[llvm] [X86] Narrow BT/BTC/BTR/BTS compare + RMW patterns on very large integers (PR #165540)

[Simon Pilgrim via llvm-commits]

================
@@ -53479,6 +53479,79 @@ static SDValue combineMaskedStore(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+// Look for a RMW operation that only touches one bit of a larger than legal
+// type and fold it to a BTC/BTR/BTS pattern acting on a single i32 sub value.
+static SDValue narrowBitOpRMW(StoreSDNode *St, const SDLoc &DL,
+                              SelectionDAG &DAG,
+                              const X86Subtarget &Subtarget) {
+  using namespace SDPatternMatch;
+
+  // Check if the previous op in the chain was a matching normal load.
+  auto *Ld = dyn_cast<LoadSDNode>(St->getChain());
+  if (!Ld || !ISD::isNormalLoad(Ld) || !Ld->isSimple() ||
+      Ld->getBasePtr() != St->getBasePtr() ||
+      Ld->getOffset() != St->getOffset())
+    return SDValue();
+
+  SDValue LoadVal(Ld, 0);
+  SDValue StoredVal = St->getValue();
+  EVT VT = StoredVal.getValueType();
+
+  // Only narrow normal stores of larger than legal scalar integers.
+  if (!ISD::isNormalStore(St) || !St->isSimple() || !VT.isScalarInteger() ||
+      VT.getSizeInBits() <= (Subtarget.is64Bit() ? 64 : 32))
+    return SDValue();
+
+  // BTR: X & ~(1 << ShAmt)
+  // BTS: X | (1 << ShAmt)
+  // BTC: X ^ (1 << ShAmt)
+  SDValue ShAmt;
+  if (!StoredVal.hasOneUse() ||
+      !(sd_match(StoredVal, m_And(m_Specific(LoadVal),
+                                  m_Not(m_Shl(m_One(), m_Value(ShAmt))))) ||
+        sd_match(StoredVal,
+                 m_Or(m_Specific(LoadVal), m_Shl(m_One(), m_Value(ShAmt)))) ||
+        sd_match(StoredVal,
+                 m_Xor(m_Specific(LoadVal), m_Shl(m_One(), m_Value(ShAmt))))))
+    return SDValue();
+
+  // Ensure the shift amount is in bounds.
+  KnownBits KnownAmt = DAG.computeKnownBits(ShAmt);
+  if (KnownAmt.getMaxValue().uge(VT.getSizeInBits()))
+    return SDValue();
----------------
RKSimon wrote:

I think we can only do that if we know the shift amount is an out of bounds constant (which I'd expect other folds to handle - although it might create a poison value for the store) - knownbits might not know enough to determine the max value even if its safe.

https://github.com/llvm/llvm-project/pull/165540