[PATCH] D46384: Reapply "[SelectionDAG] Selection of DBG_VALUE using a PHI node result (pt 2)"

[Vedant Kumar via Phabricator via llvm-commits]

vsk added inline comments.


================
Comment at: llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp:5282
+              if (Offset >= BitsToDescribe)
+                break;
+              unsigned FragmentSize = (Offset + RegisterSize > BitsToDescribe)
----------------
bjope wrote:
> vsk wrote:
> > Just curious -- could this be simplified to 'assert(Offset <= BitsToDescribe)'?
> No.
> 
> If we for example have a IR level PHI node for i64, and i64 isn't legal we can get two PHI nodes in MIR. Each of them defining a 32-bit vreg. So the loop will do two iterations (I=0 and I=1), with RegisterSize=32.
> 
> If the dbg.value only describes a fragment of a variable, then for example we could get BitsToDescibe=24
> So for I=0,Offset=0 we will create a DBG_VALUE with a 24-bit fragment.
> And then for I=0,Offset=32 we will bail out due to Offset>=BitsToDescribe. And we do not need to create a DBG_VALUE for the second register since we already described all 24-bits that the dbg.value described.
> 
Got it, thanks for explaining.


Repository:
  rL LLVM

https://reviews.llvm.org/D46384