[llvm] [win][x64] Windows x64 unwind v3: Update epilog inheritance per spec clarification (PR #202778)

[Evgenii Kudriashov via llvm-commits]

================
@@ -579,16 +579,22 @@ static void EmitUnwindInfoV3(MCStreamer &Streamer, WinEH::FrameInfo *Info) {
                           Twine(EpilogInfos.size()));
 
   // --- Inheritance decisions ---
-  // An epilog can use the inherited (3-byte) descriptor when FirstOp,
-  // NumberOfOps, NeedsLarge, and relative IP offsets all match the previous
-  // epilog.
-  for (unsigned I = 1; I < EpilogInfos.size(); ++I) {
-    auto &Prev = EpilogInfos[I - 1];
+  // Per the V3 spec, an epilog descriptor with NumberOfOps == 0 inherits its
+  // effective NumberOfOps, FirstOp, IpOffsetOfLastInstruction, and IP offset
+  // array from the first preceding descriptor with NumberOfOps != 0 (the
+  // "base"), not necessarily the immediately preceding descriptor. An epilog
+  // can therefore use the inherited (3-byte) descriptor when FirstOp,
+  // NumberOfOps, NeedsLarge, and relative IP offsets all match that base.
+  for (unsigned I = 1, Base = 0; I < EpilogInfos.size(); ++I) {
+    auto &BaseEI = EpilogInfos[Base];
     auto &Curr = EpilogInfos[I];
-    if (Curr.FirstOp == Prev.FirstOp && Curr.NumberOfOps == Prev.NumberOfOps &&
-        Curr.NeedsLarge == Prev.NeedsLarge &&
-        EpilogIpOffsetsMatch(*Curr.Epilog, *Prev.Epilog, OS->getAssembler()))
+    if (Curr.FirstOp == BaseEI.FirstOp &&
+        Curr.NumberOfOps == BaseEI.NumberOfOps &&
+        Curr.NeedsLarge == BaseEI.NeedsLarge &&
+        EpilogIpOffsetsMatch(*Curr.Epilog, *BaseEI.Epilog, OS->getAssembler()))
----------------
e-kud wrote:

I wonder how it works. So during encoding we want the Large flag to be matching when during decoding we take it independently. It looks like we have a limitation in case of Large/Non-large inheritance during encoding but I see that `EPILOG_INFO_V3` is the same for large and non-large epilogs. Which makes me think that Large equality is redundant here.

Strictly speaking it must be matching otherwise we have a bug.

https://github.com/llvm/llvm-project/pull/202778