[llvm] [LLVM][CodeGen] Teach SelectionDAG how to expand FREM to a vector math call. (PR #83859)

[Paul Walker via llvm-commits]

================
@@ -1842,6 +1859,116 @@ void VectorLegalizer::ExpandREM(SDNode *Node,
   Results.push_back(Result);
 }
 
+// Try to expand libm nodes into a call to a vector math. Callers provide the
+// LibFunc equivalent of the passed in Node, which is used to lookup mappings
+// within TargetLibraryInfo. Only simply mappings are considered whereby only
+// matching vector operands are allowed and masked functions are passed an all
+// true vector (i.e. Node cannot be a predicated operation).
+bool VectorLegalizer::tryExpandVecMathCall(SDNode *Node, RTLIB::Libcall LC,
+                                           SmallVectorImpl<SDValue> &Results) {
+  // Chain must be propagated but currently strict fp operations are down
+  // converted to their none strict counterpart.
+  assert(!Node->isStrictFPOpcode() && "Unexpected strict fp operation!");
+
+  const char *LCName = TLI.getLibcallName(LC);
+  if (!LCName)
+    return false;
+  LLVM_DEBUG(dbgs() << "Looking for vector variant of " << LCName << "\n");
+
+  EVT VT = Node->getValueType(0);
+  ElementCount VL = VT.getVectorElementCount();
+
+  // Lookup a vector function equivalent to the specified libcall. Prefer
+  // unmasked variants but we will generate a mask if need be.
+  const TargetLibraryInfo &TLibInfo = DAG.getLibInfo();
+  const VecDesc *VD = TLibInfo.getVectorMappingInfo(LCName, VL, false);
----------------
paulwalker-arm wrote:

Depends on how strongly you feel. My rational for not adding the comment is that it causes the line to wrap and I figured having a matching call just below that does comment this parameter offers enough readability.

https://github.com/llvm/llvm-project/pull/83859