[llvm] [SjLjEHPrepare] Fix callsite problem (PR #67264)

[Kazushi Marukawa via llvm-commits]

================
@@ -442,9 +442,19 @@ bool SjLjEHPrepare::setupEntryBlockAndCallSites(Function &F) {
   for (BasicBlock &BB : F) {
     if (&BB == &F.front())
       continue;
-    for (Instruction &I : BB)
-      if (I.mayThrow())
-        insertCallSiteStore(&I, -1);
+    for (Instruction &I : BB) {
+      // Partially revert b61fd7f modifications.  Stop using "I.mayThrow()"
----------------
kaz7 wrote:

Ok.  It's little tough for me.  I still don't understand the detail of mayThrow.  I'll try to explain the problem when I have time.  Thank you.

The situation is something like this.  Existing code inserts no-action mark (callsite == -1) just before "invoke __cxa_bad_cast" instructions because mayThrow function return true for such invoking instructions.  As a result, following instructions sequence is generated.
```
                                          # call _Unwind_SjLj_Register
        lea %s0, _Unwind_SjLj_Register at lo
        and %s0, %s0, (32)0
        lea.sl %s12, _Unwind_SjLj_Register at hi(, %s0)
        lea %s0, -104(, %s9)
        bsic %s10, (, %s12)
...
.LBB0_1:                                # %dynamic_cast.bad_cast
        or %s0, 1, (0)1              # store 1 by "insertCallSiteStore(Invokes[I], I + 1);" in previous loop.
        st %s0, -96(, %s9)
        or %s0, -1, (0)1             # store -1 by "insertCallSiteStore(&I, -1);" by loop I try to modify.
        st %s0, -96(, %s9)
.Ltmp0:
        lea %s0, __cxa_bad_cast at lo     # callsite is -1, so nobody catch this bad cast
        and %s0, %s0, (32)0
        lea.sl %s12, __cxa_bad_cast at hi(, %s0)
        bsic %s10, (, %s12)
```

I just notice that this problem was starting after b61fd7fc6da57c261a3d2154eecff7c59dfda9b0.  It describes "No functional change is intended.", so I just thought reverting it partially is good enough.  But not.  I'll try to explain the problem later like I said above. 

https://github.com/llvm/llvm-project/pull/67264