[llvm] [AMDGPU][SDAG] Legalise v2i32 or/xor/and instructions to make use of 64-bit wide instructions (PR #140694)

[Chris Jackson via llvm-commits]

================
@@ -3225,29 +3225,51 @@ bool AMDGPUDAGToDAGISel::SelectVOP3ModsImpl(SDValue In, SDValue &Src,
   if (IsCanonicalizing)
     return true;
 
-  unsigned Opc = Src->getOpcode();
+  // v2i32 xor/or/and are legal. A vselect using these instructions as operands
+  // is scalarised into two selects with EXTRACT_VECTOR_ELT operands. Peek
+  // through the extract to the bitwise op.
+  SDValue PeekSrc =
+      Src->getOpcode() == ISD::EXTRACT_VECTOR_ELT ? Src->getOperand(0) : Src;
+  // Convert various sign-bit masks to src mods. Currently disabled for 16-bit
+  // types as the codegen replaces the operand without adding a srcmod.
+  // This is intentionally finding the cases where we are performing float neg
+  // and abs on int types, the goal is not to obtain two's complement neg or
+  // abs.
+  // TODO: Add 16-bit support.
+  unsigned Opc = PeekSrc.getOpcode();
   EVT VT = Src.getValueType();
   if ((Opc != ISD::AND && Opc != ISD::OR && Opc != ISD::XOR) ||
-      (VT != MVT::i32 && VT != MVT::i64))
+      (VT != MVT::i32 && VT != MVT::v2i32 && VT != MVT::i64))
     return true;
 
-  ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Src->getOperand(1));
+  ConstantSDNode *CRHS = isConstOrConstSplat(PeekSrc ? PeekSrc->getOperand(1)
+                                                     : Src->getOperand(1));
----------------
chrisjbris wrote:

I didn't test for splat as the vselect is lowered to two v_cndmask_32 instructions, so at this point we don't need to know if it is a const splat. As @rampitec pointed out in https://github.com/llvm/llvm-project/pull/152119, the check for splat would become very important if we were able to emit a "v_cndmask_64". But at the moment we can't.

https://github.com/llvm/llvm-project/pull/140694