[llvm] [AMDGPU] Add amdgpu-lower-special-lds pass to lower named-barrier LDS (PR #165692)

[Pierre van Houtryve via llvm-commits]

================
@@ -0,0 +1,234 @@
+//===-- AMDGPULowerSpecialLDS.cpp -----------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass lowers the named barriers LDS globals which needs
+// special address assignment.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "AMDGPUMemoryUtils.h"
+#include "AMDGPUTargetMachine.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/Analysis/CallGraph.h"
+#include "llvm/CodeGen/TargetPassConfig.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/ReplaceConstant.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Pass.h"
+
+#include <algorithm>
+
+#define DEBUG_TYPE "amdgpu-lower-special-lds"
+
+using namespace llvm;
+using namespace AMDGPU;
+
+namespace {
+
+// If GV is also used directly by other kernels, create a new GV
+// used only by this kernel and its function.
+static GlobalVariable *uniquifyGVPerKernel(Module &M, GlobalVariable *GV,
+                                           Function *KF) {
+  bool NeedsReplacement = false;
+  for (Use &U : GV->uses()) {
+    if (auto *I = dyn_cast<Instruction>(U.getUser())) {
+      Function *F = I->getFunction();
+      if (isKernelLDS(F) && F != KF) {
+        NeedsReplacement = true;
+        break;
+      }
+    }
+  }
+  if (!NeedsReplacement)
+    return GV;
+  // Create a new GV used only by this kernel and its function
+  GlobalVariable *NewGV = new GlobalVariable(
+      M, GV->getValueType(), GV->isConstant(), GV->getLinkage(),
+      GV->getInitializer(), GV->getName() + "." + KF->getName(), nullptr,
+      GV->getThreadLocalMode(), GV->getType()->getAddressSpace());
+  NewGV->copyAttributesFrom(GV);
+  for (Use &U : make_early_inc_range(GV->uses())) {
+    if (auto *I = dyn_cast<Instruction>(U.getUser())) {
+      Function *F = I->getFunction();
+      if (!isKernelLDS(F) || F == KF) {
+        U.getUser()->replaceUsesOfWith(GV, NewGV);
+      }
+    }
+  }
+  return NewGV;
+}
+
+// Write the specified address into metadata where it can be retrieved by
+// the assembler. Format is a half open range, [Address Address+1)
+static void recordLDSAbsoluteAddress(Module *M, GlobalVariable *GV,
+                                     uint32_t Address) {
+  LLVMContext &Ctx = M->getContext();
+  auto *IntTy = M->getDataLayout().getIntPtrType(Ctx, AMDGPUAS::LOCAL_ADDRESS);
+  auto *MinC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address));
+  auto *MaxC = ConstantAsMetadata::get(ConstantInt::get(IntTy, Address + 1));
+  GV->setMetadata(LLVMContext::MD_absolute_symbol,
+                  MDNode::get(Ctx, {MinC, MaxC}));
+}
+
+template <typename T> std::vector<T> sortByName(std::vector<T> &&V) {
+  llvm::sort(V, [](const auto *L, const auto *R) {
+    return L->getName() < R->getName();
+  });
+  return {std::move(V)};
+}
+
+// Main utility function for special LDS variables lowering.
+static bool lowerSpecialLDSVariables(
+    Module &M, LDSUsesInfoTy &LDSUsesInfo,
+    VariableFunctionMap &LDSToKernelsThatNeedToAccessItIndirectly) {
+  bool Changed = false;
+  const DataLayout &DL = M.getDataLayout();
+  // The 1st round: give module-absolute assignments
+  int NumAbsolutes = 0;
+  std::vector<GlobalVariable *> OrderedGVs;
+  for (auto &K : LDSToKernelsThatNeedToAccessItIndirectly) {
+    GlobalVariable *GV = K.first;
+    if (!isNamedBarrier(*GV))
+      continue;
+    // give a module-absolute assignment if it is indirectly accessed by
+    // multiple kernels. This is not precise, but we don't want to duplicate
+    // a function when it is called by multiple kernels.
+    if (LDSToKernelsThatNeedToAccessItIndirectly[GV].size() > 1) {
+      OrderedGVs.push_back(GV);
+    } else {
+      // leave it to the 2nd round, which will give a kernel-relative
+      // assignment if it is only indirectly accessed by one kernel
+      LDSUsesInfo.direct_access[*K.second.begin()].insert(GV);
+    }
+    LDSToKernelsThatNeedToAccessItIndirectly.erase(GV);
+  }
+  OrderedGVs = sortByName(std::move(OrderedGVs));
+  for (GlobalVariable *GV : OrderedGVs) {
+    unsigned BarrierScope = llvm::AMDGPU::Barrier::BARRIER_SCOPE_WORKGROUP;
----------------
Pierre-vh wrote:

```suggestion
    unsigned BarrierScope = AMDGPU::Barrier::BARRIER_SCOPE_WORKGROUP;
```

https://github.com/llvm/llvm-project/pull/165692