[llvm] [PreISelIntrinsicLowering] Expand binary elementwise intrinsics (#193552) (PR #193580)

[David Sherwood via llvm-commits]

================
@@ -810,6 +810,22 @@ bool PreISelIntrinsicLowering::lowerIntrinsics(Module &M) const {
         return lowerUnaryVectorIntrinsicAsLoop(M, CI);
       });
       break;
+    case Intrinsic::atan2:
+    case Intrinsic::ldexp:
+    case Intrinsic::pow:
+    case Intrinsic::powi:
+      Changed |= forEachCall(F, [&](CallInst *CI) {
+        Type *Ty = CI->getArgOperand(0)->getType();
+        if (!TM || !isa<ScalableVectorType>(Ty))
+          return false;
+        const TargetLowering *TL = TM->getSubtargetImpl(F)->getTargetLowering();
+        unsigned Op = TL->IntrinsicIDToISD(F.getIntrinsicID());
+        assert(Op != ISD::DELETED_NODE && "unsupported intrinsic");
+        if (!TL->isOperationExpand(Op, EVT::getEVT(Ty)))
----------------
david-arm wrote:

I'm not sure this is right. When SVE is available in the AArch64 backend we mark ISD::FPOW as EXPAND, however this doesn't mean the code will really be unrolled when a vector math library is available. See this code in `VectorLegalizer::Expand`:

```
  case ISD::FPOW: {
    RTLIB::Libcall LC = RTLIB::getPOW(Node->getValueType(0));
    if (tryExpandVecMathCall(Node, LC, Results))
      return;

    // TODO: Try to see if there's a narrower call available to use before
    // scalarizing.
    break;
  }
```

I need to confirm, but I suspect this may lead to regressions.

https://github.com/llvm/llvm-project/pull/193580