byval if they
care, before llvm-gcc support goes in.
-Chris
_______________________________________________
llvm-commits mailing list
llvm-commits@cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
