[llvm] [AMDGPU] Add AMDGPU-specific module splitting (PR #89245)
Pierre van Houtryve via llvm-commits
llvm-commits at lists.llvm.org
Thu Apr 18 07:51:34 PDT 2024
https://github.com/Pierre-vh created https://github.com/llvm/llvm-project/pull/89245
(See #83128 to review first commit)
This enables the --lto-partitions option to work more consistently.
This module splitting logic is fully aware of AMDGPU modules and their specificities and takes advantage of
them to split modules in a way that avoids compilation issue (such as resource usage being incorrectly represented).
This also includes a logging system that's more elaborate than just LLVM_DEBUG which allows
printing logs to uniquely named files, and optionally with all value names hidden so they can be safely shared without leaking informatiton about the source. Logs can also be enabled through an environment variable, which avoids the sometimes complicated process of passing a -mllvm option all the way from clang driver to the offload linker that handles full LTO codegen.
>From 6bb50651090d6dc474198491b8aa4b2f6dad8dda Mon Sep 17 00:00:00 2001
From: pvanhout <pierre.vanhoutryve at amd.com>
Date: Tue, 27 Feb 2024 13:26:18 +0100
Subject: [PATCH 1/2] [RFC][LTO] Allow target-specific module splittting
Allow targets to implement custom module splitting logic for
--lto-partitions.
---
llvm/include/llvm/Target/TargetMachine.h | 12 +++
llvm/lib/LTO/LTOBackend.cpp | 13 ++-
llvm/tools/llvm-split/CMakeLists.txt | 7 ++
llvm/tools/llvm-split/llvm-split.cpp | 101 +++++++++++++++++------
4 files changed, 104 insertions(+), 29 deletions(-)
diff --git a/llvm/include/llvm/Target/TargetMachine.h b/llvm/include/llvm/Target/TargetMachine.h
index ceb371bdc73480..48ea3cfe02775b 100644
--- a/llvm/include/llvm/Target/TargetMachine.h
+++ b/llvm/include/llvm/Target/TargetMachine.h
@@ -418,6 +418,18 @@ class TargetMachine {
virtual unsigned getAddressSpaceForPseudoSourceKind(unsigned Kind) const {
return 0;
}
+
+ /// Entry point for module splitting. Targets can implement custom module
+ /// splitting logic, mainly used by LTO for --lto-partitions.
+ ///
+ /// \returns `true` if the module was split, `false` otherwise. When `false`
+ /// is returned, it is assumed that \p ModuleCallback has never been called
+ /// and \p M has not been modified.
+ virtual bool splitModule(
+ Module &M, unsigned NumParts,
+ function_ref<void(std::unique_ptr<Module> MPart)> ModuleCallback) const {
+ return false;
+ }
};
/// This class describes a target machine that is implemented with the LLVM
diff --git a/llvm/lib/LTO/LTOBackend.cpp b/llvm/lib/LTO/LTOBackend.cpp
index 71e8849dc3cc91..d4b89ede2d7134 100644
--- a/llvm/lib/LTO/LTOBackend.cpp
+++ b/llvm/lib/LTO/LTOBackend.cpp
@@ -436,8 +436,7 @@ static void splitCodeGen(const Config &C, TargetMachine *TM,
unsigned ThreadCount = 0;
const Target *T = &TM->getTarget();
- SplitModule(
- Mod, ParallelCodeGenParallelismLevel,
+ const auto HandleModulePartition =
[&](std::unique_ptr<Module> MPart) {
// We want to clone the module in a new context to multi-thread the
// codegen. We do it by serializing partition modules to bitcode
@@ -469,8 +468,14 @@ static void splitCodeGen(const Config &C, TargetMachine *TM,
// Pass BC using std::move to ensure that it get moved rather than
// copied into the thread's context.
std::move(BC), ThreadCount++);
- },
- false);
+ };
+
+ // Try target-specific module splitting first, then fallback to the default.
+ if (!TM->splitModule(Mod, ParallelCodeGenParallelismLevel,
+ HandleModulePartition)) {
+ SplitModule(Mod, ParallelCodeGenParallelismLevel, HandleModulePartition,
+ false);
+ }
// Because the inner lambda (which runs in a worker thread) captures our local
// variables, we need to wait for the worker threads to terminate before we
diff --git a/llvm/tools/llvm-split/CMakeLists.txt b/llvm/tools/llvm-split/CMakeLists.txt
index 52eedeb9f53f32..0579298462d113 100644
--- a/llvm/tools/llvm-split/CMakeLists.txt
+++ b/llvm/tools/llvm-split/CMakeLists.txt
@@ -1,9 +1,16 @@
set(LLVM_LINK_COMPONENTS
+ AllTargetsAsmParsers
+ AllTargetsCodeGens
+ AllTargetsDescs
+ AllTargetsInfos
TransformUtils
BitWriter
+ CodeGen
Core
IRReader
+ MC
Support
+ Target
)
add_llvm_tool(llvm-split
diff --git a/llvm/tools/llvm-split/llvm-split.cpp b/llvm/tools/llvm-split/llvm-split.cpp
index c6e20e0373c717..252e04e77941f2 100644
--- a/llvm/tools/llvm-split/llvm-split.cpp
+++ b/llvm/tools/llvm-split/llvm-split.cpp
@@ -1,4 +1,4 @@
-//===-- llvm-split: command line tool for testing module splitter ---------===//
+//===-- llvm-split: command line tool for testing module splitting --------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
@@ -6,7 +6,8 @@
//
//===----------------------------------------------------------------------===//
//
-// This program can be used to test the llvm::SplitModule function.
+// This program can be used to test the llvm::SplitModule and
+// TargetMachine::splitModule functions.
//
//===----------------------------------------------------------------------===//
@@ -15,12 +16,17 @@
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Verifier.h"
#include "llvm/IRReader/IRReader.h"
+#include "llvm/MC/TargetRegistry.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/FileSystem.h"
+#include "llvm/Support/InitLLVM.h"
#include "llvm/Support/SourceMgr.h"
+#include "llvm/Support/TargetSelect.h"
#include "llvm/Support/ToolOutputFile.h"
-#include "llvm/Support/raw_ostream.h"
#include "llvm/Support/WithColor.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetMachine.h"
+#include "llvm/TargetParser/Triple.h"
#include "llvm/Transforms/Utils/SplitModule.h"
using namespace llvm;
@@ -47,12 +53,45 @@ static cl::opt<bool>
cl::desc("Split without externalizing locals"),
cl::cat(SplitCategory));
+static cl::opt<std::string>
+ MTarget("mtarget",
+ cl::desc("Target triple. When present, a TargetMachine is created "
+ "and TargetMachine::splitModule is used instead of the "
+ "common SplitModule logic."),
+ cl::value_desc("triple"), cl::cat(SplitCategory));
+
+static cl::opt<std::string>
+ MCPU("mcpu", cl::desc("Target CPU, ignored if -mtarget is not used"),
+ cl::value_desc("cpu"), cl::cat(SplitCategory));
+
int main(int argc, char **argv) {
+ InitLLVM X(argc, argv);
+
+ // NOTE: If mtarget is not present we could avoid initializing targets to save
+ // time, but this is a testing tool and it's likely not worth the added
+ // complexity.
+ InitializeAllTargets();
+ InitializeAllTargetMCs();
+
LLVMContext Context;
SMDiagnostic Err;
cl::HideUnrelatedOptions({&SplitCategory, &getColorCategory()});
cl::ParseCommandLineOptions(argc, argv, "LLVM module splitter\n");
+ TargetMachine *TM = nullptr;
+ if (!MTarget.empty()) {
+ std::string Error;
+ const Target *T = TargetRegistry::lookupTarget(MTarget, Error);
+ if (!T) {
+ errs() << "unknown target '" << MTarget << "': " << Error << "\n";
+ return 1;
+ }
+
+ TargetOptions Options;
+ TM = T->createTargetMachine(MTarget, MCPU, /*FS*/ "", Options, std::nullopt,
+ std::nullopt);
+ }
+
std::unique_ptr<Module> M = parseIRFile(InputFilename, Err, Context);
if (!M) {
@@ -61,28 +100,40 @@ int main(int argc, char **argv) {
}
unsigned I = 0;
- SplitModule(
- *M, NumOutputs,
- [&](std::unique_ptr<Module> MPart) {
- std::error_code EC;
- std::unique_ptr<ToolOutputFile> Out(new ToolOutputFile(
- OutputFilename + utostr(I++), EC, sys::fs::OF_None));
- if (EC) {
- errs() << EC.message() << '\n';
- exit(1);
- }
-
- if (verifyModule(*MPart, &errs())) {
- errs() << "Broken module!\n";
- exit(1);
- }
-
- WriteBitcodeToFile(*MPart, Out->os());
-
- // Declare success.
- Out->keep();
- },
- PreserveLocals);
+ const auto HandleModulePart = [&](std::unique_ptr<Module> MPart) {
+ std::error_code EC;
+ std::unique_ptr<ToolOutputFile> Out(
+ new ToolOutputFile(OutputFilename + utostr(I++), EC, sys::fs::OF_None));
+ if (EC) {
+ errs() << EC.message() << '\n';
+ exit(1);
+ }
+
+ if (verifyModule(*MPart, &errs())) {
+ errs() << "Broken module!\n";
+ exit(1);
+ }
+
+ WriteBitcodeToFile(*MPart, Out->os());
+
+ // Declare success.
+ Out->keep();
+ };
+
+ if (TM) {
+ if (PreserveLocals) {
+ errs() << "warning: -preserve-locals has no effect when using "
+ "TargetMachine::splitModule\n";
+ }
+
+ if (TM->splitModule(*M, NumOutputs, HandleModulePart))
+ return 0;
+
+ errs() << "warning:"
+ "TargetMachine::splitModule failed, falling back to default "
+ "splitModule implementation\n";
+ }
+ SplitModule(*M, NumOutputs, HandleModulePart, PreserveLocals);
return 0;
}
>From 6c25662356692df32229bcd935669c56d571c771 Mon Sep 17 00:00:00 2001
From: pvanhout <pierre.vanhoutryve at amd.com>
Date: Thu, 18 Apr 2024 16:48:52 +0200
Subject: [PATCH 2/2] [AMDGPU] Add AMDGPU-specific module splitting
This enables the --lto-partitions option to work more consistently.
This module splitting logic is fully aware of AMDGPU modules and their specificities and takes advantage of
them to split modules in a way that avoids compilation issue (such as resource usage being incorrectly represented).
This also includes a logging system that's more elaborate than just LLVM_DEBUG which allows
printing logs to uniquely named files, and optionally with all value names hidden so they can be safely shared without leaking informatiton about the source. Logs can also be enabled through an environment variable, which avoids the sometimes complicated process of passing a -mllvm option all the way from clang driver to the offload linker that handles full LTO codegen.
---
llvm/lib/Target/AMDGPU/AMDGPUSplitModule.cpp | 733 ++++++++++++++++++
llvm/lib/Target/AMDGPU/AMDGPUSplitModule.h | 30 +
.../lib/Target/AMDGPU/AMDGPUTargetMachine.cpp | 8 +
llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.h | 4 +
llvm/lib/Target/AMDGPU/CMakeLists.txt | 1 +
.../address-taken-externalize-with-call.ll | 46 ++
.../AMDGPU/address-taken-externalize.ll | 37 +
.../llvm-split/AMDGPU/kernels-cost-ranking.ll | 54 ++
.../llvm-split/AMDGPU/kernels-dependencies.ll | 50 ++
.../AMDGPU/kernels-dependency-duplication.ll | 41 +
.../AMDGPU/kernels-dependency-external.ll | 43 +
.../AMDGPU/kernels-dependency-indirect.ll | 69 ++
.../kernels-global-variables-noexternal.ll | 42 +
.../AMDGPU/kernels-global-variables.ll | 44 ++
.../AMDGPU/kernels-load-balancing.ll | 75 ++
.../AMDGPU/kernels-no-dependencies.ll | 39 +
.../AMDGPU/large-kernels-merging.ll | 98 +++
.../tools/llvm-split/AMDGPU/lit.local.cfg | 2 +
18 files changed, 1416 insertions(+)
create mode 100644 llvm/lib/Target/AMDGPU/AMDGPUSplitModule.cpp
create mode 100644 llvm/lib/Target/AMDGPU/AMDGPUSplitModule.h
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/address-taken-externalize-with-call.ll
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/address-taken-externalize.ll
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/kernels-cost-ranking.ll
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/kernels-dependencies.ll
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-duplication.ll
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-external.ll
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-indirect.ll
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/kernels-global-variables-noexternal.ll
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/kernels-global-variables.ll
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/kernels-load-balancing.ll
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/kernels-no-dependencies.ll
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/large-kernels-merging.ll
create mode 100644 llvm/test/tools/llvm-split/AMDGPU/lit.local.cfg
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSplitModule.cpp b/llvm/lib/Target/AMDGPU/AMDGPUSplitModule.cpp
new file mode 100644
index 00000000000000..fa47d494f04148
--- /dev/null
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSplitModule.cpp
@@ -0,0 +1,733 @@
+//===- AMDGPUSplitModule.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
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file Implements a module splitting algorithm designed to support the
+/// FullLTO --lto-partitions option for parallel codegen. This is completely
+/// different from the common SplitModule pass, as this system is designed with
+/// AMDGPU in mind.
+///
+/// The basic idea of this module splitting implementation is the same as
+/// SplitModule: load-balance the module's functions across a set of N
+/// partitions to allow parallel codegen. However, it does it very
+/// differently than the target-agnostic variant:
+/// - Kernels are used as the module's "roots".
+/// They're known entry points on AMDGPU, and everything else is often
+/// internal only.
+/// - Each kernel has a set of dependencies, and when a kernel and its
+/// dependencies is considered "big", we try to put it in a partition where
+/// most dependencies are already imported, to avoid duplicating large
+/// amounts of code.
+/// - There's special care for indirect calls in order to ensure
+/// AMDGPUResourceUsageAnalysis can work correctly.
+///
+/// This file also includes a more elaborate logging system to enable
+/// users to easily generate logs that (if desired) do not include any value
+/// names, in order to not leak information about the source file.
+/// Such logs are very helpful to understand and fix potential issues with
+/// module splitting.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUSplitModule.h"
+#include "AMDGPUTargetMachine.h"
+#include "Utils/AMDGPUBaseInfo.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/StringExtras.h"
+#include "llvm/ADT/StringRef.h"
+#include "llvm/Analysis/CallGraph.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/Instruction.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/User.h"
+#include "llvm/IR/Value.h"
+#include "llvm/Support/Casting.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/FileSystem.h"
+#include "llvm/Support/Process.h"
+#include "llvm/Support/SHA256.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Transforms/Utils/Cloning.h"
+#include <algorithm>
+#include <cassert>
+#include <iterator>
+#include <memory>
+#include <utility>
+#include <vector>
+
+using namespace llvm;
+
+#define DEBUG_TYPE "amdgpu-split-module"
+
+namespace {
+
+static cl::opt<float> LargeKernelFactor(
+ "amdgpu-module-splitting-large-kernel-threshold", cl::init(2.0), cl::Hidden,
+ cl::desc(
+ "consider a kernel as large and needing special treatment when it "
+ "exceeds the average cost of a partition by this factor; e;g. 2.0 "
+ "means if the kernel and its dependencies is 2 times bigger than "
+ "an average partition; 0 disables large kernels handling entirely"));
+
+static cl::opt<float> LargeKernelOverlapForMerge(
+ "amdgpu-module-splitting-large-kernel-merge-overlap", cl::init(0.8),
+ cl::Hidden,
+ cl::desc("defines how much overlap between two large kernel's dependencies "
+ "is needed to put them in the same partition"));
+
+static cl::opt<bool> NoExternalizeGlobals(
+ "amdgpu-module-splitting-no-externalize-globals", cl::Hidden,
+ cl::desc("disables externalization of global variable with local linkage; "
+ "may cause globals to be duplicated which increases binary size"));
+
+static cl::opt<std::string>
+ LogDirOpt("amdgpu-module-splitting-log-dir", cl::Hidden,
+ cl::desc("output directory for AMDGPU module splitting logs"));
+
+static cl::opt<bool>
+ LogPrivate("amdgpu-module-splitting-log-private", cl::Hidden,
+ cl::desc("hash value names before printing them in the AMDGPU "
+ "module splitting logs"));
+
+using CostType = InstructionCost::CostType;
+using PartitionID = unsigned;
+
+static std::string getName(const Value &V) {
+ static std::optional<bool> HideNames;
+ if (!HideNames) {
+ if (LogPrivate.getNumOccurrences())
+ HideNames = LogPrivate;
+ else {
+ const auto EV = sys::Process::GetEnv("AMD_SPLIT_MODULE_LOG_PRIVATE");
+ HideNames = (EV.value_or("0") != "0");
+ }
+ }
+
+ if (!*HideNames)
+ return V.getName().str();
+ return toHex(SHA256::hash(arrayRefFromStringRef(V.getName())),
+ /*LowerCase*/ true);
+}
+
+/// Main logging helper.
+///
+/// Logging can be configured by the following environment variable.
+/// AMD_SPLIT_MODULE_LOG_DIR=<filepath>
+/// If set, uses <filepath> as the directory to write logfiles to
+/// each time module splitting is used.
+/// AMD_SPLIT_MODULE_LOG_PRIVATE
+/// If set to anything other than zero, all names are hidden.
+///
+/// Both environment variables have corresponding CL options which
+/// takes priority over them.
+///
+/// Any output printed to the log files is also printed to dbgs() when -debug is
+/// used and LLVM_DEBUG is defined.
+///
+/// This approach has a small disadvantage over LLVM_DEBUG though: logging logic
+/// cannot be removed from the code (by building without debug). This probably
+/// has a small performance cost because if some computation/formatting is
+/// needed for logging purpose, it may be done everytime only to be ignored
+/// by the logger.
+///
+/// As this pass only runs once and is not doing anything computationally
+/// expensive, this is likely a reasonable trade-off.
+///
+/// If some computation should really be avoided when unused, users of the class
+/// can check whether any logging will occur by using the bool operator.
+///
+/// \code
+/// if (SML) {
+/// // Executes only if logging to a file or if -debug is available and
+/// used.
+/// }
+/// \endcode
+class SplitModuleLogger {
+public:
+ SplitModuleLogger(const Module &M) {
+ std::string LogDir = LogDirOpt;
+ if (LogDir.empty())
+ LogDir = sys::Process::GetEnv("AMD_SPLIT_MODULE_LOG_DIR").value_or("");
+
+ // No log dir specified means we don't need to log to a file.
+ // We may still log to dbgs(), though.
+ if (LogDir.empty())
+ return;
+
+ if (!sys::fs::is_directory(LogDir)) {
+ report_fatal_error("invalid AMDGPU split module log directory: '" +
+ Twine(LogDir) + "' is not a directory",
+ /*CrashDiag=*/false);
+ }
+
+ // If a log directory is specified, create a new file with a unique name in
+ // that directory.
+ SmallString<0> FilePath;
+ int Fd;
+ std::string LogFile = (LogDir + "/" + "Module-%%-%%-%%-%%-%%-%%-%%.txt");
+ if (auto Err = sys::fs::createUniqueFile(LogFile, Fd, FilePath)) {
+ dbgs() << LogFile << "\n";
+ std::string Msg =
+ "Failed to create log file at '" + LogDir + "': " + Err.message();
+ report_fatal_error(StringRef(Msg),
+ /*CrashDiag=*/false);
+ }
+
+ FileOS = std::make_unique<raw_fd_ostream>(Fd, /*shouldClose*/ true);
+ }
+
+ bool hasLogFile() const { return FileOS != nullptr; }
+
+ raw_ostream &logfile() {
+ assert(FileOS && "no logfile!");
+ return *FileOS;
+ }
+
+ /// \returns true if this SML will log anything either to a file or dbgs().
+ /// Can be used to avoid expensive computations that are ignored when logging
+ /// is disabled.
+ operator bool() const {
+ return hasLogFile() || (DebugFlag && isCurrentDebugType(DEBUG_TYPE));
+ }
+
+private:
+ std::unique_ptr<raw_fd_ostream> FileOS;
+};
+
+template <typename Ty>
+static SplitModuleLogger &operator<<(SplitModuleLogger &SML, const Ty &Val) {
+ static_assert(
+ !std::is_same_v<Ty, Value>,
+ "do not print values to logs directly, use handleName instead!");
+ LLVM_DEBUG(dbgs() << Val);
+ if (SML.hasLogFile())
+ SML.logfile() << Val;
+ return SML;
+}
+
+/// Calculate the cost of each function in \p M
+/// \param SML Log Helper
+/// \param TM TargetMachine instance used to retrieve TargetTransformInfo.
+/// \param M Module to analyze.
+/// \param CostMap[out] Resulting Function -> Cost map.
+/// \return The module's total cost.
+static CostType
+calculateFunctionCosts(SplitModuleLogger &SML, const AMDGPUTargetMachine &TM,
+ Module &M,
+ DenseMap<const Function *, CostType> &CostMap) {
+ CostType ModuleCost = 0;
+ CostType KernelCost = 0;
+
+ for (auto &Fn : M) {
+ if (Fn.isDeclaration())
+ continue;
+
+ CostType FnCost = 0;
+ auto TTI = TM.getTargetTransformInfo(Fn);
+
+ for (auto &BB : Fn) {
+ for (auto &I : BB) {
+ auto Cost =
+ TTI.getInstructionCost(&I, TargetTransformInfo::TCK_CodeSize);
+ assert(Cost != InstructionCost::getMax());
+ // Assume expensive if we can't tell the cost of an instruction.
+ CostType CostVal =
+ Cost.getValue().value_or(TargetTransformInfo::TCC_Expensive);
+ assert((FnCost + CostVal) >= FnCost && "Overflow!");
+ FnCost += CostVal;
+ }
+ }
+
+ assert(FnCost != 0);
+
+ CostMap[&Fn] = FnCost;
+ assert((ModuleCost + FnCost) >= ModuleCost && "Overflow!");
+ ModuleCost += FnCost;
+
+ if (AMDGPU::isKernelCC(&Fn))
+ KernelCost += FnCost;
+ }
+
+ CostType FnCost = (ModuleCost - KernelCost);
+ SML << "=> Total Module Cost: " << ModuleCost << "\n"
+ << " => KernelCost: " << KernelCost << " ("
+ << format("%0.2f", (float(KernelCost) / ModuleCost) * 100) << "%)\n"
+ << " => FnsCost: " << FnCost << " ("
+ << format("%0.2f", (float(FnCost) / ModuleCost) * 100) << "%)\n";
+
+ return ModuleCost;
+}
+
+/// When a kernel or any of its callees performs an indirect call, this function
+/// takes over \ref addAllDependencies and adds all potentially callable
+/// functions to \p Fns so they can be counted as dependencies of the kernel.
+///
+/// This is needed due to how AMDGPUResourceUsageAnalysis operates: in the
+/// presence of an indirect call, the function's resource usage is the same as
+/// the most expensive function in the module.
+/// \param M The module.
+/// \param Fns[out] Resulting list of functions.
+static void addAllIndirectCallDependencies(const Module &M,
+ DenseSet<const Function *> &Fns) {
+ for (const auto &Fn : M) {
+ if (!Fn.isDeclaration() && !AMDGPU::isEntryFunctionCC(Fn.getCallingConv()))
+ Fns.insert(&Fn);
+ }
+}
+
+/// Adds the functions that \p Fn may call to \p Fns, then recurses into each
+/// callee until all reachable functions have been gathered.
+///
+/// \param SML Log Helper
+/// \param CG Call graph for \p Fn's module.
+/// \param Fn Current function to look at.
+/// \param Fns[out] Resulting list of functions.
+/// \param HadIndirectCall[out] Set to true if an indirect call was seen at some
+/// point, either in \p Fn or in one of the function it calls. When that
+/// happens, we fall back to adding all callable functions inside \p Fn's module
+/// to \p Fns.
+/// \param HadExternalCall[out] Set to true if a call to an external function
+/// was seen at some point, either in \p Fn or in one of the function it calls
+static void addAllDependencies(SplitModuleLogger &SML, const CallGraph &CG,
+ const Function &Fn,
+ DenseSet<const Function *> &Fns,
+ bool &HadIndirectCall, bool &HadExternalCall) {
+ assert(!Fn.isDeclaration());
+
+ const Module &M = *Fn.getParent();
+ SmallVector<const Function *> WorkList({&Fn});
+ while (!WorkList.empty()) {
+ const auto &CurFn = *WorkList.pop_back_val();
+
+ // Scan for an indirect call. If such a call is found, we have to
+ // conservatively assume this can call all non-entrypoint functions in the
+ // module.
+ for (const auto &BB : CurFn) {
+ for (const auto &I : BB) {
+ const auto *CB = dyn_cast<CallBase>(&I);
+ if (!CB || !CB->isIndirectCall())
+ continue;
+
+ SML << "Indirect call detected in " << getName(CurFn)
+ << " - treating all non-entrypoint functions as "
+ "potential dependencies\n";
+
+ // TODO: Print an ORE as well ?
+ addAllIndirectCallDependencies(M, Fns);
+ HadIndirectCall = true;
+ return;
+ }
+ }
+
+ for (auto &CGEntry : *CG[&CurFn]) {
+ auto *Callee = CGEntry.second->getFunction();
+ if (!Callee)
+ continue;
+
+ assert(!AMDGPU::isKernelCC(Callee));
+
+ if (Callee->isDeclaration())
+ continue;
+
+ if (Callee->hasExternalLinkage())
+ HadExternalCall = true;
+
+ auto [It, Inserted] = Fns.insert(Callee);
+ if (Inserted)
+ WorkList.push_back(Callee);
+ }
+ }
+}
+
+/// Contains information about a kernel and its dependencies.
+struct KernelWithDependencies {
+ KernelWithDependencies(SplitModuleLogger &SML, CallGraph &CG,
+ const DenseMap<const Function *, CostType> &FnCosts,
+ const Function *Fn)
+ : Fn(Fn) {
+ addAllDependencies(SML, CG, *Fn, Dependencies, HasIndirectCall,
+ HasExternalCall);
+ TotalCost = FnCosts.at(Fn);
+ for (const auto *Dep : Dependencies)
+ TotalCost += FnCosts.at(Dep);
+ }
+
+ const Function *Fn = nullptr;
+ DenseSet<const Function *> Dependencies;
+ /// Whether \p Fn or any of its \ref Dependencies contains an indirect call.
+ bool HasIndirectCall = false;
+ /// Whether \p Fn or any of its \ref Dependencies contains a call to a
+ /// function with external linkage.
+ bool HasExternalCall = false;
+
+ CostType TotalCost = 0;
+
+ /// \returns true if this kernel and its dependencies can be considered large
+ /// according to \p Threshold.
+ bool isLarge(CostType Threshold) const {
+ return TotalCost > Threshold && !Dependencies.empty();
+ }
+};
+
+/// Calculates how much overlap there is between \p A and \p B.
+/// \return A number between 0.0 and 1.0, where 1.0 means A == B and 0.0 means A
+/// and B have no shared elements. Kernels do not count in overlap calculation.
+static float calculateOverlap(const DenseSet<const Function *> &A,
+ const DenseSet<const Function *> &B) {
+ DenseSet<const Function *> Total;
+ for (const auto *F : A) {
+ if (!AMDGPU::isKernelCC(F))
+ Total.insert(F);
+ }
+
+ if (Total.empty())
+ return 0.0f;
+
+ unsigned NumCommon = 0;
+ for (const auto *F : B) {
+ if (AMDGPU::isKernelCC(F))
+ continue;
+
+ auto [It, Inserted] = Total.insert(F);
+ if (!Inserted)
+ ++NumCommon;
+ }
+
+ return float(NumCommon) / Total.size();
+}
+
+/// Performs all of the partitioning work on \p M.
+/// \param SML Log Helper
+/// \param M Module to partition.
+/// \param NumParts Number of partitions to create.
+/// \param ModuleCost Total cost of all functions in \p M.
+/// \param FnCosts Map of Function -> Cost
+/// \param WorkList Kernels and their dependencies to process in order.
+/// \returns The created partitions (a vector of size \p NumParts )
+static std::vector<DenseSet<const Function *>>
+doPartitioning(SplitModuleLogger &SML, Module &M, unsigned NumParts,
+ CostType ModuleCost,
+ const DenseMap<const Function *, CostType> &FnCosts,
+ const SmallVector<KernelWithDependencies> &WorkList) {
+
+ SML << "\n--Partitioning Starts--\n";
+
+ // Calculate a "large kernel threshold". When more than one kernel's total
+ // import cost exceeds this value, we will try to merge it with other,
+ // similarly large kernels.
+ //
+ // e.g. let two kernels X and Y have a import cost of ~10% of the module, we
+ // assign X to a partition as usual, but when we get to Y, we check if it's
+ // worth also putting it in Y's partition.
+ const CostType LargeKernelThreshold =
+ LargeKernelFactor ? ((ModuleCost / NumParts) * LargeKernelFactor)
+ : std::numeric_limits<CostType>::max();
+
+ std::vector<DenseSet<const Function *>> Partitions;
+ Partitions.resize(NumParts);
+
+ // Assign a partition to each kernel, and try to keep the partitions more or
+ // less balanced. We do that through a priority queue sorted in reverse, so we
+ // can always look at the partition with the least content.
+ //
+ // There are some cases where we will be deliberately unbalanced though.
+ // - Large kernels: we try to merge with existing partitions to reduce code
+ // duplication.
+ // - Kernels with indirect or external calls always go in the first partition
+ // (P0).
+ auto ComparePartitions = [](const std::pair<PartitionID, CostType> &a,
+ const std::pair<PartitionID, CostType> &b) {
+ // 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 (a.second == b.second)
+ return a.first < b.first;
+ return a.second > b.second;
+ };
+
+ // We can't use priority_queue here because we need to be able to access any
+ // element. This makes this a bit inefficient as we need to sort it again
+ // everytime we change it, but it's a very small array anyway (likely under 64
+ // partitions) so it's a cheap operation.
+ std::vector<std::pair<PartitionID, CostType>> BalancingQueue;
+ for (unsigned I = 0; I < NumParts; ++I)
+ BalancingQueue.push_back(std::make_pair(I, 0));
+
+ // Helper function to handle assigning a kernel to a partition. This takes
+ // care of updating the balancing queue.
+ const auto AssignToPartition = [&](PartitionID PID,
+ const KernelWithDependencies &KWD) {
+ auto &FnsInPart = Partitions[PID];
+ FnsInPart.insert(KWD.Fn);
+ FnsInPart.insert(KWD.Dependencies.begin(), KWD.Dependencies.end());
+
+ SML << "assign " << getName(*KWD.Fn) << " to P" << PID << "\n -> ";
+ if (!KWD.Dependencies.empty()) {
+ SML << KWD.Dependencies.size() << " dependencies added\n";
+ };
+
+ // Update the balancing queue. we scan backwards because in the common case
+ // the partition is at the end.
+ for (auto &[QueuePID, Cost] : reverse(BalancingQueue)) {
+ if (QueuePID == PID) {
+ CostType NewCost = 0;
+ for (auto *Fn : Partitions[PID])
+ NewCost += FnCosts.at(Fn);
+
+ SML << "[Updating P" << PID << " Cost]:" << Cost << " -> " << NewCost;
+ if (Cost) {
+ SML << " (" << unsigned(((float(NewCost) / Cost) - 1) * 100)
+ << "% increase)";
+ }
+ SML << "\n";
+
+ Cost = NewCost;
+ }
+ }
+
+ sort(BalancingQueue, ComparePartitions);
+ };
+
+ for (auto &CurKernel : WorkList) {
+ // When a kernel has indirect calls, it must stay in the first partition
+ // alongside every reachable non-entry function. This is a nightmare case
+ // for splitting as it severely limits what we can do.
+ if (CurKernel.HasIndirectCall) {
+ SML << "Kernel with indirect call(s): " << getName(*CurKernel.Fn)
+ << " defaulting to P0\n";
+ AssignToPartition(0, CurKernel);
+ continue;
+ }
+
+ // When a kernel calls external functions, we have to keep it in the first
+ // partition as well. This is because we cannot duplicate external functions
+ // into multiple modules. To avoid duplicating accidentally, we
+ // conservatively put every external function in P0.
+ if (CurKernel.HasExternalCall) {
+ SML << "Kernel with external call(s): " << getName(*CurKernel.Fn)
+ << " defaulting to P0\n";
+ AssignToPartition(0, CurKernel);
+ continue;
+ }
+
+ // Be smart with large kernels to avoid duplicating their dependencies.
+ if (CurKernel.isLarge(LargeKernelThreshold)) {
+ assert(LargeKernelOverlapForMerge >= 0.0f &&
+ LargeKernelOverlapForMerge <= 1.0f);
+ SML << "Large Kernel: " << getName(*CurKernel.Fn)
+ << " - looking for partition with at least "
+ << format("%0.2f", LargeKernelOverlapForMerge * 100) << "% overlap\n";
+
+ bool Assigned = false;
+ for (const auto &[PID, Fns] : enumerate(Partitions)) {
+ float Overlap = calculateOverlap(CurKernel.Dependencies, Fns);
+ SML << " => " << format("%0.2f", Overlap * 100) << "% overlap with P"
+ << PID << "\n";
+ if (Overlap > LargeKernelOverlapForMerge) {
+ SML << " selecting P" << PID << "\n";
+ AssignToPartition(PID, CurKernel);
+ Assigned = true;
+ }
+ }
+
+ if (Assigned)
+ continue;
+ }
+
+ // Normal "load-balancing", assign to partition with least pressure.
+ auto [PID, CurCost] = BalancingQueue.back();
+ AssignToPartition(PID, CurKernel);
+ }
+
+ // Work is mostly done now, verify the partioning and add all functions we may
+ // have missed (= unreachable, or we don't understand how they're reached) to
+ // P0.
+ DenseSet<const Function *> AllFunctions;
+ for (const auto &[Idx, Part] : enumerate(Partitions)) {
+ [[maybe_unused]] CostType Cost = 0;
+ for (auto *Fn : Part) {
+ // external linkage functions should exclusively be in the first partition
+ // at this stage. In theory, we should only ever see external linkage
+ // functions here if they're kernels, or if they've been added due to a
+ // kernel using indirect calls somewhere in its CallGraph.
+ assert(Idx == 0 || (!Fn->hasExternalLinkage() || AMDGPU::isKernelCC(Fn)));
+ Cost += FnCosts.at(Fn);
+ }
+ SML << "P" << Idx << " has a total cost of " << Cost << " ("
+ << format("%0.2f", (float(Cost) / ModuleCost) * 100)
+ << "% of source module)\n";
+ AllFunctions.insert(Part.begin(), Part.end());
+ }
+
+ // Add missed functions to P0. This will take care of adding things like
+ // external functions with no callers in the module to P0. This should be
+ // fairly rare as AMDGPU internalizes everything in most cases, so unused
+ // internal functions would get removed.
+ for (auto &Fn : M) {
+ if (!Fn.isDeclaration() && !AllFunctions.contains(&Fn)) {
+ SML << getName(Fn) << " has no partition assigned, defaulting to P0\n";
+ Partitions[0].insert(&Fn);
+ }
+ }
+
+ SML << "--Partitioning Done--\n\n";
+
+ return Partitions;
+}
+
+static void externalize(GlobalValue &GV) {
+ if (GV.hasLocalLinkage()) {
+ GV.setLinkage(GlobalValue::ExternalLinkage);
+ GV.setVisibility(GlobalValue::HiddenVisibility);
+ }
+
+ // Unnamed entities must be named consistently between modules. setName will
+ // give a distinct name to each such entity.
+ if (!GV.hasName())
+ GV.setName("__llvmsplit_unnamed");
+}
+} // end anonymous namespace
+
+void llvm::splitAMDGPUModule(
+ const AMDGPUTargetMachine &TM, Module &M, unsigned N,
+ function_ref<void(std::unique_ptr<Module> MPart)> ModuleCallback) {
+
+ SplitModuleLogger SML(M);
+
+ CallGraph CG(M);
+
+ // Externalize functions whose address are taken.
+ //
+ // This is needed because partitioning is purely based on calls, but sometimes
+ // a kernel/function may just look at the address of another local function
+ // and not do anything (no calls). After partitioning, that local function may
+ // end up in a different module (so it's just a declaration in the module
+ // where its address is taken), which emits a "undefined hidden symbol" linker
+ // error.
+ //
+ // Additionally, it guides partitioning to not duplicate this function if it's
+ // called directly at some point.
+ for (auto &Fn : M) {
+ if (Fn.hasAddressTaken()) {
+ if (Fn.hasLocalLinkage()) {
+ SML << "[externalize] " << Fn.getName()
+ << " because its address is taken\n";
+ }
+ externalize(Fn);
+ }
+ }
+
+ // Externalize local GVs, which avoids duplicating their initializers, which
+ // in turns helps keep code size in check.
+ if (!NoExternalizeGlobals) {
+ for (auto &GV : M.globals()) {
+ if (GV.hasLocalLinkage())
+ SML << "[externalize] GV " << GV.getName() << "\n";
+ externalize(GV);
+ }
+ }
+
+ // Start by calculating the cost of every function in the module, as well as
+ // the module's overall cost.
+ DenseMap<const Function *, CostType> FnCosts;
+ const CostType ModuleCost = calculateFunctionCosts(SML, TM, M, FnCosts);
+
+ // Gather every kernel into a WorkList, then sort it by descending total cost
+ // of the kernel so the biggest kernels are seen first.
+ SmallVector<KernelWithDependencies> WorkList;
+ for (auto &Fn : M) {
+ if (AMDGPU::isKernelCC(&Fn) && !Fn.isDeclaration())
+ WorkList.emplace_back(SML, CG, FnCosts, &Fn);
+ }
+ sort(WorkList, [&](auto &A, auto &B) {
+ // Sort by total cost, and if the total cost is identical, sort
+ // alphabetically.
+ if (A.TotalCost == B.TotalCost)
+ return A.Fn->getName() < B.Fn->getName();
+ return A.TotalCost > B.TotalCost;
+ });
+
+ if (SML) {
+ SML << "Worklist\n";
+ for (const auto &KWD : WorkList) {
+ SML << "[Kernel] " << getName(*KWD.Fn) << " (totalCost:" << KWD.TotalCost
+ << " indirect:" << KWD.HasIndirectCall
+ << " external:" << KWD.HasExternalCall << ")\n";
+ for (const auto *Dep : KWD.Dependencies)
+ SML << " [Dep] " << getName(*Dep) << "\n";
+ }
+ }
+
+ // This performs all of the partitioning work.
+ auto Partitions = doPartitioning(SML, M, N, ModuleCost, FnCosts, WorkList);
+ assert(Partitions.size() == N);
+
+ // If we didn't externalize GVs, then local GVs need to be conservatively
+ // imported into every module (including their initializers), and then cleaned
+ // up afterwards.
+ const auto NeedsConservativeImport = [&](const GlobalValue *GV) {
+ // We conservatively import private/internal GVs into every module and clean
+ // them up afterwards.
+ const auto *Var = dyn_cast<GlobalVariable>(GV);
+ return Var && Var->hasLocalLinkage();
+ };
+
+ SML << "Creating " << N << " modules...\n";
+ unsigned TotalFnImpls = 0;
+ for (unsigned I = 0; I < N; ++I) {
+ const auto &FnsInPart = Partitions[I];
+
+ ValueToValueMapTy VMap;
+ std::unique_ptr<Module> MPart(
+ CloneModule(M, VMap, [&](const GlobalValue *GV) {
+ // Functions go in their assigned partition.
+ if (const auto *Fn = dyn_cast<Function>(GV)) {
+// Check we don't import an external linkage function in any
+// partition other than P0.
+#ifndef NDEBUG
+ if (Fn->hasExternalLinkage() && !AMDGPU::isKernelCC(Fn)) {
+ assert((I == 0) == FnsInPart.contains(Fn));
+ }
+#endif
+ return FnsInPart.contains(Fn);
+ }
+
+ if (NeedsConservativeImport(GV))
+ return true;
+
+ // Everything else goes in the first partition.
+ return I == 0;
+ }));
+ if (I != 0)
+ MPart->setModuleInlineAsm("");
+
+ // Clean-up conservatively imported GVs without any users.
+ for (auto &GV : make_early_inc_range(MPart->globals())) {
+ if (NeedsConservativeImport(&GV) && GV.use_empty())
+ GV.eraseFromParent();
+ }
+
+ unsigned NumAllFns = 0, NumKernels = 0;
+ for (auto &Cur : *MPart) {
+ if (!Cur.isDeclaration()) {
+ ++NumAllFns;
+ if (AMDGPU::isKernelCC(&Cur))
+ ++NumKernels;
+ }
+ }
+ TotalFnImpls += NumAllFns;
+ SML << " - Module " << I << " with " << NumAllFns << " functions ("
+ << NumKernels << " kernels)\n";
+ ModuleCallback(std::move(MPart));
+ }
+
+ SML << TotalFnImpls << " function definitions across all modules ("
+ << format("%0.2f", (float(TotalFnImpls) / FnCosts.size()) * 100)
+ << "% of original module)\n";
+}
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSplitModule.h b/llvm/lib/Target/AMDGPU/AMDGPUSplitModule.h
new file mode 100644
index 00000000000000..6171643bd4adc4
--- /dev/null
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSplitModule.h
@@ -0,0 +1,30 @@
+//===- AMDGPUSplitModule.h -------------------------------------*- C++ -*-===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_TARGET_AMDGPUSPLITMODULE_H
+#define LLVM_TARGET_AMDGPUSPLITMODULE_H
+
+#include "llvm/ADT/STLFunctionalExtras.h"
+#include <memory>
+
+namespace llvm {
+
+class Module;
+class AMDGPUTargetMachine;
+
+/// Splits the module M into N linkable partitions. The function ModuleCallback
+/// is called N times passing each individual partition as the MPart argument.
+void splitAMDGPUModule(
+ const AMDGPUTargetMachine &TM, Module &M, unsigned N,
+ function_ref<void(std::unique_ptr<Module> MPart)> ModuleCallback);
+
+} // end namespace llvm
+
+#endif // LLVM_TARGET_AMDGPUSPLITMODULE_H
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index 305a6c8c3b9262..72d77f0a2f7067 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -20,6 +20,7 @@
#include "AMDGPUIGroupLP.h"
#include "AMDGPUMacroFusion.h"
#include "AMDGPURegBankSelect.h"
+#include "AMDGPUSplitModule.h"
#include "AMDGPUTargetObjectFile.h"
#include "AMDGPUTargetTransformInfo.h"
#include "AMDGPUUnifyDivergentExitNodes.h"
@@ -806,6 +807,13 @@ AMDGPUTargetMachine::getAddressSpaceForPseudoSourceKind(unsigned Kind) const {
return AMDGPUAS::FLAT_ADDRESS;
}
+bool AMDGPUTargetMachine::splitModule(
+ Module &M, unsigned NumParts,
+ function_ref<void(std::unique_ptr<Module> MPart)> ModuleCallback) const {
+ splitAMDGPUModule(*this, M, NumParts, ModuleCallback);
+ return true;
+}
+
//===----------------------------------------------------------------------===//
// GCN Target Machine (SI+)
//===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.h b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.h
index 30ab388c7d52e2..5f6233d4378c83 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.h
@@ -67,6 +67,10 @@ class AMDGPUTargetMachine : public LLVMTargetMachine {
getPredicatedAddrSpace(const Value *V) const override;
unsigned getAddressSpaceForPseudoSourceKind(unsigned Kind) const override;
+
+ bool splitModule(Module &M, unsigned NumParts,
+ function_ref<void(std::unique_ptr<Module> MPart)>
+ ModuleCallback) const override;
};
//===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/AMDGPU/CMakeLists.txt b/llvm/lib/Target/AMDGPU/CMakeLists.txt
index 48325a0928f93d..d991cc5b1b79d5 100644
--- a/llvm/lib/Target/AMDGPU/CMakeLists.txt
+++ b/llvm/lib/Target/AMDGPU/CMakeLists.txt
@@ -97,6 +97,7 @@ add_llvm_target(AMDGPUCodeGen
AMDGPURewriteOutArguments.cpp
AMDGPURewriteUndefForPHI.cpp
AMDGPUSetWavePriority.cpp
+ AMDGPUSplitModule.cpp
AMDGPUSubtarget.cpp
AMDGPUTargetMachine.cpp
AMDGPUTargetObjectFile.cpp
diff --git a/llvm/test/tools/llvm-split/AMDGPU/address-taken-externalize-with-call.ll b/llvm/test/tools/llvm-split/AMDGPU/address-taken-externalize-with-call.ll
new file mode 100644
index 00000000000000..0e361df9597ca4
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/address-taken-externalize-with-call.ll
@@ -0,0 +1,46 @@
+; RUN: llvm-split -o %t %s -j 3 -mtarget amdgcn-amd-amdhsa -amdgpu-module-splitting-large-kernel-threshold=0
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+
+; 3 kernels:
+; - A does a direct call to HelperA
+; - B is storing @HelperA
+; - C does a direct call to HelperA
+;
+; The helper functions will get externalized, which will force A and C into P0 as
+; external functions cannot be duplicated.
+
+; CHECK0: define hidden void @HelperA()
+; CHECK0: define amdgpu_kernel void @A()
+; CHECK0: declare amdgpu_kernel void @B(ptr)
+; CHECK0: define amdgpu_kernel void @C()
+
+; CHECK1: declare hidden void @HelperA()
+; CHECK1: declare amdgpu_kernel void @A()
+; CHECK1: declare amdgpu_kernel void @B(ptr)
+; CHECK1: declare amdgpu_kernel void @C()
+
+; CHECK2: declare hidden void @HelperA()
+; CHECK2: declare amdgpu_kernel void @A()
+; CHECK2: define amdgpu_kernel void @B(ptr %dst)
+; CHECK2: declare amdgpu_kernel void @C()
+
+define internal void @HelperA() {
+ ret void
+}
+
+define amdgpu_kernel void @A() {
+ call void @HelperA()
+ ret void
+}
+
+define amdgpu_kernel void @B(ptr %dst) {
+ store ptr @HelperA, ptr %dst
+ ret void
+}
+
+define amdgpu_kernel void @C() {
+ call void @HelperA()
+ ret void
+}
diff --git a/llvm/test/tools/llvm-split/AMDGPU/address-taken-externalize.ll b/llvm/test/tools/llvm-split/AMDGPU/address-taken-externalize.ll
new file mode 100644
index 00000000000000..bc20d52c81015b
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/address-taken-externalize.ll
@@ -0,0 +1,37 @@
+; RUN: llvm-split -o %t %s -j 2 -mtarget amdgcn-amd-amdhsa -amdgpu-module-splitting-large-kernel-threshold=0
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+
+; 2 kernels:
+; - A is isolated
+; - B is storing @HelperA/B's address
+;
+; The helper functions should get externalized (become hidden w/ external linkage)
+
+; CHECK0: define hidden void @HelperA()
+; CHECK0: define hidden void @HelperB()
+; CHECK0: define amdgpu_kernel void @A()
+; CHECK0: declare amdgpu_kernel void @B(i1, ptr)
+
+; CHECK1: declare hidden void @HelperA()
+; CHECK1: declare hidden void @HelperB()
+; CHECK1: declare amdgpu_kernel void @A()
+; CHECK1: define amdgpu_kernel void @B(i1 %cond, ptr %dst)
+
+define internal void @HelperA() {
+ ret void
+}
+
+define internal void @HelperB() {
+ ret void
+}
+
+define amdgpu_kernel void @A() {
+ ret void
+}
+
+define amdgpu_kernel void @B(i1 %cond, ptr %dst) {
+ %addr = select i1 %cond, ptr @HelperA, ptr @HelperB
+ store ptr %addr, ptr %dst
+ ret void
+}
diff --git a/llvm/test/tools/llvm-split/AMDGPU/kernels-cost-ranking.ll b/llvm/test/tools/llvm-split/AMDGPU/kernels-cost-ranking.ll
new file mode 100644
index 00000000000000..f57c6e271556ea
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/kernels-cost-ranking.ll
@@ -0,0 +1,54 @@
+; RUN: llvm-split -o %t %s -j 3 -mtarget amdgcn-amd-amdhsa
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+
+; 3 kernels with each their own dependencies should go into 3
+; distinct partitions. The most expensive kernel should be
+; seen first and go into the last partition.
+
+; CHECK0-NOT: define
+; CHECK0: define amdgpu_kernel void @C
+; CHECK0: define internal void @HelperC
+; CHECK0-NOT: define
+
+; CHECK1-NOT: define
+; CHECK1: define amdgpu_kernel void @A
+; CHECK1: define internal void @HelperA
+; CHECK1-NOT: define
+
+; CHECK2-NOT: define
+; CHECK2: define amdgpu_kernel void @B
+; CHECK2: define internal void @HelperB
+; CHECK2-NOT: define
+
+
+define amdgpu_kernel void @A() {
+ call void @HelperA()
+ ret void
+}
+
+define internal void @HelperA() {
+ ret void
+}
+
+define amdgpu_kernel void @B() {
+ store i64 42, ptr undef
+ store i64 43, ptr undef
+ store i64 44, ptr undef
+ call void @HelperB()
+ ret void
+}
+
+define internal void @HelperB() {
+ ret void
+}
+
+define amdgpu_kernel void @C() {
+ call void @HelperC()
+ ret void
+}
+
+define internal void @HelperC() {
+ ret void
+}
diff --git a/llvm/test/tools/llvm-split/AMDGPU/kernels-dependencies.ll b/llvm/test/tools/llvm-split/AMDGPU/kernels-dependencies.ll
new file mode 100644
index 00000000000000..e501356a60aa0d
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/kernels-dependencies.ll
@@ -0,0 +1,50 @@
+; RUN: llvm-split -o %t %s -j 3 -mtarget amdgcn-amd-amdhsa
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+
+; 3 kernels with each their own dependencies should go into 3
+; distinct partitions.
+
+; CHECK0-NOT: define
+; CHECK0: define amdgpu_kernel void @C
+; CHECK0: define internal void @HelperC
+; CHECK0-NOT: define
+
+; CHECK1-NOT: define
+; CHECK1: define amdgpu_kernel void @B
+; CHECK1: define internal void @HelperB
+; CHECK1-NOT: define
+
+; CHECK2-NOT: define
+; CHECK2: define amdgpu_kernel void @A
+; CHECK2: define internal void @HelperA
+; CHECK2-NOT: define
+
+
+define amdgpu_kernel void @A() {
+ call void @HelperA()
+ ret void
+}
+
+define internal void @HelperA() {
+ ret void
+}
+
+define amdgpu_kernel void @B() {
+ call void @HelperB()
+ ret void
+}
+
+define internal void @HelperB() {
+ ret void
+}
+
+define amdgpu_kernel void @C() {
+ call void @HelperC()
+ ret void
+}
+
+define internal void @HelperC() {
+ ret void
+}
diff --git a/llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-duplication.ll b/llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-duplication.ll
new file mode 100644
index 00000000000000..6727af70a48783
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-duplication.ll
@@ -0,0 +1,41 @@
+; RUN: llvm-split -o %t %s -j 3 -mtarget amdgcn-amd-amdhsa
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+
+; 3 kernels share a common helper, that helper should be
+; cloned in all partitions.
+
+; CHECK0-NOT: define
+; CHECK0: define internal void @Helper
+; CHECK0: define amdgpu_kernel void @C
+; CHECK0-NOT: define
+
+; CHECK1-NOT: define
+; CHECK1: define internal void @Helper
+; CHECK1: define amdgpu_kernel void @B
+; CHECK1-NOT: define
+
+; CHECK2-NOT: define
+; CHECK2: define internal void @Helper
+; CHECK2: define amdgpu_kernel void @A
+; CHECK2-NOT: define
+
+define internal void @Helper() {
+ ret void
+}
+
+define amdgpu_kernel void @A() {
+ call void @Helper()
+ ret void
+}
+
+define amdgpu_kernel void @B() {
+ call void @Helper()
+ ret void
+}
+
+define amdgpu_kernel void @C() {
+ call void @Helper()
+ ret void
+}
diff --git a/llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-external.ll b/llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-external.ll
new file mode 100644
index 00000000000000..47fac7f1f7b16a
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-external.ll
@@ -0,0 +1,43 @@
+; RUN: llvm-split -o %t %s -j 3 -mtarget amdgcn-amd-amdhsa
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+
+; 3 kernels use private/internal global variables.
+; The GVs should be copied in each partition as needed.
+
+; CHECK0-NOT: define
+; CHECK0: define void @ExternalHelper
+; CHECK0: define amdgpu_kernel void @A
+; CHECK0: define amdgpu_kernel void @B
+; CHECK0-NOT: define
+
+; CHECK1-NOT: define
+; CHECK1: define amdgpu_kernel void @D
+; CHECK1-NOT: define
+
+; CHECK2-NOT: define
+; CHECK2: define amdgpu_kernel void @C
+; CHECK2-NOT: define
+
+define void @ExternalHelper() {
+ ret void
+}
+
+define amdgpu_kernel void @A() {
+ call void @ExternalHelper()
+ ret void
+}
+
+define amdgpu_kernel void @B() {
+ call void @ExternalHelper()
+ ret void
+}
+
+define amdgpu_kernel void @C() {
+ ret void
+}
+
+define amdgpu_kernel void @D() {
+ ret void
+}
diff --git a/llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-indirect.ll b/llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-indirect.ll
new file mode 100644
index 00000000000000..336677c33309ae
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/kernels-dependency-indirect.ll
@@ -0,0 +1,69 @@
+; RUN: llvm-split -o %t %s -j 3 -mtarget amdgcn-amd-amdhsa
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+
+; We have 4 kernels:
+; - Each kernel has an internal helper
+; - @A and @B's helpers does an indirect call.
+;
+; We default to putting A/B in P0, alongside a copy
+; of all helpers. The other kernels can still go into
+; separate partitions.
+
+; CHECK0-NOT: define
+; CHECK0: define internal void @HelperA
+; CHECK0: define internal void @HelperB
+; CHECK0: define internal void @HelperC
+; CHECK0: define internal void @HelperD
+; CHECK0: define amdgpu_kernel void @A
+; CHECK0: define amdgpu_kernel void @B
+; CHECK0-NOT: define
+
+; CHECK1-NOT: define
+; CHECK1: define internal void @HelperD
+; CHECK1: define amdgpu_kernel void @D
+; CHECK1-NOT: define
+
+; CHECK2-NOT: define
+; CHECK2: define internal void @HelperC
+; CHECK2: define amdgpu_kernel void @C
+; CHECK2-NOT: define
+
+define internal void @HelperA(ptr %call) {
+ call void %call()
+ ret void
+}
+
+define internal void @HelperB(ptr %call) {
+ call void %call()
+ ret void
+}
+
+define internal void @HelperC() {
+ ret void
+}
+
+define internal void @HelperD() {
+ ret void
+}
+
+define amdgpu_kernel void @A(ptr %call) {
+ call void @HelperA(ptr %call)
+ ret void
+}
+
+define amdgpu_kernel void @B(ptr %call) {
+ call void @HelperB(ptr %call)
+ ret void
+}
+
+define amdgpu_kernel void @C() {
+ call void @HelperC()
+ ret void
+}
+
+define amdgpu_kernel void @D() {
+ call void @HelperD()
+ ret void
+}
diff --git a/llvm/test/tools/llvm-split/AMDGPU/kernels-global-variables-noexternal.ll b/llvm/test/tools/llvm-split/AMDGPU/kernels-global-variables-noexternal.ll
new file mode 100644
index 00000000000000..406f23150db03e
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/kernels-global-variables-noexternal.ll
@@ -0,0 +1,42 @@
+; RUN: llvm-split -o %t %s -j 3 -mtarget amdgcn-amd-amdhsa -amdgpu-module-splitting-no-externalize-globals
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+
+; 3 kernels use private/internal global variables.
+; The GVs should be copied in each partition as needed.
+
+; CHECK0-NOT: define
+; CHECK0: @bar = internal constant ptr
+; CHECK0: define amdgpu_kernel void @C
+; CHECK0-NOT: define
+
+; CHECK1-NOT: define
+; CHECK1: @foo = private constant ptr
+; CHECK1: define amdgpu_kernel void @A
+; CHECK1-NOT: define
+
+; CHECK2-NOT: define
+; CHECK2: @foo = private constant ptr
+; CHECK2: @bar = internal constant ptr
+; CHECK2: define amdgpu_kernel void @B
+; CHECK2-NOT: define
+
+ at foo = private constant ptr undef
+ at bar = internal constant ptr undef
+
+define amdgpu_kernel void @A() {
+ store i32 42, ptr @foo
+ ret void
+}
+
+define amdgpu_kernel void @B() {
+ store i32 42, ptr @foo
+ store i32 42, ptr @bar
+ ret void
+}
+
+define amdgpu_kernel void @C() {
+ store i32 42, ptr @bar
+ ret void
+}
diff --git a/llvm/test/tools/llvm-split/AMDGPU/kernels-global-variables.ll b/llvm/test/tools/llvm-split/AMDGPU/kernels-global-variables.ll
new file mode 100644
index 00000000000000..3f7960fa3c381e
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/kernels-global-variables.ll
@@ -0,0 +1,44 @@
+; RUN: llvm-split -o %t %s -j 3 -mtarget amdgcn-amd-amdhsa
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+
+; 3 kernels use private/internal global variables.
+; The GVs should be copied in each partition as needed.
+
+; CHECK0-NOT: define
+; CHECK0: @foo = hidden constant ptr undef
+; CHECK0: @bar = hidden constant ptr undef
+; CHECK0: define amdgpu_kernel void @C
+; CHECK0-NOT: define
+
+; CHECK1-NOT: define
+; CHECK1: @foo = external hidden constant ptr{{$}}
+; CHECK1: @bar = external hidden constant ptr{{$}}
+; CHECK1: define amdgpu_kernel void @A
+; CHECK1-NOT: define
+
+; CHECK2-NOT: define
+; CHECK2: @foo = external hidden constant ptr{{$}}
+; CHECK2: @bar = external hidden constant ptr{{$}}
+; CHECK2: define amdgpu_kernel void @B
+; CHECK2-NOT: define
+
+ at foo = private constant ptr undef
+ at bar = internal constant ptr undef
+
+define amdgpu_kernel void @A() {
+ store i32 42, ptr @foo
+ ret void
+}
+
+define amdgpu_kernel void @B() {
+ store i32 42, ptr @foo
+ store i32 42, ptr @bar
+ ret void
+}
+
+define amdgpu_kernel void @C() {
+ store i32 42, ptr @bar
+ ret void
+}
diff --git a/llvm/test/tools/llvm-split/AMDGPU/kernels-load-balancing.ll b/llvm/test/tools/llvm-split/AMDGPU/kernels-load-balancing.ll
new file mode 100644
index 00000000000000..ee4e9583fc937d
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/kernels-load-balancing.ll
@@ -0,0 +1,75 @@
+; RUN: llvm-split -o %t %s -j 3 -mtarget amdgcn-amd-amdhsa
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+
+; Test load balancing logic with 6 kernels.
+;
+; Kernels go from most expensive (A == 6) to least expensive (F == 1)
+;
+; Load balancing should work like this (current partition cost is in parens)
+;
+; Initial -> [P0(0), P1(0), P2(0)]
+;
+; A(6) goes in 2 -> [P2(6), P0(0), P1(0)]
+; B(5) goes in 1 -> [P2(6), P1(5), P0(4)]
+; C(4) goes in 0 -> [P2(6), P1(5), P0(4)]
+
+; D(3) goes in 0 -> [P0(7), P2(6), P1(5)]
+; E(2) goes in 1 -> [P0(7), P1(7), P2(6)]
+; F(1) goes in 2 -> [P0(7), P1(7), P2(7)]
+
+; CHECK0-NOT: define
+; CHECK0: define amdgpu_kernel void @C
+; CHECK0: define amdgpu_kernel void @D
+; CHECK0-NOT: define
+
+; CHECK1-NOT: define
+; CHECK1: define amdgpu_kernel void @B
+; CHECK1: define amdgpu_kernel void @E
+; CHECK1-NOT: define
+
+; CHECK2-NOT: define
+; CHECK2: define amdgpu_kernel void @A
+; CHECK2: define amdgpu_kernel void @F
+; CHECK2-NOT: define
+
+
+define amdgpu_kernel void @A() {
+ store i64 42, ptr undef
+ store i64 43, ptr undef
+ store i64 44, ptr undef
+ store i64 45, ptr undef
+ store i64 46, ptr undef
+ ret void
+}
+
+define amdgpu_kernel void @B() {
+ store i64 42, ptr undef
+ store i64 43, ptr undef
+ store i64 44, ptr undef
+ store i64 45, ptr undef
+ ret void
+}
+
+define amdgpu_kernel void @C() {
+ store i64 42, ptr undef
+ store i64 43, ptr undef
+ store i64 44, ptr undef
+ ret void
+}
+
+define amdgpu_kernel void @D() {
+ store i64 42, ptr undef
+ store i64 43, ptr undef
+ ret void
+}
+
+define amdgpu_kernel void @E() {
+ store i64 42, ptr undef
+ ret void
+}
+
+define amdgpu_kernel void @F() {
+ ret void
+}
diff --git a/llvm/test/tools/llvm-split/AMDGPU/kernels-no-dependencies.ll b/llvm/test/tools/llvm-split/AMDGPU/kernels-no-dependencies.ll
new file mode 100644
index 00000000000000..720e73a8dbe7c2
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/kernels-no-dependencies.ll
@@ -0,0 +1,39 @@
+; RUN: llvm-split -o %t %s -j 4 -mtarget amdgcn-amd-amdhsa
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+; RUN: llvm-dis -o - %t3 | FileCheck --check-prefix=CHECK3 %s
+
+; Check that 4 independent kernels get put into 4 different partitions.
+
+; CHECK0-NOT: define
+; CHECK0: define amdgpu_kernel void @D
+; CHECK0-NOT: define
+
+; CHECK1-NOT: define
+; CHECK1: define amdgpu_kernel void @C
+; CHECK1-NOT: define
+
+; CHECK2-NOT: define
+; CHECK2: define amdgpu_kernel void @B
+; CHECK2-NOT: define
+
+; CHECK3-NOT: define
+; CHECK3: define amdgpu_kernel void @A
+; CHECK3-NOT: define
+
+define amdgpu_kernel void @A() {
+ ret void
+}
+
+define amdgpu_kernel void @B() {
+ ret void
+}
+
+define amdgpu_kernel void @C() {
+ ret void
+}
+
+define amdgpu_kernel void @D() {
+ ret void
+}
diff --git a/llvm/test/tools/llvm-split/AMDGPU/large-kernels-merging.ll b/llvm/test/tools/llvm-split/AMDGPU/large-kernels-merging.ll
new file mode 100644
index 00000000000000..a3fddab73a14ca
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/large-kernels-merging.ll
@@ -0,0 +1,98 @@
+; RUN: llvm-split -o %t %s -j 3 -mtarget amdgcn-amd-amdhsa -amdgpu-module-splitting-large-kernel-threshold=1.2 -amdgpu-module-splitting-large-kernel-merge-overlap=0.5
+; RUN: llvm-dis -o - %t0 | FileCheck --check-prefix=CHECK0 %s
+; RUN: llvm-dis -o - %t1 | FileCheck --check-prefix=CHECK1 %s
+; RUN: llvm-dis -o - %t2 | FileCheck --check-prefix=CHECK2 %s
+
+; RUN: llvm-split -o %t.nolarge %s -j 3 -mtarget amdgcn-amd-amdhsa -amdgpu-module-splitting-large-kernel-threshold=0
+; RUN: llvm-dis -o - %t.nolarge0 | FileCheck --check-prefix=NOLARGEKERNELS-CHECK0 %s
+; RUN: llvm-dis -o - %t.nolarge1 | FileCheck --check-prefix=NOLARGEKERNELS-CHECK1 %s
+; RUN: llvm-dis -o - %t.nolarge2 | FileCheck --check-prefix=NOLARGEKERNELS-CHECK2 %s
+
+; 2 kernels (A/B) are large and share all their dependencies.
+; They should go in the same partition, the remaining kernel should
+; go somewhere else, and one partition should be empty.
+;
+; Also check w/o large kernels processing to verify they are indeed handled
+; differently.
+
+; CHECK0-NOT: define
+
+; CHECK1-NOT: define
+; CHECK1: define internal void @HelperC()
+; CHECK1: define amdgpu_kernel void @C
+; CHECK1-NOT: define
+
+; CHECK2-NOT: define
+; CHECK2: define internal void @large2()
+; CHECK2: define internal void @large1()
+; CHECK2: define internal void @large0()
+; CHECK2: define internal void @HelperA()
+; CHECK2: define internal void @HelperB()
+; CHECK2: define amdgpu_kernel void @A
+; CHECK2: define amdgpu_kernel void @B
+; CHECK2-NOT: define
+
+; NOLARGEKERNELS-CHECK0-NOT: define
+; NOLARGEKERNELS-CHECK0: define internal void @HelperC()
+; NOLARGEKERNELS-CHECK0: define amdgpu_kernel void @C
+; NOLARGEKERNELS-CHECK0-NOT: define
+
+; NOLARGEKERNELS-CHECK1: define internal void @large2()
+; NOLARGEKERNELS-CHECK1: define internal void @large1()
+; NOLARGEKERNELS-CHECK1: define internal void @large0()
+; NOLARGEKERNELS-CHECK1: define internal void @HelperB()
+; NOLARGEKERNELS-CHECK1: define amdgpu_kernel void @B
+
+; NOLARGEKERNELS-CHECK2: define internal void @large2()
+; NOLARGEKERNELS-CHECK2: define internal void @large1()
+; NOLARGEKERNELS-CHECK2: define internal void @large0()
+; NOLARGEKERNELS-CHECK2: define internal void @HelperA()
+; NOLARGEKERNELS-CHECK2: define amdgpu_kernel void @A
+
+define internal void @large2() {
+ store i32 42, ptr undef
+ call void @large2()
+ ret void
+}
+
+define internal void @large1() {
+ call void @large1()
+ call void @large2()
+ ret void
+}
+
+define internal void @large0() {
+ call void @large0()
+ call void @large1()
+ call void @large2()
+ ret void
+}
+
+define internal void @HelperA() {
+ call void @large0()
+ ret void
+}
+
+define internal void @HelperB() {
+ call void @large0()
+ ret void
+}
+
+define amdgpu_kernel void @A() {
+ call void @HelperA()
+ ret void
+}
+
+define amdgpu_kernel void @B() {
+ call void @HelperB()
+ ret void
+}
+
+define internal void @HelperC() {
+ ret void
+}
+
+define amdgpu_kernel void @C() {
+ call void @HelperC()
+ ret void
+}
diff --git a/llvm/test/tools/llvm-split/AMDGPU/lit.local.cfg b/llvm/test/tools/llvm-split/AMDGPU/lit.local.cfg
new file mode 100644
index 00000000000000..7c492428aec761
--- /dev/null
+++ b/llvm/test/tools/llvm-split/AMDGPU/lit.local.cfg
@@ -0,0 +1,2 @@
+if not "AMDGPU" in config.root.targets:
+ config.unsupported = True
More information about the llvm-commits
mailing list