[llvm] [LoongArch][MC] Handle more PseudoLA* instructions with la-global-with-abs feature (PR #112858)

[via llvm-commits]

================
@@ -1023,14 +1021,42 @@ void LoongArchAsmParser::emitLoadAddressPcrelLarge(MCInst &Inst, SMLoc IDLoc,
 void LoongArchAsmParser::emitLoadAddressGot(MCInst &Inst, SMLoc IDLoc,
                                             MCStreamer &Out) {
   // la.got $rd, sym
-  // expands to:
-  //   pcalau12i $rd, %got_pc_hi20(sym)
-  //   ld.w/d    $rd, $rd, %got_pc_lo12(sym)
   MCRegister DestReg = Inst.getOperand(0).getReg();
   const MCExpr *Symbol = Inst.getOperand(1).getExpr();
   InstSeq Insts;
   unsigned LD = is64Bit() ? LoongArch::LD_D : LoongArch::LD_W;
 
+  if (getSTI().hasFeature(LoongArch::LaGlobalWithAbs)) {
+    // with feature: +la-glabal-with-abs
+    // for 32bit:
+    //   lu12i.w $rd, %got_hi20(sym)
+    //   ori     $rd, $rd, %got_lo12(sym)
+    //   ld.w    $rd, $rd, 0
+    //
+    // for 64bit:
+    //   lu12i.w $rd, %got_hi20(sym)
+    //   ori     $rd, $rd, %got_lo12(sym)
+    //   lu32i.d $rd, %got64_lo20(sym)
+    //   lu52i.d $rd, $rd, %got64_hi12(sym)
+    //   ld.d    $rd, $rd, 0
----------------
wangleiat wrote:

This feature is only used for `macro instruction` expansion and is not related to the code model. It will only affect user-written assembly and will not impact the `LoongArch[PreRA]ExpandPseudo` pass in `codegen`.

https://github.com/llvm/llvm-project/pull/112858