[llvm] [LoopFusion] Reform LCSSA after peelFusionCandidate's peelLoop (PR #200442)

[Ehsan Amiri via llvm-commits]

================
@@ -665,7 +666,13 @@ struct LoopFuser {
                       << " iterations of the first loop. \n");
 
     ValueToValueMapTy VMap;
-    peelLoop(FC0.L, PeelCount, false, &LI, &SE, DT, &AC, true, VMap);
+    // peelLoop can invalidate LCSSA on the surrounding nest; reform it
+    // explicitly below instead of asking peelLoop to preserve it.
+    peelLoop(FC0.L, PeelCount, /*PeelLast=*/false, &LI, &SE, DT, &AC,
+             /*PreserveLCSSA=*/false, VMap);
+    Loop *AffectedNest =
+        FC0.L->getParentLoop() ? FC0.L->getParentLoop() : FC0.L;
+    formLCSSARecursively(*AffectedNest, DT, &LI, &SE);
----------------
amehsan wrote:

The patch is likely fine, I just have a question, quite likely I am missing something here:

Why do we need to call `formLCSSARecursively`?  This is what I know:


1. There is a call to simplifyLoop very early in loop fusion (in `LoopFusePass::run`) that I copy below. It seems that loop fusion doesn't assume incoming loops are in LCSSA form
```
  for (auto &L : LI) {
    Changed |=
        simplifyLoop(L, &DT, &LI, &SE, &AC, nullptr, false /* PreserveLCSSA */);
  }
```

2. I believe loop pass manager requires loop passes to preserve  LCSSA. but loop fusion is not a loop pass, it is a function pass. So it has no commitment to preserve LCSSA and it doesn't expect the loops to be in LCSSA a form either.

So why do we need to call `formLCSSARecursively` and flipping the flag in line 672 is not enough?

https://github.com/llvm/llvm-project/pull/200442