[LLVMdev] More Back-End Porting Troubles

[Fabian Scheler]

Hi,

first of all: thanks for your kind, very helpful and unbelievable fast response!

>> as mentioned in an earlier post
>> (http://lists.cs.uiuc.edu/pipermail/llvmdev/2012-July/051677.html) I
>> am currently working on a Back-End for the TriCore processor.
>> Currently, I am struggling as LLVM could not select zext and load, for
>> instance, so some of the testcases in test/CodeGen/Generic are not
>> successfully compiled by my back-end.
>> Furthermore, I am completely puzzled by the error messages generated.
>> The testcase 'test/CodeGen/Generic/pr12507.ll' for instance yields:
>>
>> LLVM ERROR: Cannot select: 0x229dbd0: i64,ch = load 0x226ee70,
>> 0x229d8d0, 0x229d9d0<LD4[FixedStack-1](align=8), zext from i32>
>> [ID=14]
>>   0x229d8d0: i64 = FrameIndex<-1> [ORD=1] [ID=3]
>>   0x229d9d0: i64 = undef [ORD=1] [ID=4]
> [Villmow, Micah] This is a zero extending load and not just a normal load. Do you handle the pattern fragment zextload in your tablegen?
> You can try to do setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, Expand); in your ISelLowering constructor if you don't support this natively.
>

This really brought me forward as I just missed setLoadExtAction & Co
up to now, I just was aware of setOperationAction.

>> What exactly does this mean? OK, it means that some load instruction
>> could not be selected. But why? Is the instruction not supported as it
>> is (I doubt this as there are such instructions defined according with
>> appropriate patterns)? Is the addressing mode node supported (how
>> could I derive this information)? Is there a way how I can extract
>> that information from these error messages in an "easy" manner?
>>
>> My other problem is the zext instruction from i32 to i64. In
>> principle, this is very easy on the TriCore. Zext the argument in the
>> lower part of the 64bit register pair and write 0 to higher part.
>> Finally, these parts are glued together using ISD::BUILD_PAIR.
>> However, when I try to insert such nodes into the DAG within
>> TriCoreTargetLowering, I run into an assertion:
>>
>> llc:
>> /home/scheler/git/llvm/lib/CodeGen/SelectionDAG/InstrEmitter.cpp:704:
>> void llvm::InstrEmitter::EmitMachineNode(llvm::SDNode*, bool, bool,
>> llvm::DenseMap<llvm::SDValue, unsigned int>&): Assertion
>> `NumMIOperands >= II.getNumOperands() && NumMIOperands <=
>> II.getNumOperands()+II.getNumImplicitDefs() && "#operands for dag node
>> doesn't match .td file!"' failed.
>>
>> Currently, I am not even able to find out which instruction is messed
>> up here  (dumping the node via the dump-Method yields "<<Unknown
>> Machine Node #65434>>"). Can I use "machine nodes" and "normal nodes"
>> when lowering a specific instruction within the TargetLowering?
>
> If you do dump(CurDAG) then it will resolve machine nodes for you.
> See llvm/include /llvm/CodeGen/SelectionDAGNodes.h
>

By using dump(CurDAG) I was finally able to find out the problematic
instruction. It was defined in TriCoreInstrInfo.td but did not have an
associated pattern (the pattern is empty '[]'). Furthermore, this
instruction was not used anywhere else in the TriCore back-end, so I
commented it out. This resulted in the same assertion being triggered
and again no pattern was associated with the problematic instruction.
After commenting out enough of these instructions the assertion was
not triggered any more, this is the current state.

Is '[]' supposed to be a catch-all-pattern or a pattern that never
matches? Is this a problem that results from the back-end
implementation or is it a general problem (there are more back-ends
having '[]'-patterns).

> [Villmow, Micah] Have you tried using a tablegen pattern here? I find it is easier for simple conversions like this than using C++ code.
> For example our backend does this with:
> def uitoli64rr:Pat < (i64 (zext GPRI32:$src)), (LCREATEi64rr GPRI32:$src, (LOADCONSTi32 0)) >;
> Where LCREATE is a machine instruction that does similar to ISD::BUILD_PAIR from two i32's and outputs a i64.
>

I already tried to use a pattern, but for some reason I messed it up.
Giving it another try now, I finally succeeded :-) And yes, it is much
easier than doing it in C++.

Ciao, Fabian