[llvm] AMDGPU: Add NextUseAnalysis Pass (PR #178873)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Thu Apr 16 10:23:00 PDT 2026
================
@@ -0,0 +1,2614 @@
+//===---------------------- AMDGPUNextUseAnalysis.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 file implements the AMDGPUNextUseAnalysis pass, a machine-level analysis
+// that computes the distance from each instruction to the "nearest" next use of
+// every live virtual register. These distances guide register spilling
+// decisions by identifying which live values are furthest from their next use
+// and are therefore the best candidates to spill.
+//
+// The analysis is based on the Braun & Hack CC'09 paper "Register Spilling and
+// Live-Range Splitting for SSA-Form Programs."
+//
+// Key concepts:
+//
+// NextUseDistance A loop-depth-weighted instruction count representing
+// how far away a register's next use is. Distances
+// through deeper loops are scaled by fromLoopDepth() so
+// that uses inside hot loops appear closer.
+//
+// Inter-block Pre-computed shortest weighted distances between all
+// distances pairs of basic blocks, used to efficiently answer
+// cross-block next-use queries. Each intermediate block
+// is weighted by fromLoopDepth() applied once per loop
+// boundary crossing relative to the destination.
+//
+// Configuration flags (see Config struct in the header):
+//
+// CountPhis Count PHI instructions toward distance and block size.
+// ForwardOnly Restrict inter-block distances to forward-reachable
+// paths.
+// PreciseUseModeling Model PHI uses at their incoming edge block and filter
+// uses with intermediate redefinitions.
+// PromoteToPreheader Route loop-entry and inner-loop uses to the preheader.
+//
+// This file contains:
+//
+// - Command-line options for configuration and debug output
+// - LiveRegUse / JSON helpers
+// - AMDGPUNextUseAnalysisImpl (the main analysis implementation)
+// - Instruction ID assignment and block size computation
+// - CFG path pre-computation (reachability, loop depth, back-edges)
+// - Inter-block distance computation
+// - Per-register next-use distance queries and caching
+// - AMDGPUNextUseAnalysis (public facade, pimpl)
+// - Legacy and new pass manager wrappers
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUNextUseAnalysis.h"
+#include "AMDGPU.h"
+#include "GCNRegPressure.h"
+#include "GCNSubtarget.h"
+
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/PostOrderIterator.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/IR/ModuleSlotTracker.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Support/FileSystem.h"
+#include "llvm/Support/JSON.h"
+#include "llvm/Support/Timer.h"
+#include "llvm/Support/ToolOutputFile.h"
+#include "llvm/Support/raw_ostream.h"
+
+#include <algorithm>
+#include <limits>
+#include <string>
+
+using namespace llvm;
+
+#define DEBUG_TYPE "amdgpu-next-use-analysis"
+
+//==============================================================================
+// Options etc
+//==============================================================================
+namespace {
+
+cl::opt<bool>
+ DistanceCacheEnabled("amdgpu-next-use-analysis-distance-cache",
+ cl::init(true), cl::Hidden,
+ cl::desc("Enable live-reg-use distance cache"));
+
+cl::opt<std::string>
+ DumpNextUseDistanceAsJson("amdgpu-next-use-analysis-dump-distance-as-json",
+ cl::Hidden);
+
+cl::opt<bool> DumpNextUseDistanceDefToUse(
+ "amdgpu-next-use-analysis-dump-distance-def-to-use", cl::init(false),
+ cl::Hidden);
+
+cl::opt<bool>
+ DumpNextUseDistanceVerbose("amdgpu-next-use-analysis-dump-distance-verbose",
+ cl::init(false), cl::Hidden);
+
+// 'graphics' and 'compute' modes arose due to initial competing implementations
+// of next-use analysis that emphasized different types of workloads. This
+// implementation is a compromise that combines aspects of both. Over time, the
+// hope is we will be able to remove some of these differences and settle on a
+// more unified implementation.
+cl::opt<std::string>
+ ConfigPresetOpt("amdgpu-next-use-analysis-config", cl::Hidden,
+ cl::init("graphics"),
+ cl::desc("Config preset: 'graphics' or 'compute'"));
+
+cl::opt<bool> ConfigCountPhisOpt(
+ "amdgpu-next-use-analysis-count-phis", cl::Hidden,
+ cl::desc("Count PHI instructions toward distance and block size"));
+cl::opt<bool> ConfigForwardOnlyOpt(
+ "amdgpu-next-use-analysis-forward-only", cl::Hidden,
+ cl::desc("Restrict inter-block distances to forward-reachable paths"));
+cl::opt<bool> ConfigPreciseUseModelingOpt(
+ "amdgpu-next-use-analysis-precise-use-modeling", cl::Hidden,
+ cl::desc("Model PHI uses via incoming edge block with loop-aware "
+ "reachability filtering"));
+cl::opt<bool> ConfigPromoteToPreheaderOpt(
+ "amdgpu-next-use-analysis-use-preheader-model", cl::Hidden,
+ cl::desc("Promote loop-entry and inner-loop uses to the loop preheader"));
+} // namespace
+
+//==============================================================================
+// LiveRegUse - Represents a live register use with its distance. Used for
+// tracking and sorting register uses by distance.
+//==============================================================================
+namespace {
+using UseDistancePair = AMDGPUNextUseAnalysis::UseDistancePair;
+struct LiveRegUse : public UseDistancePair {
+ // 'nullptr' indicates an unset/invalid state.
+ LiveRegUse() : UseDistancePair(nullptr, 0) {}
+ LiveRegUse(const MachineOperand *Use, NextUseDistance Dist)
+ : UseDistancePair(Use, Dist) {}
+ LiveRegUse(const UseDistancePair &P) : UseDistancePair(P) {}
+
+ bool isUnset() const { return Use == nullptr; }
+
+ Register getReg() const { return Use->getReg(); }
+ unsigned getSubReg() const { return Use->getSubReg(); }
+ LaneBitmask getLaneMask(const SIRegisterInfo *TRI) const {
+ return TRI->getSubRegIndexLaneMask(Use->getSubReg());
+ }
+
+ bool isCloserThan(const LiveRegUse &X) const {
+ if (Dist < X.Dist)
+ return true;
+
+ if (Dist > X.Dist)
+ return false;
+
+ if (Use == X.Use)
+ return false;
+
+ // Ugh. When !CountPhis, PHIs and the first non-PHI instruction have id
+ // 0. In this case, consider PHIs as less than the first non-PHI
+ // instruction.
+ const MachineInstr *ThisMI = Use->getParent();
+ const MachineInstr *XMI = X.Use->getParent();
+ const MachineBasicBlock *ThisMBB = ThisMI->getParent();
+ if (ThisMBB == XMI->getParent()) {
+ if (ThisMI->isPHI() && !XMI->isPHI() &&
+ XMI == &(*ThisMBB->getFirstNonPHI()))
+ return true;
+ }
+
+ // Ensure deterministic results
+ return X.getReg() < getReg();
+ }
+
+ void print(raw_ostream &OS, const TargetRegisterInfo *TRI = nullptr,
+ const MachineRegisterInfo *MRI = nullptr) const {
+ if (isUnset()) {
+ OS << "<unset>";
+ return;
+ }
+ Dist.print(OS);
+ OS << " [" << printReg(getReg(), TRI, getSubReg(), MRI) << "]";
+ }
+
+ LLVM_DUMP_METHOD void dump() const {
+ print(dbgs());
+ dbgs() << '\n';
+ }
+};
+
+inline bool updateClosest(LiveRegUse &Closest, const LiveRegUse &X) {
+ if (!Closest.Use || X.isCloserThan(Closest)) {
+ Closest = X;
+ return true;
+ }
+ return false;
+}
+
+inline bool updateFurthest(LiveRegUse &Furthest, const LiveRegUse &X) {
+ if (!Furthest.Use || Furthest.isCloserThan(X)) {
+ Furthest = X;
+ return true;
+ }
+ return false;
+}
+} // namespace
+
+//==============================================================================
+// JSON helpers
+//==============================================================================
+namespace {
+template <typename Lambda>
+void printStringAttr(json::OStream &J, const char *Name, Lambda L) {
+ J.attributeBegin(Name);
+ raw_ostream &OS = J.rawValueBegin();
+ OS << '"';
+ L(OS);
+ OS << '"';
+ J.rawValueEnd();
+ J.attributeEnd();
+}
+void printStringAttr(json::OStream &J, const char *Name, Printable P) {
+ printStringAttr(J, Name, [&](raw_ostream &OS) { OS << P; });
+}
+
+void printStringAttr(json::OStream &J, const char *Name, const MachineInstr &MI,
+ ModuleSlotTracker &MST) {
+ printStringAttr(J, Name, [&](raw_ostream &OS) {
+ MI.print(OS, MST,
+ /* IsStandalone */ false,
+ /* SkipOpers */ false,
+ /* SkipDebugLoc */ false,
+ /* AddNewLine ---> */ false,
+ /* TargetInstrInfo */ nullptr);
+ });
+}
+
+void printMBBNameAttr(json::OStream &J, const char *Name,
+ const MachineBasicBlock &MBB, ModuleSlotTracker &MST) {
+ printStringAttr(J, Name, [&](raw_ostream &OS) {
+ MBB.printName(OS, MachineBasicBlock::PrintNameIr, &MST);
+ });
+}
+
+template <typename NameLambda, typename ValueT>
+void printAttr(json::OStream &J, NameLambda NL, ValueT V) {
+ std::string Name;
+ raw_string_ostream NameOS(Name);
+ NL(NameOS);
+ J.attribute(NameOS.str(), V);
+}
+
+template <typename ValueT>
+void printAttr(json::OStream &J, const Printable &P, ValueT V) {
+ printAttr(J, [&](raw_ostream &OS) { OS << P; }, V);
+}
+
+} // namespace
+
+//==============================================================================
+// AMDGPUNextUseAnalysisImpl
+//==============================================================================
+class llvm::AMDGPUNextUseAnalysisImpl {
+public:
+ struct CacheableNextUseDistance {
+ bool IsInstrRelative;
+ NextUseDistance Distance;
+ };
+ static constexpr bool InstrRelative = true;
+ static constexpr bool InstrInvariant = false;
+
+private:
+ const MachineFunction *MF = nullptr;
+ const SIRegisterInfo *TRI = nullptr;
+ const SIInstrInfo *TII = nullptr;
+ const MachineLoopInfo *MLI = nullptr;
+ const MachineRegisterInfo *MRI = nullptr;
+
+ using InstrIdTy = unsigned;
+ using InstrToIdMap = DenseMap<const MachineInstr *, InstrIdTy>;
+ InstrToIdMap InstrToId;
+ AMDGPUNextUseAnalysis::Config Cfg;
+
+ void initializeTables() {
+ for (const MachineBasicBlock &BB : *MF)
+ calcInstrIds(&BB, InstrToId);
+ initializeCfgPaths();
+ initializeInterBlockDistances();
+ }
+
+ void clearTables() {
+ InstrToId.clear();
+ RegUseMap.clear();
+ Paths.clear();
+
+ resetDistanceCache();
+ }
+
+ //~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ // Instruction Ids
+ //~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+private:
+ unsigned sizeOf(const MachineInstr &MI) const {
+ // When !Cfg.CountPhis, PHIs do not contribute to distances/sizes since they
+ // generally don't result in the generation of a machine instruction.
+ // FIXME: Consider using MI.isPseudo() or maybe MI.isMetaInstruction().
+ return Cfg.CountPhis ? 1 : !MI.isPHI();
+ }
+
+ void calcInstrIds(const MachineBasicBlock *BB,
+ InstrToIdMap &MutableInstrToId) const {
+ InstrIdTy Id = 0;
+ for (auto &MI : BB->instrs()) {
----------------
arsenm wrote:
No auto
https://github.com/llvm/llvm-project/pull/178873
More information about the llvm-commits
mailing list