[LLD][AArch64] Detach Landing Pad creation from Thunk creation (#116402)

Move Landing Pad Creation to a new function that checks each thunk every
pass to see if it needs a landing pad. This permits a thunk to be
created without needing a landing pad, but later needing one due to
drifting out of direct branch range and requiring an indirect branch.

We record all the Thunks created so far in a new vector rather than
trying to iterate over the DenseMap as we need a deterministic order of
adding LandingPadThunks due to the short branch fall through. We cannot
use normalizeExistingThunk() either as that only iterates through live
thunks.

Fixes: https://crbug.com/377438309
Original PR: https://github.com/llvm/llvm-project/pull/108989

Sending without a new test case to fix existing test. A new regression
test will come in a separate PR as coming up with a small enough
reproducer for this case is non-trivial.

lld/ELF/Relocations.cpp

@@ -2293,6 +2293,30 @@ bool ThunkCreator::normalizeExistingThunk(Relocation &rel, uint64_t src) {
   return false;
 }
 
+// When indirect branches are restricted, such as AArch64 BTI Thunks may need
+// to target a linker generated landing pad instead of the target. This needs
+// to be done once per pass as the need for a BTI thunk is dependent whether
+// a thunk is short or long. We iterate over all the thunks to make sure we
+// catch thunks that have been created but are no longer live. Non-live thunks
+// are not reachable via normalizeExistingThunk() but are still written.
+bool ThunkCreator::addSyntheticLandingPads() {
+  bool addressesChanged = false;
+  for (Thunk *t : allThunks) {
+    if (!t->needsSyntheticLandingPad())
+      continue;
+    Thunk *lpt;
+    bool isNew;
+    auto &dr = cast<Defined>(t->destination);
+    std::tie(lpt, isNew) = getSyntheticLandingPad(dr, t->addend);
+    if (isNew) {
+      addressesChanged = true;
+      getISThunkSec(cast<InputSection>(dr.section))->addThunk(lpt);
+    }
+    t->landingPad = lpt->getThunkTargetSym();
+  }
+  return addressesChanged;
+}
+
 // Process all relocations from the InputSections that have been assigned
 // to InputSectionDescriptions and redirect through Thunks if needed. The
 // function should be called iteratively until it returns false.
@@ -2326,6 +2350,9 @@ bool ThunkCreator::createThunks(uint32_t pass,
   if (pass == 0 && ctx.target->getThunkSectionSpacing())
     createInitialThunkSections(outputSections);
 
+  if (ctx.arg.emachine == EM_AARCH64)
+    addressesChanged = addSyntheticLandingPads();
+
   // Create all the Thunks and insert them into synthetic ThunkSections. The
   // ThunkSections are later inserted back into InputSectionDescriptions.
   // We separate the creation of ThunkSections from the insertion of the
@@ -2360,20 +2387,7 @@ bool ThunkCreator::createThunks(uint32_t pass,
                 ts = getISDThunkSec(os, isec, isd, rel, src);
               ts->addThunk(t);
               thunks[t->getThunkTargetSym()] = t;
-
-              // When indirect branches are restricted, such as AArch64 BTI
-              // Thunks may need to target a linker generated landing pad
-              // instead of the target.
-              if (t->needsSyntheticLandingPad()) {
-                Thunk *lpt;
-                auto &dr = cast<Defined>(t->destination);
-                std::tie(lpt, isNew) = getSyntheticLandingPad(dr, t->addend);
-                if (isNew) {
-                  ts = getISThunkSec(cast<InputSection>(dr.section));
-                  ts->addThunk(lpt);
-                }
-                t->landingPad = lpt->getThunkTargetSym();
-              }
+              allThunks.push_back(t);
             }
 
             // Redirect relocation to Thunk, we never go via the PLT to a Thunk

lld/ELF/Relocations.h

@@ -183,6 +183,8 @@ private:
 
   bool normalizeExistingThunk(Relocation &rel, uint64_t src);
 
+  bool addSyntheticLandingPads();
+
   Ctx &ctx;
 
   // Record all the available Thunks for a (Symbol, addend) pair, where Symbol
@@ -216,6 +218,9 @@ private:
                  Thunk *>
       landingPadsBySectionAndAddend;
 
+  // All the nonLandingPad thunks that have been created, in order of creation.
+  std::vector<Thunk *> allThunks;
+
   // The number of completed passes of createThunks this permits us
   // to do one time initialization on Pass 0 and put a limit on the
   // number of times it can be called to prevent infinite loops.