[clang] [llvm] [RISCV][LLVM] Enable atomics for 'Zalrsc' (PR #163672)

[Craig Topper via cfe-commits]

================
@@ -24483,6 +24483,26 @@ ISD::NodeType RISCVTargetLowering::getExtendForAtomicCmpSwapArg() const {
   return Subtarget.hasStdExtZacas() ? ISD::ANY_EXTEND : ISD::SIGN_EXTEND;
 }
 
+ISD::NodeType RISCVTargetLowering::getExtendForAtomicRMWArg(unsigned Op) const {
+  // Zaamo will use amo<op>.w which does not require extension.
+  if (Subtarget.hasStdExtZaamo() || Subtarget.hasForcedAtomics())
+    return ISD::ANY_EXTEND;
+
+  // Zalasr pseudo expansions with comparison
+  assert(Subtarget.hasStdExtZalrsc());
+  switch (Op) {
+  case ISD::ATOMIC_LOAD_MIN:
+  case ISD::ATOMIC_LOAD_MAX:
+    return ISD::SIGN_EXTEND;
+  case ISD::ATOMIC_LOAD_UMIN:
+  case ISD::ATOMIC_LOAD_UMAX:
+    return ISD::ZERO_EXTEND;
----------------
topperc wrote:

It would be ok to sign extend umin/umax as long as we also sign extend the value read from memory in ExpandAtomicPseudos. Which we would if we used the LR result directly.

https://github.com/llvm/llvm-project/pull/163672