[llvm] [InstCombine] Add fold combining adjacent extracted bit fields to restore optimization for separate truncations (PR #219598)

[Ryunosuke Yanashita via llvm-commits]

================
@@ -4319,6 +4319,29 @@ Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) {
           return BinaryOperator::CreateAnd(Or, C01);
         }
       }
+
+      // (trunc (lshr X, S) & C0) | (lshr (trunc X), S & C1)
+      // --> (trunc (lshr X, S) & (C0 | C1) (and similar cases)
+      // A = trunc (lshr X, S) B = lshr (trunc X), S
+      const APInt *ShiftAmt;
+      if (match(A, m_Trunc(m_LShr(m_Value(X), m_APInt(ShiftAmt)))) &&
+          match(B, m_LShr(m_Trunc(m_Specific(X)), m_SpecificInt(*ShiftAmt))) &&
+          ShiftAmt->ult(A->getType()->getScalarSizeInBits()) &&
+          !C1->intersects(APInt::getHighBitsSet(
+              A->getType()->getScalarSizeInBits(), ShiftAmt->getZExtValue()))) {
+        return BinaryOperator::CreateAnd(
+            A, ConstantInt::getIntegerValue(I.getType(), *C0 | *C1));
----------------
ryanashita wrote:

I initially used ```ConstantInt::get``` but the lack of support for vectors caused a crash in Archer, so to support results that are vector types I switched to using ```ConstantInt::getIntegerValue``` .  I need the more general ```getIntegerValue``` to support vectors. Should this transform support vectors, or not? Some clarification would be appreciated

https://github.com/llvm/llvm-project/pull/219598