[AMDGPU] Teach iterative schedulers about IGLP (llvm#134953)

jrbyrnes · jrbyrnes · commit ba0142e0bf3b · 2025-04-13T12:46:25.000-07:00
This adds IGLP mutation to the iterative schedulers
(`gcn-iterative-max-occupancy-experimental`, `gcn-iterative-minreg`, and
`gcn-iterative-ilp`).

The `gcn-iterative-minreg` and `gcn-iterative-ilp` schedulers never
actually applied the mutations added, so this also has the effect of
teaching them about mutations in general. The
`gcn-iterative-max-occupancy-experimental` scheduler has calls to
`ScheduleDAGMILive::schedule()`, so, before this, mutations were applied
at this point. Now this is done during calls to `BuildDAG`, with IGLP
superseding other mutations (similar to the other schedulers). We may
end up scheduling regions multiple times, with mutations being applied
each time, so we need to track for
`AMDGPU::SchedulingPhase::PreRAReentry`
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -563,12 +563,15 @@ createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
   if (ST.shouldClusterStores())
     DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
+  DAG->addMutation(createIGroupLPDAGMutation(AMDGPU::SchedulingPhase::Initial));
   return DAG;
 }
 
 static ScheduleDAGInstrs *createMinRegScheduler(MachineSchedContext *C) {
-  return new GCNIterativeScheduler(C,
-    GCNIterativeScheduler::SCHEDULE_MINREGFORCED);
+  auto *DAG = new GCNIterativeScheduler(
+      C, GCNIterativeScheduler::SCHEDULE_MINREGFORCED);
+  DAG->addMutation(createIGroupLPDAGMutation(AMDGPU::SchedulingPhase::Initial));
+  return DAG;
 }
 
 static ScheduleDAGInstrs *
@@ -580,6 +583,7 @@ createIterativeILPMachineScheduler(MachineSchedContext *C) {
   if (ST.shouldClusterStores())
     DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
   DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
+  DAG->addMutation(createIGroupLPDAGMutation(AMDGPU::SchedulingPhase::Initial));
   return DAG;
 }
 
diff --git a/llvm/lib/Target/AMDGPU/GCNIterativeScheduler.cpp b/llvm/lib/Target/AMDGPU/GCNIterativeScheduler.cpp
@@ -12,6 +12,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "GCNIterativeScheduler.h"
+#include "AMDGPUIGroupLP.h"
 #include "GCNSchedStrategy.h"
 #include "SIMachineFunctionInfo.h"
 
@@ -118,21 +119,42 @@ void GCNIterativeScheduler::printSchedRP(raw_ostream &OS,
 }
 #endif
 
+void GCNIterativeScheduler::swapIGLPMutations(const Region &R, bool IsReentry) {
+  bool HasIGLPInstrs = false;
+  const SIInstrInfo *SII = static_cast<const SIInstrInfo *>(TII);
+  for (MachineBasicBlock::iterator I = R.Begin; I != R.End; I++) {
+    if (SII->isIGLPMutationOnly(I->getOpcode())) {
+      HasIGLPInstrs = true;
+      break;
+    }
+  }
+
+  if (HasIGLPInstrs) {
+    SavedMutations.clear();
+    SavedMutations.swap(Mutations);
+    auto SchedPhase = IsReentry ? AMDGPU::SchedulingPhase::PreRAReentry
+                                : AMDGPU::SchedulingPhase::Initial;
+
+    addMutation(createIGroupLPDAGMutation(SchedPhase));
+  }
+}
+
 // DAG builder helper
 class GCNIterativeScheduler::BuildDAG {
   GCNIterativeScheduler &Sch;
   SmallVector<SUnit *, 8> TopRoots;
 
   SmallVector<SUnit*, 8> BotRoots;
 public:
-  BuildDAG(const Region &R, GCNIterativeScheduler &_Sch)
-    : Sch(_Sch) {
-    auto BB = R.Begin->getParent();
+  BuildDAG(const Region &R, GCNIterativeScheduler &_Sch, bool IsReentry = false)
+      : Sch(_Sch) {
+    auto *BB = R.Begin->getParent();
     Sch.BaseClass::startBlock(BB);
     Sch.BaseClass::enterRegion(BB, R.Begin, R.End, R.NumRegionInstrs);
-
+    Sch.swapIGLPMutations(R, IsReentry);
     Sch.buildSchedGraph(Sch.AA, nullptr, nullptr, nullptr,
                         /*TrackLaneMask*/true);
+    Sch.postProcessDAG();
     Sch.Topo.InitDAGTopologicalSorting();
     Sch.findRootsAndBiasEdges(TopRoots, BotRoots);
   }
@@ -432,13 +454,15 @@ unsigned GCNIterativeScheduler::tryMaximizeOccupancy(unsigned TargetOcc) {
 
   auto NewOcc = TargetOcc;
   for (auto *R : Regions) {
+    // Always build the DAG to add mutations
+    BuildDAG DAG(*R, *this);
+
     if (R->MaxPressure.getOccupancy(ST) >= NewOcc)
-      break;
+      continue;
 
     LLVM_DEBUG(printRegion(dbgs(), R->Begin, R->End, LIS, 3);
                printLivenessInfo(dbgs(), R->Begin, R->End, LIS));
 
-    BuildDAG DAG(*R, *this);
     const auto MinSchedule = makeMinRegSchedule(DAG.getTopRoots(), *this);
     const auto MaxRP = getSchedulePressure(*R, MinSchedule);
     LLVM_DEBUG(dbgs() << "Occupancy improvement attempt:\n";
@@ -469,8 +493,11 @@ void GCNIterativeScheduler::scheduleLegacyMaxOccupancy(
   sortRegionsByPressure(TgtOcc);
   auto Occ = Regions.front()->MaxPressure.getOccupancy(ST);
 
-  if (TryMaximizeOccupancy && Occ < TgtOcc)
+  bool IsReentry = false;
+  if (TryMaximizeOccupancy && Occ < TgtOcc) {
     Occ = tryMaximizeOccupancy(TgtOcc);
+    IsReentry = true;
+  }
 
   // This is really weird but for some magic scheduling regions twice
   // gives performance improvement
@@ -489,7 +516,8 @@ void GCNIterativeScheduler::scheduleLegacyMaxOccupancy(
     LStrgy.setTargetOccupancy(I == 0 ? 0 : TgtOcc);
     for (auto *R : Regions) {
       OverrideLegacyStrategy Ovr(*R, LStrgy, *this);
-
+      IsReentry |= I > 0;
+      swapIGLPMutations(*R, IsReentry);
       Ovr.schedule();
       const auto RP = getRegionPressure(*R);
       LLVM_DEBUG(printSchedRP(dbgs(), R->MaxPressure, RP));
@@ -556,8 +584,11 @@ void GCNIterativeScheduler::scheduleILP(
   sortRegionsByPressure(TgtOcc);
   auto Occ = Regions.front()->MaxPressure.getOccupancy(ST);
 
-  if (TryMaximizeOccupancy && Occ < TgtOcc)
+  bool IsReentry = false;
+  if (TryMaximizeOccupancy && Occ < TgtOcc) {
     Occ = tryMaximizeOccupancy(TgtOcc);
+    IsReentry = true;
+  }
 
   TgtOcc = std::min(Occ, TgtOcc);
   LLVM_DEBUG(dbgs() << "Scheduling using default scheduler, "
@@ -566,7 +597,7 @@ void GCNIterativeScheduler::scheduleILP(
 
   unsigned FinalOccupancy = std::min(Occ, MFI->getOccupancy());
   for (auto *R : Regions) {
-    BuildDAG DAG(*R, *this);
+    BuildDAG DAG(*R, *this, IsReentry);
     const auto ILPSchedule = makeGCNILPScheduler(DAG.getBottomRoots(), *this);
 
     const auto RP = getSchedulePressure(*R, ILPSchedule);
diff --git a/llvm/lib/Target/AMDGPU/GCNIterativeScheduler.h b/llvm/lib/Target/AMDGPU/GCNIterativeScheduler.h
@@ -77,6 +77,8 @@ class GCNIterativeScheduler : public ScheduleDAGMILive {
   const StrategyKind Strategy;
   mutable GCNUpwardRPTracker UPTracker;
 
+  std::vector<std::unique_ptr<ScheduleDAGMutation>> SavedMutations;
+
   class BuildDAG;
   class OverrideLegacyStrategy;
 
@@ -91,6 +93,7 @@ class GCNIterativeScheduler : public ScheduleDAGMILive {
     return getRegionPressure(R.Begin, R.End);
   }
 
+  void swapIGLPMutations(const Region &R, bool IsReentry);
   void setBestSchedule(Region &R,
                        ScheduleRef Schedule,
                        const GCNRegPressure &MaxRP = GCNRegPressure());
diff --git a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
@@ -188,12 +188,6 @@ static void getRegisterPressures(
   Pressure[AMDGPU::RegisterPressureSets::AGPR_32] = NewPressure.getAGPRNum();
 }
 
-// Return true if the instruction is mutually exclusive with all non-IGLP DAG
-// mutations, requiring all other mutations to be disabled.
-static bool isIGLPMutationOnly(unsigned Opcode) {
-  return Opcode == AMDGPU::SCHED_GROUP_BARRIER || Opcode == AMDGPU::IGLP_OPT;
-}
-
 void GCNSchedStrategy::initCandidate(SchedCandidate &Cand, SUnit *SU,
                                      bool AtTop,
                                      const RegPressureTracker &RPTracker,
@@ -1018,9 +1012,10 @@ bool GCNSchedStage::initGCNRegion() {
   Unsched.reserve(DAG.NumRegionInstrs);
   if (StageID == GCNSchedStageID::OccInitialSchedule ||
       StageID == GCNSchedStageID::ILPInitialSchedule) {
+    const SIInstrInfo *SII = static_cast<const SIInstrInfo *>(DAG.TII);
     for (auto &I : DAG) {
       Unsched.push_back(&I);
-      if (isIGLPMutationOnly(I.getOpcode()))
+      if (SII->isIGLPMutationOnly(I.getOpcode()))
         DAG.RegionsWithIGLPInstrs[RegionIdx] = true;
     }
   } else {
@@ -1754,8 +1749,9 @@ void GCNScheduleDAGMILive::updateRegionBoundaries(
 }
 
 static bool hasIGLPInstrs(ScheduleDAGInstrs *DAG) {
-  return any_of(*DAG, [](MachineBasicBlock::iterator MI) {
-    return isIGLPMutationOnly(MI->getOpcode());
+  const SIInstrInfo *SII = static_cast<const SIInstrInfo *>(DAG->TII);
+  return any_of(*DAG, [SII](MachineBasicBlock::iterator MI) {
+    return SII->isIGLPMutationOnly(MI->getOpcode());
   });
 }
 
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.h b/llvm/lib/Target/AMDGPU/SIInstrInfo.h
@@ -986,6 +986,12 @@ class SIInstrInfo final : public AMDGPUGenInstrInfo {
 
   bool isIGLP(const MachineInstr &MI) const { return isIGLP(MI.getOpcode()); }
 
+  // Return true if the instruction is mutually exclusive with all non-IGLP DAG
+  // mutations, requiring all other mutations to be disabled.
+  bool isIGLPMutationOnly(unsigned Opcode) const {
+    return Opcode == AMDGPU::SCHED_GROUP_BARRIER || Opcode == AMDGPU::IGLP_OPT;
+  }
+
   static unsigned getNonSoftWaitcntOpcode(unsigned Opcode) {
     switch (Opcode) {
     case AMDGPU::S_WAITCNT_soft:
diff --git a/llvm/test/CodeGen/AMDGPU/iglp.opt.reentry.ll b/llvm/test/CodeGen/AMDGPU/iglp.opt.reentry.ll
@@ -1,4 +1,6 @@
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx908 -O3 < %s | FileCheck %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx908 -O3 -misched=gcn-iterative-max-occupancy-experimental < %s | FileCheck %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx908 -O3 -misched=gcn-iterative-ilp < %s | FileCheck %s
 
 ; Test should not result in build failure
 ; CHECK-LABEL: shouldNotReApply
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.sched.group.barrier.iterative.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.sched.group.barrier.iterative.ll
diff --git a/llvm/test/CodeGen/AMDGPU/schedule-regpressure-limit2.ll b/llvm/test/CodeGen/AMDGPU/schedule-regpressure-limit2.ll
