[llvm] [LoongArch] Ensure pcaddu18i and jirl adjacency in tail calls for correct relocation (PR #113932)

[via llvm-commits]

================
@@ -763,6 +744,54 @@ bool LoongArchExpandPseudo::expandCopyCFR(
   return true;
 }
 
+bool LoongArchExpandPseudo::expandFunctionCALL(
+    MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
+    MachineBasicBlock::iterator &NextMBBI, bool IsTailCall) {
+  MachineFunction *MF = MBB.getParent();
+  MachineInstr &MI = *MBBI;
+  DebugLoc DL = MI.getDebugLoc();
+  const MachineOperand &Func = MI.getOperand(0);
+  MachineInstrBuilder CALL;
+  unsigned Opcode;
+
+  switch (MF->getTarget().getCodeModel()) {
+  default:
+    report_fatal_error("Unsupported code model");
+    break;
+  case CodeModel::Medium: {
+    // CALL:
+    // pcaddu18i $ra, %call36(func)
+    // jirl      $ra, $ra, 0
+    // TAIL:
+    // pcaddu18i $t8, %call36(func)
+    // jirl      $r0, $t8, 0
+    Opcode =
+        IsTailCall ? LoongArch::PseudoJIRL_TAIL : LoongArch::PseudoJIRL_CALL;
+    Register ScratchReg = IsTailCall ? LoongArch::R20 : LoongArch::R1;
----------------
heiher wrote:

In the case of a tail call, the lifetime of the callee-saved, non-argument registers has expired, allowing any such register to be used as scratch. Using a statically designated scratch register here is simpler and more effective than a dynamic allocation mechanism.

https://github.com/llvm/llvm-project/pull/113932