[PATCH] D102817: [Windows SEH]: HARDWARE EXCEPTION HANDLING (MSVC -EHa) - Part 2

[Ten Tzen via Phabricator via llvm-commits]

tentzen added inline comments.


================
Comment at: llvm/lib/CodeGen/AsmPrinter/AsmPrinter.cpp:1316
+        //  Div with variable opnds won't be the first instruction in
+        //  an EH region as it must be led by at least a Load
+        {
----------------
efriedma wrote:
> If the issue here that the "end" label for a region can accidentally include the following instruction after the region?  Would it make sense to explicitly insert a nop earlier, in that case?
I don't see the value of emitting it earlier. I think the closer to emission is safer. Or there is always a risk of being changed or reordered by other phases. Besides, there are a couple of similar cases of emitting Nop in that file (like functionPrfix). I think inserting the nop where the EH_Label is emitted is safest and most straight-forward.


================
Comment at: llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp:2916
+      if (CleanupMBB) // a block referenced by EH table
+        CleanupMBB->setHasAddressTaken(); // so dtor-funclet not removed by opts
       break;
----------------
efriedma wrote:
> I assume this is trying to trick optimizations into avoiding optimizations on the cleanup block if the __try involves multiple basic blocks.
> 
> Have you considered trying to fix the control flow graph?  I mean, I understand the difficulty of expressing the control flow at the IR level, given we don't have IR versions of "load" and "store" with an unwind edge, but those concerns don't really apply in the same way in a MachineFunction.  You can mark the cleanup block as a successor of any block that can throw an exception, without explicitly marking up the control flow on each individual instruction.  (This is how exceptions normally work.)
The Cleanup block is turned into a dtor funclet at the end. And the beginning address of that funclet will be recorded in EH table in .xdata section. So it's really address-taken by definition. If you verify this by checking Linker relocation records or the EH table in .asm file.  IMO this flag should be set regardless of the SEH feature.


================
Comment at: llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp:1323
+    MachineBasicBlock *MBB = &*MBBI;
+    const BasicBlock *BB = MBB->getBasicBlock();
+    int State = EHInfo->BlockToStateMap[BB];
----------------
efriedma wrote:
> getBasicBlock() can return null, in general.  Maybe since this is running during isel, that isn't an issue, though.  Maybe worth adding a comment.
ok, will add a comment. thanks.


================
Comment at: llvm/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp:1341
+      MachineInstr *MIe = &*(--MBBe);
+      // insert before (possible multiple) terminators
+      while (MIe->isTerminator())
----------------
efriedma wrote:
> If a load is folded into the terminator, it can fault... but I guess that's unlikely to come up in practice.  The most likely scenario probably involves indirectbr.  Most other indirect jumps probably aren't terminators. I guess a jump table could theoretically fault, but if you cared you could probably just disable those.
Right, it's unlikely in practice. I don't think it's concern here. 
thank you for pointing out.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D102817/new/

https://reviews.llvm.org/D102817