[llvm] [SelectionDAG][RISCV] Teach computeKnownBits to use range metadata for atomic_load. (PR #137119)

[Matt Arsenault via llvm-commits]

================
@@ -4382,6 +4382,38 @@ KnownBits SelectionDAG::computeKnownBits(SDValue Op, const APInt &DemandedElts,
     Known.Zero |= APInt::getBitsSetFrom(BitWidth, VT.getScalarSizeInBits());
     break;
   }
+  case ISD::ATOMIC_LOAD: {
+    // If we are looking at the loaded value.
+    if (Op.getResNo() == 0) {
+      auto *AT = cast<AtomicSDNode>(Op);
+      unsigned ScalarMemorySize = AT->getMemoryVT().getScalarSizeInBits();
+      KnownBits KnownScalarMemory(ScalarMemorySize);
+      if (const MDNode *MD = AT->getRanges())
+        computeKnownBitsFromRangeMetadata(*MD, KnownScalarMemory);
+
+      switch (AT->getExtensionType()) {
+      case ISD::ZEXTLOAD:
+        Known = KnownScalarMemory.zext(BitWidth);
+        break;
+      case ISD::SEXTLOAD:
+        Known = KnownScalarMemory.sext(BitWidth);
+        break;
+      case ISD::EXTLOAD:
+        if (TLI->getExtendForAtomicOps() == ISD::ZERO_EXTEND)
+          Known = KnownScalarMemory.zext(BitWidth);
+        else if (TLI->getExtendForAtomicOps() == ISD::SIGN_EXTEND)
+          Known = KnownScalarMemory.sext(BitWidth);
----------------
arsenm wrote:

Avoid repeated getExtendForAtomicOps, switch over it? We should get rid of it eventually 

https://github.com/llvm/llvm-project/pull/137119