[RISCV] Improve codegen for i8/i16 'atomicrmw xchg a, {0,-1}'

As noted in <#64090>, it's
more efficient to lower a partword 'atomicrmw xchg a, 0` to and amoand
with appropriate mask. There are a range of possible ways to go about
this - e.g. writing a combine based on the
`llvm.riscv.masked.atomicrmw.xchg` intrinsic, or introducing a new
interface to AtomicExpandPass to allow target-specific atomics
conversions, or trying to lift the conversion into AtomicExpandPass
itself based on querying some target hook. Ultimately I've gone with
what appears to be the simplest approach - just covering this case in
emitMaskedAtomicRMWIntrinsic. I perhaps should have given that hook a
different name way back when it was introduced.

This also handles the `atomicrmw xchg a, -1` case suggested by Craig
during review.

Fixes #64090

Differential Revision: https://reviews.llvm.org/D156801

Author: asb

Diff for: llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -16747,6 +16747,22 @@ getIntrinsicForMaskedAtomicRMWBinOp(unsigned XLen, AtomicRMWInst::BinOp BinOp) {
 Value *RISCVTargetLowering::emitMaskedAtomicRMWIntrinsic(
     IRBuilderBase &Builder, AtomicRMWInst *AI, Value *AlignedAddr, Value *Incr,
     Value *Mask, Value *ShiftAmt, AtomicOrdering Ord) const {
+  // In the case of an atomicrmw xchg with a constant 0/-1 operand, replace
+  // the atomic instruction with an AtomicRMWInst::And/Or with appropriate
+  // mask, as this produces better code than the LR/SC loop emitted by
+  // int_riscv_masked_atomicrmw_xchg.
+  if (AI->getOperation() == AtomicRMWInst::Xchg &&
+      isa<ConstantInt>(AI->getValOperand())) {
+    ConstantInt *CVal = cast<ConstantInt>(AI->getValOperand());
+    if (CVal->isZero())
+      return Builder.CreateAtomicRMW(AtomicRMWInst::And, AlignedAddr,
+                                     Builder.CreateNot(Mask, "Inv_Mask"),
+                                     AI->getAlign(), Ord);
+    if (CVal->isMinusOne())
+      return Builder.CreateAtomicRMW(AtomicRMWInst::Or, AlignedAddr, Mask,
+                                     AI->getAlign(), Ord);
+  }
+
   unsigned XLen = Subtarget.getXLen();
   Value *Ordering =
       Builder.getIntN(XLen, static_cast<uint64_t>(AI->getOrdering()));