[llvm] [BOLT] Add support for Linux kernel .smp_locks section (PR #90798)

[Davide Italiano via llvm-commits]

================
@@ -439,79 +408,73 @@ void LinuxKernelRewriter::processLKKSymtab(bool IsGPL) {
 
 /// .smp_locks section contains PC-relative references to instructions with LOCK
 /// prefix. The prefix can be converted to NOP at boot time on non-SMP systems.
-void LinuxKernelRewriter::processLKSMPLocks() {
-  ErrorOr<BinarySection &> SectionOrError =
+Error LinuxKernelRewriter::processSMPLocks() {
+  ErrorOr<BinarySection &> SMPLocksSection =
       BC.getUniqueSectionByName(".smp_locks");
-  if (!SectionOrError)
-    return;
+  if (!SMPLocksSection)
+    return Error::success();
 
-  uint64_t SectionSize = SectionOrError->getSize();
-  const uint64_t SectionAddress = SectionOrError->getAddress();
-  assert((SectionSize % 4) == 0 &&
-         "The size of the .smp_locks section should be a multiple of 4");
+  const uint64_t SectionSize = SMPLocksSection->getSize();
+  const uint64_t SectionAddress = SMPLocksSection->getAddress();
+  if (SectionSize % SMP_LOCKS_ENTRY_SIZE)
+    return createStringError(errc::executable_format_error,
+                             "bad size of .smp_locks section");
 
-  for (uint64_t I = 0; I < SectionSize; I += 4) {
-    const uint64_t EntryAddress = SectionAddress + I;
-    ErrorOr<uint64_t> Offset = BC.getSignedValueAtAddress(EntryAddress, 4);
-    assert(Offset && "Reading valid PC-relative offset for a .smp_locks entry");
-    int32_t SignedOffset = *Offset;
-    uint64_t RefAddress = EntryAddress + SignedOffset;
+  DataExtractor DE = DataExtractor(SMPLocksSection->getContents(),
+                                   BC.AsmInfo->isLittleEndian(),
+                                   BC.AsmInfo->getCodePointerSize());
+  DataExtractor::Cursor Cursor(0);
+  while (Cursor && Cursor.tell() < SectionSize) {
+    const uint64_t Offset = Cursor.tell();
+    const uint64_t IP = SectionAddress + Offset + (int32_t)DE.getU32(Cursor);
+
+    // Consume the status of the cursor.
+    if (!Cursor)
+      return createStringError(errc::executable_format_error,
+                               "error while reading .smp_locks: %s",
+                               toString(Cursor.takeError()).c_str());
 
-    BinaryFunction *ContainingBF =
-        BC.getBinaryFunctionContainingAddress(RefAddress);
-    if (!ContainingBF)
+    if (opts::DumpSMPLocks)
+      BC.outs() << "SMP lock at 0x: " << Twine::utohexstr(IP) << '\n';
+
+    BinaryFunction *BF = BC.getBinaryFunctionContainingAddress(IP);
+    if (!BF || !BC.shouldEmit(*BF))
       continue;
 
-    insertLKMarker(RefAddress, I, SignedOffset, true, ".smp_locks");
-  }
-}
+    MCInst *Inst = BF->getInstructionAtOffset(IP - BF->getAddress());
+    if (!Inst)
+      return createStringError(errc::executable_format_error,
+                               "no instruction matches lock at 0x%" PRIx64, IP);
 
-void LinuxKernelRewriter::updateLKMarkers() {
-  if (LKMarkers.size() == 0)
-    return;
+    // Check for duplicate entries.
+    if (BC.MIB->hasAnnotation(*Inst, "SMPLock"))
----------------
dcci wrote:

is this a defensive check or something that can actually happen?

https://github.com/llvm/llvm-project/pull/90798