llvm/llvm/lib/Transforms/HipStdPar/HipStdPar.cpp

//===----- HipStdPar.cpp - HIP C++ Standard Parallelism Support Passes ----===//
//
// 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 file implements two passes that enable HIP C++ Standard Parallelism
// Support:
//
// 1. AcceleratorCodeSelection (required): Given that only algorithms are
//    accelerated, and that the accelerated implementation exists in the form of
//    a compute kernel, we assume that only the kernel, and all functions
//    reachable from it, constitute code that the user expects the accelerator
//    to execute. Thus, we identify the set of all functions reachable from
//    kernels, and then remove all unreachable ones. This last part is necessary
//    because it is possible for code that the user did not expect to execute on
//    an accelerator to contain constructs that cannot be handled by the target
//    BE, which cannot be provably demonstrated to be dead code in general, and
//    thus can lead to mis-compilation. The degenerate case of this is when a
//    Module contains no kernels (the parent TU had no algorithm invocations fit
//    for acceleration), which we handle by completely emptying said module.
//    **NOTE**: The above does not handle indirectly reachable functions i.e.
//              it is possible to obtain a case where the target of an indirect
//              call is otherwise unreachable and thus is removed; this
//              restriction is aligned with the current `-hipstdpar` limitations
//              and will be relaxed in the future.
//
// 2. AllocationInterposition (required only when on-demand paging is
//    unsupported): Some accelerators or operating systems might not support
//    transparent on-demand paging. Thus, they would only be able to access
//    memory that is allocated by an accelerator-aware mechanism. For such cases
//    the user can opt into enabling allocation / deallocation interposition,
//    whereby we replace calls to known allocation / deallocation functions with
//    calls to runtime implemented equivalents that forward the requests to
//    accelerator-aware interfaces. We also support freeing system allocated
//    memory that ends up in one of the runtime equivalents, since this can
//    happen if e.g. a library that was compiled without interposition returns
//    an allocation that can be validly passed to `free`.
//===----------------------------------------------------------------------===//

#include "llvm/Transforms/HipStdPar/HipStdPar.h"

#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/Analysis/CallGraph.h"
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/Module.h"
#include "llvm/Transforms/Utils/ModuleUtils.h"

#include <cassert>
#include <string>
#include <utility>

usingnamespacellvm;

template<typename T>
static inline void eraseFromModule(T &ToErase) { … }

static inline bool checkIfSupported(GlobalVariable &G) { … }

static inline void clearModule(Module &M) { … }

static inline void maybeHandleGlobals(Module &M) { … }

template<unsigned N>
static inline void removeUnreachableFunctions(
  const SmallPtrSet<const Function *, N>& Reachable, Module &M) { … }

static inline bool isAcceleratorExecutionRoot(const Function *F) { … }

static inline bool checkIfSupported(const Function *F, const CallBase *CB) { … }

PreservedAnalyses
  HipStdParAcceleratorCodeSelectionPass::run(Module &M,
                                             ModuleAnalysisManager &MAM) { … }

static constexpr std::pair<StringLiteral, StringLiteral> ReplaceMap[]{ … };

PreservedAnalyses
HipStdParAllocationInterpositionPass::run(Module &M, ModuleAnalysisManager&) { … }