[PATCH] D113351: [DAG] Add BuildVectorSDNode::getConstantRawBits helper

[Simon Pilgrim via Phabricator via llvm-commits]

RKSimon planned changes to this revision.
RKSimon added inline comments.


================
Comment at: llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp:10932-10933
+
+  if (!isConstant())
+    return false;
+
----------------
pengfei wrote:
> Move ahead?
Yes I can do that - the BuildVectorSDNode methods are inconsistent as to whether the output variables are canonicalized or not on failure.


================
Comment at: llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp:10942
+      for (unsigned J = 0; J != Scale; ++J) {
+        SDValue Op = getOperand((I * Scale) + (IsLE ? J : (Scale - J - 1)));
+        if (Op.isUndef())
----------------
pengfei wrote:
> Do we need to check `IsLE` for `I`?
Not AFAICT - we don't do that for other vectors-through-bitcasts (the DAG code this was pulled out of, shuffles-of-bitcasts, etc.) - its purely the elements that alias within the bitcast that need swapping..


================
Comment at: llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp:10951
+            CInt ? CInt->getAPIntValue().zextOrTrunc(SrcEltSizeInBits)
+                 : CFP->getValueAPF().bitcastToAPInt(),
+            J * SrcEltSizeInBits);
----------------
pengfei wrote:
> Do we need to `zextOrTrunc` it?
For integer - a truncOrSelf would suffice. For float the bitwidths should match. I'll add an assert to be sure.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D113351/new/

https://reviews.llvm.org/D113351