================
@@ -0,0 +1,360 @@
+#include "llvm/Transforms/Utils/SplitModuleCG.h"
+#include "llvm/Bitcode/BitcodeReader.h"
+#include "llvm/Bitcode/BitcodeWriter.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/GlobalValue.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Value.h"
+#include "llvm/Support/MD5.h"
+#include "llvm/Transforms/Utils/Cloning.h"
+#include "llvm/Transforms/Utils/ModuleUtils.h"
+#include <thread>
+using namespace llvm;
+
+#define DEBUG_TYPE "split-module-cg"
+
+namespace {
+
+static cl::opt<bool>
+ enablePrintSimplifiedCallGraph("enable-print-simplified-callgraph",
+ cl::Hidden, cl::init(false),
+ cl::desc("print SimplifiedCallGraph"));
+
+using PartitionID = unsigned;
+
+/// Returns whether duplicate definitions of \p F across partitions may be
+/// downgraded to available_externally. This is safe for external functions
+/// (either originally external or promoted by externalize), and for
+/// weak_odr/linkonce_odr functions whose equivalent definitions can be
+/// deduplicated to reduce codegen. Interposable linkages (weak/linkonce
+/// non-ODR) are excluded since downgrading them would change their
+/// optimization semantics.
+static bool canDowngradeToAvailableExternally(const Function &F) {
+ return !F.isDeclaration() &&
+ (F.hasExternalLinkage() || F.hasWeakODRLinkage() ||
+ F.hasLinkOnceODRLinkage());
+}
+
+} // namespace
+
+std::vector<DenseSet<const Function *>> SplitModuleCG::doPartitioning() {
+ LLVM_DEBUG(dbgs() << "\n--Partitioning Starts--\n");
+ // Performs all of the partitioning work on M.
+ assert(N != 0 && "Partition count must be at least 1");
+ std::vector<DenseSet<const Function *>> Partitions;
+ Partitions.resize(N);
+
+ auto ComparePartitions = [](const std::pair<PartitionID, CostType> &LHS,
+ const std::pair<PartitionID, CostType> &RHS) {
+ // When two partitions have the same cost, assign to the one with the
+ // biggest ID first. This allows us to put things in P0 last, because P0
may
+ // have other stuff added later.
+ if (LHS.second == RHS.second)
+ return LHS.first < RHS.first;
+ return LHS.second > RHS.second;
+ };
+
+ std::vector<std::pair<PartitionID, CostType>> BalancingQueue;
+ for (unsigned I = 0; I < N; ++I)
+ BalancingQueue.emplace_back(I, 0);
+
+ for (auto &CurFn : FWDWorkList) {
+ // Normal "load-balancing", assign to partition with least pressure.
+ auto [PID, _] = BalancingQueue.back();
+
+ // Insert the root function and its dependencies into the partition,
+ // tracking the cost of newly inserted functions so the balancing queue
+ // can be updated. CurFn.Dependencies includes the root F itself.
+ auto &FnsInPart = Partitions[PID];
+ CostType AddedCost = 0;
+ for (const Function *Dep : CurFn.Dependencies)
+ if (FnsInPart.insert(Dep).second)
+ AddedCost += FuncsCosts.lookup(Dep);
+
+ // Update the balancing queue. We scan backwards because in the common
+ // case the target partition is at the end of the sorted queue.
----------------
mmjjpp wrote:
Done
https://github.com/llvm/llvm-project/pull/198702
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
