[llvm-branch-commits] [clang] [lldb] [libcxx] [mlir] [openmp] [clang-tools-extra] [flang] [libc] [llvm] [BOLT] Deduplicate equal offsets in BAT (PR #76905)

[Amir Ayupov via llvm-branch-commits]

================
@@ -110,6 +111,34 @@ void BoltAddressTranslation::write(const BinaryContext &BC, raw_ostream &OS) {
   outs() << "BOLT-INFO: Wrote " << Maps.size() << " BAT maps\n";
 }
 
+APInt BoltAddressTranslation::calculateBranchEntriesBitMask(MapTy &Map,
+                                                            size_t EqualElems) {
+  APInt BitMask(alignTo(EqualElems, 8), 0);
+  size_t Index = 0;
+  for (std::pair<const uint32_t, uint32_t> &KeyVal : Map) {
+    if (Index == EqualElems)
+      break;
+    const uint32_t OutputOffset = KeyVal.second;
+    if (OutputOffset & BRANCHENTRY)
+      BitMask.setBit(Index);
+    ++Index;
+  }
+  return BitMask;
+}
+
+size_t BoltAddressTranslation::getNumEqualOffsets(const MapTy &Map) const {
+  size_t EqualOffsets = 0;
+  for (const std::pair<const uint32_t, uint32_t> &KeyVal : Map) {
+    const uint32_t OutputOffset = KeyVal.first;
+    const uint32_t InputOffset = KeyVal.second >> 1;
+    if (OutputOffset == InputOffset)
+      ++EqualOffsets;
+    else
+      break;
----------------
aaupov wrote:

Yes, that's by design. The idea is that if OutputOffset==InputOffset, only emit one of them. It's expected that if offsets start to diverge at some point, they're not reconverging.

https://github.com/llvm/llvm-project/pull/76905