[GISel][RISCV]Implement indirect parameter passing (#95429)

Some targets like RISC-V pass scalars wider than 2×XLEN bits by
reference, so those arguments are replaced in the argument list with an
address (See RISC-V ABIs Specification 1.0 section 2.1).

This commit implements this indirect parameter passing in GlobalISel.

---------

Co-authored-by: Gabor Spaits <Gabor.Spaits@hightec-rt.com>

Author: spaits

llvm/lib/CodeGen/GlobalISel/CallLowering.cpp

@@ -36,6 +36,7 @@ void CallLowering::anchor() {}
 static void
 addFlagsUsingAttrFn(ISD::ArgFlagsTy &Flags,
                     const std::function<bool(Attribute::AttrKind)> &AttrFn) {
+  // TODO: There are missing flags. Add them here.
   if (AttrFn(Attribute::SExt))
     Flags.setSExt();
   if (AttrFn(Attribute::ZExt))
@@ -756,6 +757,8 @@ bool CallLowering::handleAssignments(ValueHandler &Handler,
       continue;
     }
 
+    auto AllocaAddressSpace = MF.getDataLayout().getAllocaAddrSpace();
+
     const MVT ValVT = VA.getValVT();
     const MVT LocVT = VA.getLocVT();
 
@@ -764,6 +767,8 @@ bool CallLowering::handleAssignments(ValueHandler &Handler,
     const LLT NewLLT = Handler.isIncomingArgumentHandler() ? LocTy : ValTy;
     const EVT OrigVT = EVT::getEVT(Args[i].Ty);
     const LLT OrigTy = getLLTForType(*Args[i].Ty, DL);
+    const LLT PointerTy = LLT::pointer(
+        AllocaAddressSpace, DL.getPointerSizeInBits(AllocaAddressSpace));
 
     // Expected to be multiple regs for a single incoming arg.
     // There should be Regs.size() ArgLocs per argument.
@@ -778,31 +783,76 @@ bool CallLowering::handleAssignments(ValueHandler &Handler,
       // intermediate values.
       Args[i].Regs.resize(NumParts);
 
-      // For each split register, create and assign a vreg that will store
-      // the incoming component of the larger value. These will later be
-      // merged to form the final vreg.
-      for (unsigned Part = 0; Part < NumParts; ++Part)
-        Args[i].Regs[Part] = MRI.createGenericVirtualRegister(NewLLT);
+      // When we have indirect parameter passing we are receiving a pointer,
+      // that points to the actual value, so we need one "temporary" pointer.
+      if (VA.getLocInfo() == CCValAssign::Indirect) {
+        if (Handler.isIncomingArgumentHandler())
+          Args[i].Regs[0] = MRI.createGenericVirtualRegister(PointerTy);
+      } else {
+        // For each split register, create and assign a vreg that will store
+        // the incoming component of the larger value. These will later be
+        // merged to form the final vreg.
+        for (unsigned Part = 0; Part < NumParts; ++Part)
+          Args[i].Regs[Part] = MRI.createGenericVirtualRegister(NewLLT);
+      }
     }
 
     assert((j + (NumParts - 1)) < ArgLocs.size() &&
            "Too many regs for number of args");
 
     // Coerce into outgoing value types before register assignment.
-    if (!Handler.isIncomingArgumentHandler() && OrigTy != ValTy) {
+    if (!Handler.isIncomingArgumentHandler() && OrigTy != ValTy &&
+        VA.getLocInfo() != CCValAssign::Indirect) {
       assert(Args[i].OrigRegs.size() == 1);
       buildCopyToRegs(MIRBuilder, Args[i].Regs, Args[i].OrigRegs[0], OrigTy,
                       ValTy, extendOpFromFlags(Args[i].Flags[0]));
     }
 
+    bool IndirectParameterPassingHandled = false;
     bool BigEndianPartOrdering = TLI->hasBigEndianPartOrdering(OrigVT, DL);
     for (unsigned Part = 0; Part < NumParts; ++Part) {
+      assert((VA.getLocInfo() != CCValAssign::Indirect || Part == 0) &&
+             "Only the first parameter should be processed when "
+             "handling indirect passing!");
       Register ArgReg = Args[i].Regs[Part];
       // There should be Regs.size() ArgLocs per argument.
       unsigned Idx = BigEndianPartOrdering ? NumParts - 1 - Part : Part;
       CCValAssign &VA = ArgLocs[j + Idx];
       const ISD::ArgFlagsTy Flags = Args[i].Flags[Part];
 
+      // We found an indirect parameter passing, and we have an
+      // OutgoingValueHandler as our handler (so we are at the call site or the
+      // return value). In this case, start the construction of the following
+      // GMIR, that is responsible for the preparation of indirect parameter
+      // passing:
+      //
+      // %1(indirectly passed type) = The value to pass
+      // %3(pointer) = G_FRAME_INDEX %stack.0
+      // G_STORE %1, %3 :: (store (s128), align 8)
+      //
+      // After this GMIR, the remaining part of the loop body will decide how
+      // to get the value to the caller and we break out of the loop.
+      if (VA.getLocInfo() == CCValAssign::Indirect &&
+          !Handler.isIncomingArgumentHandler()) {
+        Align AlignmentForStored = DL.getPrefTypeAlign(Args[i].Ty);
+        MachineFrameInfo &MFI = MF.getFrameInfo();
+        // Get some space on the stack for the value, so later we can pass it
+        // as a reference.
+        int FrameIdx = MFI.CreateStackObject(OrigTy.getScalarSizeInBits(),
+                                             AlignmentForStored, false);
+        Register PointerToStackReg =
+            MIRBuilder.buildFrameIndex(PointerTy, FrameIdx).getReg(0);
+        MachinePointerInfo StackPointerMPO =
+            MachinePointerInfo::getFixedStack(MF, FrameIdx);
+        // Store the value in the previously created stack space.
+        MIRBuilder.buildStore(Args[i].OrigRegs[Part], PointerToStackReg,
+                              StackPointerMPO,
+                              inferAlignFromPtrInfo(MF, StackPointerMPO));
+
+        ArgReg = PointerToStackReg;
+        IndirectParameterPassingHandled = true;
+      }
+
       if (VA.isMemLoc() && !Flags.isByVal()) {
         // Individual pieces may have been spilled to the stack and others
         // passed in registers.
@@ -812,14 +862,21 @@ bool CallLowering::handleAssignments(ValueHandler &Handler,
         LLT MemTy = Handler.getStackValueStoreType(DL, VA, Flags);
 
         MachinePointerInfo MPO;
-        Register StackAddr = Handler.getStackAddress(
-            MemTy.getSizeInBytes(), VA.getLocMemOffset(), MPO, Flags);
-
-        Handler.assignValueToAddress(Args[i], Part, StackAddr, MemTy, MPO, VA);
-        continue;
-      }
-
-      if (VA.isMemLoc() && Flags.isByVal()) {
+        Register StackAddr =
+            Handler.getStackAddress(VA.getLocInfo() == CCValAssign::Indirect
+                                        ? PointerTy.getSizeInBytes()
+                                        : MemTy.getSizeInBytes(),
+                                    VA.getLocMemOffset(), MPO, Flags);
+
+        // Finish the handling of indirect passing from the passers
+        // (OutgoingParameterHandler) side.
+        // This branch is needed, so the pointer to the value is loaded onto the
+        // stack.
+        if (VA.getLocInfo() == CCValAssign::Indirect)
+          Handler.assignValueToAddress(ArgReg, StackAddr, PointerTy, MPO, VA);
+        else
+          Handler.assignValueToAddress(Args[i], Part, StackAddr, MemTy, MPO, VA);
+      } else if (VA.isMemLoc() && Flags.isByVal()) {
         assert(Args[i].Regs.size() == 1 &&
                "didn't expect split byval pointer");
 
@@ -858,30 +915,45 @@ bool CallLowering::handleAssignments(ValueHandler &Handler,
                                      DstMPO, DstAlign, SrcMPO, SrcAlign,
                                      MemSize, VA);
         }
-        continue;
-      }
-
-      assert(!VA.needsCustom() && "custom loc should have been handled already");
-
-      if (i == 0 && !ThisReturnRegs.empty() &&
-          Handler.isIncomingArgumentHandler() &&
-          isTypeIsValidForThisReturn(ValVT)) {
+      } else if (i == 0 && !ThisReturnRegs.empty() &&
+                 Handler.isIncomingArgumentHandler() &&
+                 isTypeIsValidForThisReturn(ValVT)) {
         Handler.assignValueToReg(ArgReg, ThisReturnRegs[Part], VA);
-        continue;
-      }
-
-      if (Handler.isIncomingArgumentHandler())
+      } else if (Handler.isIncomingArgumentHandler()) {
         Handler.assignValueToReg(ArgReg, VA.getLocReg(), VA);
-      else {
+      } else {
         DelayedOutgoingRegAssignments.emplace_back([=, &Handler]() {
           Handler.assignValueToReg(ArgReg, VA.getLocReg(), VA);
         });
       }
+
+      // Finish the handling of indirect parameter passing when receiving
+      // the value (we are in the called function or the caller when receiving
+      // the return value).
+      if (VA.getLocInfo() == CCValAssign::Indirect &&
+          Handler.isIncomingArgumentHandler()) {
+        Align Alignment = DL.getABITypeAlign(Args[i].Ty);
+        MachinePointerInfo MPO = MachinePointerInfo::getUnknownStack(MF);
+
+        // Since we are doing indirect parameter passing, we know that the value
+        // in the temporary register is not the value passed to the function,
+        // but rather a pointer to that value. Let's load that value into the
+        // virtual register where the parameter should go.
+        MIRBuilder.buildLoad(Args[i].OrigRegs[0], Args[i].Regs[0], MPO,
+                             Alignment);
+
+        IndirectParameterPassingHandled = true;
+      }
+
+      if (IndirectParameterPassingHandled)
+        break;
     }
 
     // Now that all pieces have been assigned, re-pack the register typed values
-    // into the original value typed registers.
-    if (Handler.isIncomingArgumentHandler() && OrigVT != LocVT) {
+    // into the original value typed registers. This is only necessary, when
+    // the value was passed in multiple registers, not indirectly.
+    if (Handler.isIncomingArgumentHandler() && OrigVT != LocVT &&
+        !IndirectParameterPassingHandled) {
       // Merge the split registers into the expected larger result vregs of
       // the original call.
       buildCopyFromRegs(MIRBuilder, Args[i].OrigRegs, Args[i].Regs, OrigTy,

llvm/lib/Target/RISCV/GISel/RISCVCallLowering.cpp

@@ -341,10 +341,8 @@ static bool isLegalElementTypeForRVV(Type *EltTy,
 // TODO: Remove IsLowerArgs argument by adding support for vectors in lowerCall.
 static bool isSupportedArgumentType(Type *T, const RISCVSubtarget &Subtarget,
                                     bool IsLowerArgs = false) {
-  // TODO: Integers larger than 2*XLen are passed indirectly which is not
-  // supported yet.
   if (T->isIntegerTy())
-    return T->getIntegerBitWidth() <= Subtarget.getXLen() * 2;
+    return true;
   if (T->isHalfTy() || T->isFloatTy() || T->isDoubleTy())
     return true;
   if (T->isPointerTy())