[llvm] [RISCV] Implement trampolines for rv64 (PR #96309)

[Roger Ferrer Ibáñez via llvm-commits]

================
@@ -7170,6 +7179,123 @@ SDValue RISCVTargetLowering::emitFlushICache(SelectionDAG &DAG, SDValue InChain,
   return CallResult.second;
 }
 
+SDValue RISCVTargetLowering::lowerINIT_TRAMPOLINE(SDValue Op,
+                                                  SelectionDAG &DAG) const {
+  if (!Subtarget.is64Bit())
+    llvm::report_fatal_error("Trampolines only implemented for RV64");
+
+  SDValue Root = Op.getOperand(0);
+  SDValue Trmp = Op.getOperand(1); // trampoline
+  SDLoc dl(Op);
+
+  const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
+
+  // We store in the trampoline buffer the following instructions and data.
+  // Offset:
+  //      0: auipc   t2, 0
+  //      4: ld      t0, 24(t2)
+  //      8: ld      t2, 16(t2)
+  //     12: jalr    t0
+  //     16: <StaticChainOffset>
+  //     24: <FunctionAddressOffset>
+  //     32:
+
+  // Constants shamelessly taken from GCC.
+  constexpr unsigned Opcode_AUIPC = 0x17;
+  constexpr unsigned Opcode_LD = 0x3003;
+  constexpr unsigned Opcode_JALR = 0x67;
+  constexpr unsigned ShiftField_RD = 7;
+  constexpr unsigned ShiftField_RS1 = 15;
+  constexpr unsigned ShiftField_IMM = 20;
+  constexpr unsigned Reg_X5 = 0x5; // x5/t0 (holds the address to the function)
+  constexpr unsigned Reg_X7 = 0x7; // x7/t2 (holds the static chain)
+
+  constexpr unsigned StaticChainOffset = 16;
+  constexpr unsigned FunctionAddressOffset = 24;
+
+  SDValue OutChains[6];
+  SDValue Addr = Trmp;
+
+  // auipc t2, 0
+  // Loads the current PC into t2.
+  constexpr uint32_t AUIPC_X7_0 =
+      Opcode_AUIPC | (Reg_X7 << ShiftField_RD);
+  OutChains[0] =
+      DAG.getTruncStore(Root, dl, DAG.getConstant(AUIPC_X7_0, dl, MVT::i64),
----------------
rofirrim wrote:

Not sure, given that I'm using a truncating store perhaps it is understood that the upper 32-bits are dropped hence the combiner kicking in.

I can explicitly use two 64-bit stores if that is preferable.

https://github.com/llvm/llvm-project/pull/96309