[llvm] [RISCV] Teach getIntImmCostInst about X & -(1 << C1) & 0xffffffff) == C2 << C1 (PR #160163)

[Craig Topper via llvm-commits]

================
@@ -166,6 +166,42 @@ static bool canUseShiftPair(Instruction *Inst, const APInt &Imm) {
   return false;
 }
 
+// If this is a 64-bit AND with a mask of the form -(1 << C) in the lower 32
+// bits and the only user is an equality comparison, we might be able to use a
+// sraiw instead. This avoids the need to materialize the AND constant.
+static bool canUseSRAIWCmp(Instruction *Inst, const APInt &Imm) {
+  if (!Inst->hasOneUse())
+    return false;
+
+  // Look for equality comparison.
+  auto *Cmp = dyn_cast<ICmpInst>(*Inst->user_begin());
+  if (!Cmp || !Cmp->isEquality())
+    return false;
+
+  // Right hand side of comparison should be a constant.
+  auto *C = dyn_cast<ConstantInt>(Cmp->getOperand(1));
+  if (!C)
+    return false;
+
+  uint64_t Mask = Imm.getZExtValue();
+
+  // Mask should be of the form -(1 << C) in the lower 32 bits.
+  if (!isShiftedMask_64(Mask) || !isUInt<32>(Mask) ||
----------------
topperc wrote:

Turns out I didn't use isShiftedMask in the DAGCombine. So no change needed.

https://github.com/llvm/llvm-project/pull/160163