[MC] Don't treat altentry symbols as atoms

The current `setAtom` is inaccurate: a `.alt_entry` label can also be
recognized as an atom. This is mostly benign, but might cause two
locations only separated by an `.alt_entry` to have different atoms.

https://reviews.llvm.org/D153167 changed a `evaluateKnownAbsolute` to
`evaluateAsAbsolute` and would not fold `A-B` even if they are only
separated by a `.alt_entry` label, leading to a spurious error
`invalid CFI advance_loc expression`.

The fix is similar to #82268: add a special case for `.alt_entry`.

Fix #97116

Pull Request: https://github.com/llvm/llvm-project/pull/97479

1 files changed, 1 insertions, 1 deletions

diff --git a/llvm/lib/MC/MCMachOStreamer.cpp b/llvm/lib/MC/MCMachOStreamer.cpp
index d67b958..0b34d87 100644
--- a/llvm/lib/MC/MCMachOStreamer.cpp
+++ b/llvm/lib/MC/MCMachOStreamer.cpp
@@ -508,7 +508,7 @@ void MCMachOStreamer::finishImpl() {
   DenseMap<const MCFragment *, const MCSymbol *> DefiningSymbolMap;
   for (const MCSymbol &Symbol : getAssembler().symbols()) {
     if (getAssembler().isSymbolLinkerVisible(Symbol) && Symbol.isInSection() &&
-        !Symbol.isVariable()) {
+        !Symbol.isVariable() && !cast<MCSymbolMachO>(Symbol).isAltEntry()) {
       // An atom defining symbol should never be internal to a fragment.
       assert(Symbol.getOffset() == 0 &&
              "Invalid offset in atom defining symbol!");