Mark SRet argument as pointer in SelectionDAGISel::LowerArguments

[JamesNagurne]

Summary

Mark SRet argument as pointer in SelectionDAGISel::LowerArguments

When the concept of pointers was added to ArgFlagsTy, lowered SRet
arguments were not marked as pointers.

I don't think this affects any upstream target, but the flag should still
be set for consistency.

[efriedma]

I suspect this affects arm64_32, at least in terms of the generated SelectionDAG.

[JamesNagurne]

In D105310#2854283, @efriedma wrote:

I suspect this affects arm64_32, at least in terms of the generated SelectionDAG.

The upstream targets that use the isPointer field (through CCIfPtr in the calling convention .td files) are AArch64 and X86.

A problem occurs iff a pointer type is handled in a calling convention prior to an explicit SRet being handled. I'm unfortunately not well versed in how those targets' conventions work, and I understand the implications of this change breaking the ABI. Do you know of any experts on the topics for the given targets?

[efriedma]

For AArch64, I added @t.p.northover as reviewer; he wrote the relevant code.

For x32, adding @pengfei ?

I suspect this is doing what we want, I just want to make sure this is isn't surprising to someone more familiar with the relevant targets.

[pengfei]

I think the only thing X86 does for the pointer type is zero extend it to i64. So I think only x32 should be affected by it. Add @hvdijk who is familiar with x32 calling conversion.

[hvdijk]

In D105310#2856765, @pengfei wrote:

I think the only thing X86 does for the pointer type is zero extend it to i64. So I think only x32 should be affected by it. Add @hvdijk who is familiar with x32 calling conversion.

That's right. The same registers are used regardless of whether the type is zero extended, 32-bit integers and 32-bit pointers are returned in the same 64-bit registers, the difference is whether the caller is responsible for clearing the high bits of those registers. According to the ABI the caller is responsible for that for all pointer arguments, no exception is made for the implicit pointer argument for return values that cannot be returned in registers, so I think this change is correct for x32.

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

ComputeValueVTs(*TLI, DAG.getDataLayout(),
                     F.getReturnType()->getPointerTo(
                         DAG.getDataLayout().getAllocaAddrSpace()),
                     ValueVTs);
 
     // NOTE: Assuming that a pointer will never break down to more than one VT
     // or one register.
     ISD::ArgFlagsTy Flags;
     Flags.setSRet();
+    Flags.setPointer();
     MVT RegisterVT = TLI->getRegisterType(*DAG.getContext(), ValueVTs[0]);
     ISD::InputArg RetArg(Flags, RegisterVT, ValueVTs[0], true,
                          ISD::InputArg::NoArgIndex, 0);
     Ins.push_back(RetArg);
   }
 
   // Look for stores of arguments to static allocas. Mark such arguments with a
   // flag to ask the target to give us the memory location of that argument if