[clang] [llvm] [LoongArch] Support amcas[_db].{b/h/w/d} instructions. (PR #114189)

[via cfe-commits]

================
@@ -5743,6 +5747,14 @@ LoongArchTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
   }
 
   unsigned Size = AI->getType()->getPrimitiveSizeInBits();
+  if (Subtarget.hasLAMCAS()) {
+    if (AI->getOperation() == AtomicRMWInst::Nand ||
+        (Size < 32 /*&& (AI->getOperation() != AtomicRMWInst::Or &&
----------------
tangaac wrote:

atomicrmw nand & i8 / i16 atomicrmw add / sub / [u]min / [u]max should be converted to amcas.

i8 / i16 atomicrmw and  / or / xor shouldn't be converted to amcas because they can use am(and / or / xor).w to expand. 

But after feature LAMCAS enabled, all i8,i16 atomicrmw operations will be converted to amcas, 
how to fix it?

https://github.com/llvm/llvm-project/pull/114189