[BOLT] Add split function support for the Linux kernel (#90541)

While rewriting the Linux kernel, we try to fit optimized functions into
their original boundaries. When a function becomes larger, we skip it
during the rewrite and end up with less than optimal code layout. To
overcome that issue, add support for --split-function option so that hot
part of the function could be fit into the original space. The cold part
should go to reserved space in the binary.

Author: maksfb

bolt/lib/Passes/SplitFunctions.cpp

@@ -715,6 +715,12 @@ Error SplitFunctions::runOnFunctions(BinaryContext &BC) {
   if (!opts::SplitFunctions)
     return Error::success();
 
+  if (BC.IsLinuxKernel && BC.BOLTReserved.empty()) {
+    BC.errs() << "BOLT-ERROR: split functions require reserved space in the "
+                 "Linux kernel binary\n";
+    exit(1);
+  }
+
   // If split strategy is not CDSplit, then a second run of the pass is not
   // needed after function reordering.
   if (BC.HasFinalizedFunctionOrder &&
@@ -829,6 +835,13 @@ void SplitFunctions::splitFunction(BinaryFunction &BF, SplitStrategy &S) {
         }
       }
     }
+
+    // Outlining blocks with dynamic branches is not supported yet.
+    if (BC.IsLinuxKernel) {
+      if (llvm::any_of(
+              *BB, [&](MCInst &Inst) { return BC.MIB->isDynamicBranch(Inst); }))
+        BB->setCanOutline(false);
+    }
   }
 
   BF.getLayout().updateLayoutIndices();

bolt/lib/Rewrite/LinuxKernelRewriter.cpp

@@ -783,11 +783,9 @@ Error LinuxKernelRewriter::rewriteORCTables() {
   };
 
   // Emit new ORC entries for the emitted function.
-  auto emitORC = [&](const BinaryFunction &BF) -> Error {
-    assert(!BF.isSplit() && "Split functions not supported by ORC writer yet.");
-
+  auto emitORC = [&](const FunctionFragment &FF) -> Error {
     ORCState CurrentState = NullORC;
-    for (BinaryBasicBlock *BB : BF.getLayout().blocks()) {
+    for (BinaryBasicBlock *BB : FF) {
       for (MCInst &Inst : *BB) {
         ErrorOr<ORCState> ErrorOrState =
             BC.MIB->tryGetAnnotationAs<ORCState>(Inst, "ORC");
@@ -808,7 +806,36 @@ Error LinuxKernelRewriter::rewriteORCTables() {
     return Error::success();
   };
 
+  // Emit ORC entries for cold fragments. We assume that these fragments are
+  // emitted contiguously in memory using reserved space in the kernel. This
+  // assumption is validated in post-emit pass validateORCTables() where we
+  // check that ORC entries are sorted by their addresses.
+  auto emitColdORC = [&]() -> Error {
+    for (BinaryFunction &BF :
+         llvm::make_second_range(BC.getBinaryFunctions())) {
+      if (!BC.shouldEmit(BF))
+        continue;
+      for (FunctionFragment &FF : BF.getLayout().getSplitFragments())
+        if (Error E = emitORC(FF))
+          return E;
+    }
+
+    return Error::success();
+  };
+
+  bool ShouldEmitCold = !BC.BOLTReserved.empty();
   for (ORCListEntry &Entry : ORCEntries) {
+    if (ShouldEmitCold && Entry.IP > BC.BOLTReserved.start()) {
+      if (Error E = emitColdORC())
+        return E;
+
+      // Emit terminator entry at the end of the reserved region.
+      if (Error E = emitORCEntry(BC.BOLTReserved.end(), NullORC))
+        return E;
+
+      ShouldEmitCold = false;
+    }
+
     // Emit original entries for functions that we haven't modified.
     if (!Entry.BF || !BC.shouldEmit(*Entry.BF)) {
       // Emit terminator only if it marks the start of a function.
@@ -822,7 +849,7 @@ Error LinuxKernelRewriter::rewriteORCTables() {
     // Emit all ORC entries for a function referenced by an entry and skip over
     // the rest of entries for this function by resetting its ORC attribute.
     if (Entry.BF->hasORC()) {
-      if (Error E = emitORC(*Entry.BF))
+      if (Error E = emitORC(Entry.BF->getLayout().getMainFragment()))
         return E;
       Entry.BF->setHasORC(false);
     }
@@ -831,10 +858,9 @@ Error LinuxKernelRewriter::rewriteORCTables() {
   LLVM_DEBUG(dbgs() << "BOLT-DEBUG: emitted " << NumEmitted
                     << " ORC entries\n");
 
-  // Replicate terminator entry at the end of sections to match the original
-  // table sizes.
-  const BinaryFunction &LastBF = BC.getBinaryFunctions().rbegin()->second;
-  const uint64_t LastIP = LastBF.getAddress() + LastBF.getMaxSize();
+  // Populate ORC tables with a terminator entry with max address to match the
+  // original table sizes.
+  const uint64_t LastIP = std::numeric_limits<uint64_t>::max();
   while (UnwindWriter.bytesRemaining()) {
     if (Error E = emitORCEntry(LastIP, NullORC, nullptr, /*Force*/ true))
       return E;