[llvm] [DAG] Fold trunc(srl(extract_elt(vec,c1),c2)) -> extract_elt(bitcast(vec),c3) (PR #107987)

[Simon Pilgrim via llvm-commits]

================
@@ -15142,26 +15142,42 @@ SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
   // Note: We only run this optimization after type legalization (which often
   // creates this pattern) and before operation legalization after which
   // we need to be more careful about the vector instructions that we generate.
-  if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
-      LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
-    EVT VecTy = N0.getOperand(0).getValueType();
-    EVT ExTy = N0.getValueType();
+  if (LegalTypes && !LegalOperations && VT.isScalarInteger() && VT != MVT::i1 &&
+      N0->hasOneUse()) {
     EVT TrTy = N->getValueType(0);
+    SDValue Src = N0;
+
+    // Check for cases where we shift down an upper element before truncation.
+    int EltOffset = 0;
+    if (Src.getOpcode() == ISD::SRL && Src.getOperand(0)->hasOneUse()) {
+      if (auto ShAmt = DAG.getValidShiftAmount(Src)) {
+        if ((*ShAmt % TrTy.getSizeInBits()) == 0) {
----------------
RKSimon wrote:

ShAmt is optional<uint64_t> not TypeSize so I don't think that would work

https://github.com/llvm/llvm-project/pull/107987