[PATCH] D48467: [X86] Recognize a splat of negate in isFNEG

[Simon Pilgrim via Phabricator via llvm-commits]

RKSimon added inline comments.


================
Comment at: lib/Target/X86/X86ISelLowering.cpp:36852
+      if (auto *C = cast_or_null<ConstantFP>(
+              getTargetConstantFromNode(Op1.getOperand(0))))
+        if (IsNeg(C))
----------------
eraman wrote:
> RKSimon wrote:
> > Are there any circumstances that this isn't a ConstantFP? getTargetConstantFromNode peeks through bitcasts so don't you need to use dyn_cast_or_null?
> I have changed it to dyn_cast_or_null, but thinking about it I don't think that is needed.  First, the current code reads 
> 
> if (Op1.getOpcode() == X86ISD::VBROADCAST) {
>     if (auto *C = getTargetConstantFromNode(Op1.getOperand(0)))
>       if (isSignMask(cast<ConstantFP>(C)))
> 
> The x86 vbroadcast instruction broadcasts floating point values, so I think the cast<ConstantFP> is right. 
If you look at getTargetConstantFromNode the first thing it does is call peekThroughBitcasts so it can have any type. In fact we should probably be checking that the vector's element size is correct as well.

This is the kind of thing a fuzz test finds in 3 months time......


Repository:
  rL LLVM

https://reviews.llvm.org/D48467