[PATCH] D70582: [FPEnv][X86] Constrained FCmp intrinsics enabling on X86

[Craig Topper via Phabricator via llvm-commits]

craig.topper added inline comments.


================
Comment at: llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp:3577
     EVT VT = Node->getValueType(0);
+    SDValue Chain;
     SDValue CC = Node->getOperand(4);
----------------
pengfei wrote:
> uweigand wrote:
> > I guess this needs an input chain, right?   I think you'll probably have to add a whole new STRICT_FSELECT_CC node to get this right ...
> Since we are supporting both signaling and quite STRICT intrinsics, I think there's no need to expand the STRICT version of `SETCC`, `SELECT_CC` and `BR_CC` now.
> That's because `setCondCodeAction` doesn't distinguish signaling and quite. Obviously, some CC that is legal in signaling or quite might be not in another.
> The only way is customizing them, otherwise falling back to non STRICT node.
Is it never legal to fall back to a non strict node. Once all targets support strict fp all mutation code will be removed from the tree. We might need to enhance setCondCodeAction to understand signaling.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D70582/new/

https://reviews.llvm.org/D70582