[llvm] [SelectionDAG] Widen vector math libcalls when no routine is available (PR #218948)

[David Sherwood via llvm-commits]

=?utf-8?q?Mattéo?= Rizza Murgier,=?utf-8?q?Mattéo?= Rizza Murgier,
=?utf-8?q?Mattéo?= Rizza Murgier,=?utf-8?q?Mattéo?= Rizza Murgier,
=?utf-8?q?Mattéo?= Rizza Murgier
Message-ID:
In-Reply-To: <llvm.org/llvm/llvm-project/pull/218948 at github.com>


================
@@ -2147,24 +2142,36 @@ void VectorLegalizer::ExpandREM(SDNode *Node,
 }
 
 // Try to expand libm nodes into vector math routine calls. Callers provide the
-// LibFunc equivalent of the passed in Node, which is used to lookup mappings
-// within TargetLibraryInfo. The only mappings considered are those where the
-// result and all operands are the same vector type. While predicated nodes are
-// not supported, we will emit calls to masked routines by passing in an all
-// true mask.
-bool VectorLegalizer::tryExpandVecMathCall(SDNode *Node, RTLIB::Libcall LC,
-                                           SmallVectorImpl<SDValue> &Results) {
+// RTLIB::get<OP>(EVT) selector of the node's libcall family, which is used to
+// look up mappings within RuntimeLibcallsInfo. The only mappings considered are
+// those where the result and all operands are the same vector type. While
+// predicated nodes are not supported, we will emit calls to masked routines by
+// passing in a mask that is true for the lanes computed by the node.
+bool VectorLegalizer::tryExpandVecMathCall(
+    SDNode *Node, function_ref<RTLIB::Libcall(EVT)> GetLibcall,
+    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!");
 
-  RTLIB::LibcallImpl LCImpl = DAG.getLibcalls().getLibcallImpl(LC);
-  if (LCImpl == RTLIB::Unsupported)
+  EVT VT = Node->getValueType(0);
+  LLVMContext &Ctx = *DAG.getContext();
+  const LibcallLoweringInfo &Libcalls = DAG.getLibcalls();
+
+  // Try to widen the vector type when no libcall is available at that width.
+  EVT CallVT = VT;
+  RTLIB::LibcallImpl LCImpl = Libcalls.getLibcallImpl(GetLibcall(CallVT));
+  if (LCImpl == RTLIB::Unsupported && VT.getVectorElementCount().isScalar())
     return false;
+  while (LCImpl == RTLIB::Unsupported) {
+    CallVT = CallVT.getDoubleNumVectorElementsVT(Ctx);
----------------
david-arm wrote:

What if the vector type is <3 x float>? Also, in general I'm a bit worried that there are cases where we get stuck in an infinite loop. We should at least have tests for non-power-of-2 element counts for fixed-width and scalable.

https://github.com/llvm/llvm-project/pull/218948