[MachinePipeliner] Improve loop carried dependence analysis

ytmukai · ytmukai · commit 6201e3bed667 · 2025-01-16T17:37:42.000Z
The previous implementation had false positive/negative cases in
analysis of loop carried dependence.

A case of missed dependencies is caused by incorrect analysis of
address increments. It is fixed by a strict analysis of recursive
definitions. See added test swp-carried-dep4.mir.

Excessive detection of dependence is corrected by improving the
formula for determining overlap of address ranges to be accessed.
See added test swp-carried-dep5.mir.
diff --git a/llvm/include/llvm/CodeGen/MachinePipeliner.h b/llvm/include/llvm/CodeGen/MachinePipeliner.h
@@ -390,6 +390,9 @@ class SwingSchedulerDAG : public ScheduleDAGInstrs {
 
   const SwingSchedulerDDG *getDDG() const { return DDG.get(); }
 
+  bool mayOverlapInLaterIter(const MachineInstr *BaseMI,
+                             const MachineInstr *OtherMI) const;
+
 private:
   void addLoopCarriedDependences(AAResults *AA);
   void updatePhiDependences();
@@ -409,7 +412,7 @@ class SwingSchedulerDAG : public ScheduleDAGInstrs {
   void computeNodeOrder(NodeSetType &NodeSets);
   void checkValidNodeOrder(const NodeSetType &Circuits) const;
   bool schedulePipeline(SMSchedule &Schedule);
-  bool computeDelta(MachineInstr &MI, unsigned &Delta) const;
+  bool computeDelta(const MachineInstr &MI, int &Delta) const;
   MachineInstr *findDefInLoop(Register Reg);
   bool canUseLastOffsetValue(MachineInstr *MI, unsigned &BasePos,
                              unsigned &OffsetPos, unsigned &NewBase,
diff --git a/llvm/lib/CodeGen/MachinePipeliner.cpp b/llvm/lib/CodeGen/MachinePipeliner.cpp
@@ -2523,9 +2523,104 @@ bool SwingSchedulerDAG::schedulePipeline(SMSchedule &Schedule) {
   return scheduleFound && Schedule.getMaxStageCount() > 0;
 }
 
+static Register findUniqueOperandDefinedInLoop(const MachineInstr &MI) {
+  const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
+  Register Result;
+  for (const MachineOperand &Use : MI.all_uses()) {
+    Register Reg = Use.getReg();
+    if (!Reg.isVirtual())
+      return Register();
+    if (MRI.getVRegDef(Reg)->getParent() != MI.getParent())
+      continue;
+    if (Result)
+      return Register();
+    Result = Reg;
+  }
+  return Result;
+}
+
+/// When Op is a value that is incremented recursively in a loop and there is a
+/// unique instruction that increments it, returns true and sets Value.
+static bool findLoopIncrementValue(const MachineOperand &Op, int &Value) {
+  if (!Op.isReg() || !Op.getReg().isVirtual())
+    return false;
+
+  Register OrgReg = Op.getReg();
+  Register CurReg = OrgReg;
+  const MachineBasicBlock *LoopBB = Op.getParent()->getParent();
+  const MachineRegisterInfo &MRI = LoopBB->getParent()->getRegInfo();
+
+  const TargetInstrInfo *TII =
+      LoopBB->getParent()->getSubtarget().getInstrInfo();
+  const TargetRegisterInfo *TRI =
+      LoopBB->getParent()->getSubtarget().getRegisterInfo();
+
+  MachineInstr *Phi = nullptr;
+  MachineInstr *Increment = nullptr;
+
+  // Traverse definitions until it reaches Op or an instruction that does not
+  // satisfy the condition.
+  // Acceptable example:
+  //   bb.0:
+  //     %0 = PHI %3, %bb.0, ...
+  //     %2 = ADD %0, Value
+  //     ... = LOAD %2(Op)
+  //     %3 = COPY %2
+  while (true) {
+    if (!CurReg.isValid() || !CurReg.isVirtual())
+      return false;
+    MachineInstr *Def = MRI.getVRegDef(CurReg);
+    if (Def->getParent() != LoopBB)
+      return false;
+
+    if (Def->isCopy()) {
+      // Ignore copy instructions unless they contain subregisters
+      if (Def->getOperand(0).getSubReg() || Def->getOperand(1).getSubReg())
+        return false;
+      CurReg = Def->getOperand(1).getReg();
+    } else if (Def->isPHI()) {
+      // There must be just one Phi
+      if (Phi)
+        return false;
+      Phi = Def;
+      CurReg = getLoopPhiReg(*Def, LoopBB);
+    } else if (TII->getIncrementValue(*Def, Value)) {
+      // Potentially a unique increment
+      if (Increment)
+        // Multiple increments exist
+        return false;
+
+      const MachineOperand *BaseOp;
+      int64_t Offset;
+      bool OffsetIsScalable;
+      if (TII->getMemOperandWithOffset(*Def, BaseOp, Offset, OffsetIsScalable,
+                                       TRI)) {
+        // Pre/post increment instruction
+        CurReg = BaseOp->getReg();
+      } else {
+        // If only one of the operands is defined within the loop, it is assumed
+        // to be an incremented value.
+        CurReg = findUniqueOperandDefinedInLoop(*Def);
+        if (!CurReg.isValid())
+          return false;
+      }
+      Increment = Def;
+    } else {
+      return false;
+    }
+    if (CurReg == OrgReg)
+      break;
+  }
+
+  if (!Phi || !Increment)
+    return false;
+
+  return true;
+}
+
 /// Return true if we can compute the amount the instruction changes
 /// during each iteration. Set Delta to the amount of the change.
-bool SwingSchedulerDAG::computeDelta(MachineInstr &MI, unsigned &Delta) const {
+bool SwingSchedulerDAG::computeDelta(const MachineInstr &MI, int &Delta) const {
   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
   const MachineOperand *BaseOp;
   int64_t Offset;
@@ -2540,24 +2635,7 @@ bool SwingSchedulerDAG::computeDelta(MachineInstr &MI, unsigned &Delta) const {
   if (!BaseOp->isReg())
     return false;
 
-  Register BaseReg = BaseOp->getReg();
-
-  MachineRegisterInfo &MRI = MF.getRegInfo();
-  // Check if there is a Phi. If so, get the definition in the loop.
-  MachineInstr *BaseDef = MRI.getVRegDef(BaseReg);
-  if (BaseDef && BaseDef->isPHI()) {
-    BaseReg = getLoopPhiReg(*BaseDef, MI.getParent());
-    BaseDef = MRI.getVRegDef(BaseReg);
-  }
-  if (!BaseDef)
-    return false;
-
-  int D = 0;
-  if (!TII->getIncrementValue(*BaseDef, D) && D >= 0)
-    return false;
-
-  Delta = D;
-  return true;
+  return findLoopIncrementValue(*BaseOp, Delta);
 }
 
 /// Check if we can change the instruction to use an offset value from the
@@ -2675,6 +2753,96 @@ MachineInstr *SwingSchedulerDAG::findDefInLoop(Register Reg) {
   return Def;
 }
 
+/// Return false if there is no overlap between the region accessed by BaseMI in
+/// an iteration and the region accessed by OtherMI in subsequent iterations.
+bool SwingSchedulerDAG::mayOverlapInLaterIter(
+    const MachineInstr *BaseMI, const MachineInstr *OtherMI) const {
+  int DeltaB, DeltaO, Delta;
+  if (!computeDelta(*BaseMI, DeltaB) || !computeDelta(*OtherMI, DeltaO) ||
+      DeltaB != DeltaO)
+    return true;
+  Delta = DeltaB;
+
+  const MachineOperand *BaseOpB, *BaseOpO;
+  int64_t OffsetB, OffsetO;
+  bool OffsetBIsScalable, OffsetOIsScalable;
+  const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
+  if (!TII->getMemOperandWithOffset(*BaseMI, BaseOpB, OffsetB,
+                                    OffsetBIsScalable, TRI) ||
+      !TII->getMemOperandWithOffset(*OtherMI, BaseOpO, OffsetO,
+                                    OffsetOIsScalable, TRI))
+    return true;
+
+  if (OffsetBIsScalable || OffsetOIsScalable)
+    return true;
+
+  if (!BaseOpB->isIdenticalTo(*BaseOpO)) {
+    // Pass cases with different base operands but same initial values.
+    // Typically for when pre/post increment is used.
+
+    if (!BaseOpB->isReg() || !BaseOpO->isReg())
+      return true;
+    Register RegB = BaseOpB->getReg(), RegO = BaseOpO->getReg();
+    if (!RegB.isVirtual() || !RegO.isVirtual())
+      return true;
+
+    MachineInstr *DefB = MRI.getVRegDef(BaseOpB->getReg());
+    MachineInstr *DefO = MRI.getVRegDef(BaseOpO->getReg());
+    if (!DefB || !DefO || !DefB->isPHI() || !DefO->isPHI())
+      return true;
+
+    unsigned InitValB = 0;
+    unsigned LoopValB = 0;
+    unsigned InitValO = 0;
+    unsigned LoopValO = 0;
+    getPhiRegs(*DefB, BB, InitValB, LoopValB);
+    getPhiRegs(*DefO, BB, InitValO, LoopValO);
+    MachineInstr *InitDefB = MRI.getVRegDef(InitValB);
+    MachineInstr *InitDefO = MRI.getVRegDef(InitValO);
+
+    if (!InitDefB->isIdenticalTo(*InitDefO))
+      return true;
+  }
+
+  LocationSize AccessSizeB = (*BaseMI->memoperands_begin())->getSize();
+  LocationSize AccessSizeO = (*OtherMI->memoperands_begin())->getSize();
+
+  // This is the main test, which checks the offset values and the loop
+  // increment value to determine if the accesses may be loop carried.
+  if (!AccessSizeB.hasValue() || !AccessSizeO.hasValue())
+    return true;
+
+  LLVM_DEBUG({
+    dbgs() << "Overlap check:\n";
+    dbgs() << "  BaseMI: ";
+    BaseMI->dump();
+    dbgs() << "    Base + " << OffsetB << " + I * " << Delta
+           << ", Len: " << AccessSizeB.getValue() << "\n";
+    dbgs() << "  OtherMI: ";
+    OtherMI->dump();
+    dbgs() << "    Base + " << OffsetO << " + I * " << Delta
+           << ", Len: " << AccessSizeO.getValue() << "\n";
+  });
+
+  if (Delta < 0) {
+    int64_t BaseMinAddr = OffsetB;
+    int64_t OhterNextIterMaxAddr = OffsetO + Delta + AccessSizeO.getValue() - 1;
+    if (BaseMinAddr > OhterNextIterMaxAddr) {
+      LLVM_DEBUG(dbgs() << "  Result: No overlap\n");
+      return false;
+    }
+  } else {
+    int64_t BaseMaxAddr = OffsetB + AccessSizeB.getValue() - 1;
+    int64_t OtherNextIterMinAddr = OffsetO + Delta;
+    if (BaseMaxAddr < OtherNextIterMinAddr) {
+      LLVM_DEBUG(dbgs() << "  Result: No overlap\n");
+      return false;
+    }
+  }
+  LLVM_DEBUG(dbgs() << "  Result: Overlap\n");
+  return true;
+}
+
 /// Return true for an order or output dependence that is loop carried
 /// potentially. A dependence is loop carried if the destination defines a value
 /// that may be used or defined by the source in a subsequent iteration.
@@ -2706,61 +2874,7 @@ bool SwingSchedulerDAG::isLoopCarriedDep(
   // The conservative assumption is that a dependence between memory operations
   // may be loop carried. The following code checks when it can be proved that
   // there is no loop carried dependence.
-  unsigned DeltaS, DeltaD;
-  if (!computeDelta(*SI, DeltaS) || !computeDelta(*DI, DeltaD))
-    return true;
-
-  const MachineOperand *BaseOpS, *BaseOpD;
-  int64_t OffsetS, OffsetD;
-  bool OffsetSIsScalable, OffsetDIsScalable;
-  const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
-  if (!TII->getMemOperandWithOffset(*SI, BaseOpS, OffsetS, OffsetSIsScalable,
-                                    TRI) ||
-      !TII->getMemOperandWithOffset(*DI, BaseOpD, OffsetD, OffsetDIsScalable,
-                                    TRI))
-    return true;
-
-  assert(!OffsetSIsScalable && !OffsetDIsScalable &&
-         "Expected offsets to be byte offsets");
-
-  MachineInstr *DefS = MRI.getVRegDef(BaseOpS->getReg());
-  MachineInstr *DefD = MRI.getVRegDef(BaseOpD->getReg());
-  if (!DefS || !DefD || !DefS->isPHI() || !DefD->isPHI())
-    return true;
-
-  unsigned InitValS = 0;
-  unsigned LoopValS = 0;
-  unsigned InitValD = 0;
-  unsigned LoopValD = 0;
-  getPhiRegs(*DefS, BB, InitValS, LoopValS);
-  getPhiRegs(*DefD, BB, InitValD, LoopValD);
-  MachineInstr *InitDefS = MRI.getVRegDef(InitValS);
-  MachineInstr *InitDefD = MRI.getVRegDef(InitValD);
-
-  if (!InitDefS->isIdenticalTo(*InitDefD))
-    return true;
-
-  // Check that the base register is incremented by a constant value for each
-  // iteration.
-  MachineInstr *LoopDefS = MRI.getVRegDef(LoopValS);
-  int D = 0;
-  if (!LoopDefS || !TII->getIncrementValue(*LoopDefS, D))
-    return true;
-
-  LocationSize AccessSizeS = (*SI->memoperands_begin())->getSize();
-  LocationSize AccessSizeD = (*DI->memoperands_begin())->getSize();
-
-  // This is the main test, which checks the offset values and the loop
-  // increment value to determine if the accesses may be loop carried.
-  if (!AccessSizeS.hasValue() || !AccessSizeD.hasValue())
-    return true;
-
-  if (DeltaS != DeltaD || DeltaS < AccessSizeS.getValue() ||
-      DeltaD < AccessSizeD.getValue())
-    return true;
-
-  return (OffsetS + (int64_t)AccessSizeS.getValue() <
-          OffsetD + (int64_t)AccessSizeD.getValue());
+  return mayOverlapInLaterIter(DI, SI);
 }
 
 void SwingSchedulerDAG::postProcessDAG() {
diff --git a/llvm/test/CodeGen/Hexagon/swp-carried-dep1.mir b/llvm/test/CodeGen/Hexagon/swp-carried-dep1.mir
@@ -3,6 +3,13 @@
 
 # Test that the loop carried dependence check correctly identifies a recurrence.
 
+# CHECK: Overlap check:
+# CHECK-NEXT:   BaseMI:   S2_storerh_io %{{[0-9]+}}:intregs, 0, %{{[0-9]+}}:intregs :: (store (s16) into %ir.lsr.iv24)
+# CHECK-NEXT:     Base + 0 + I * 4, Len: 2
+# CHECK-NEXT:   OtherMI:   %{{[0-9]+}}:intregs = L2_loadrh_io %{{[0-9]+}}:intregs, -8 :: (load (s16) from %ir.cgep10)
+# CHECK-NEXT:     Base + -8 + I * 4, Len: 2
+# CHECK-NEXT:   Result: Overlap
+
 # CHECK: Rec NodeSet
 # CHECK: Rec NodeSet
 # CHECK: Rec NodeSet
diff --git a/llvm/test/CodeGen/Hexagon/swp-carried-dep2.mir b/llvm/test/CodeGen/Hexagon/swp-carried-dep2.mir
@@ -1,14 +1,15 @@
 # RUN: llc -mtriple=hexagon -run-pass pipeliner -debug-only=pipeliner %s -o /dev/null 2>&1 -pipeliner-experimental-cg=true | FileCheck %s
 # REQUIRES: asserts
 
-# Test that the loop carried dependence check correctly identifies a recurrence
+# Test that it correctly recognizes that there is no loop carried dependence
 # when the loop variable decreases and the array index offset is negative.
 
-# CHECK: Rec NodeSet
-# CHECK: Rec NodeSet
-# CHECK: SU(3)
-# CHECK: SU(4)
-# CHECK: SU(5)
+# CHECK: Overlap check:
+# CHECK-NEXT:   BaseMI:   S2_storeri_io %{{[0-9]+}}:intregs, 0, %{{[0-9]+}}:intregs :: (store (s32) into %ir.lsr.iv1)
+# CHECK-NEXT:     Base + 0 + I * -4, Len: 4
+# CHECK-NEXT:   OtherMI:   %{{[0-9]+}}:intregs = L2_loadri_io %{{[0-9]+}}:intregs, -8 :: (load (s32) from %ir.cgep)
+# CHECK-NEXT:     Base + -8 + I * -4, Len: 4
+# CHECK-NEXT:   Result: No overlap
 
 --- |
 
diff --git a/llvm/test/CodeGen/Hexagon/swp-carried-dep3.mir b/llvm/test/CodeGen/Hexagon/swp-carried-dep3.mir
@@ -7,6 +7,13 @@
 # requires to use a single CHECK-NOT to match such a Rec NodeSet. Fortunately
 # the atom '.' does not match a newline but anything else on a line.
 
+# CHECK: Overlap check:
+# CHECK-NEXT:   BaseMI:   %13:intregs = S2_storerh_pi %12:intregs(tied-def 0), 2, %20:intregs :: (store (s16))
+# CHECK-NEXT:     Base + 0 + I * 2, Len: 2
+# CHECK-NEXT:   OtherMI:   %19:intregs, %15:intregs = L2_loadrh_pi %14:intregs(tied-def 1), 2 :: (load (s16))
+# CHECK-NEXT:     Base + 0 + I * 2, Len: 2
+# CHECK-NEXT:   Result: No overlap
+
 # CHECK-NOT: Rec NodeSet{{.+[[:space:]]}} SU(5){{.+[[:space:]]}} SU(7)
 
 ...
diff --git a/llvm/test/CodeGen/Hexagon/swp-carried-dep4.mir b/llvm/test/CodeGen/Hexagon/swp-carried-dep4.mir
@@ -0,0 +1,37 @@
+# RUN: llc -mtriple=hexagon -run-pass pipeliner -debug-only=pipeliner %s -o /dev/null 2>&1 | FileCheck %s
+# REQUIRES: asserts
+
+# Loop carried dependence is assumed in cases where increment value cannot be recognized
+# (Not supported for multiple increment instruction)
+
+# CHECK:  Rec NodeSet
+# CHECK:  Rec NodeSet
+# CHECK-NEXT:    SU(1)
+# CHECK-NEXT:    SU(2)
+
+---
+name:            test
+tracksRegLiveness: true
+
+body:             |
+  bb.0:
+    successors: %bb.1
+
+    %10:intregs = IMPLICIT_DEF
+    %11:intregs = IMPLICIT_DEF
+    J2_loop0i %bb.1, 6, implicit-def $lc0, implicit-def $sa0, implicit-def $usr
+
+  bb.1 (machine-block-address-taken):
+    successors: %bb.1, %bb.2
+
+    %0:intregs = PHI %11, %bb.0, %6, %bb.1
+    %4:intregs = L2_loadri_io %0, 0 :: (load (s32))
+    S2_storeri_io %0, 0, %10 :: (store (s32))
+    %7:intregs = A2_addi %0, -8
+    %6:intregs = A2_addi %7, 4
+    ENDLOOP0 %bb.1, implicit-def $pc, implicit-def $lc0, implicit $sa0, implicit $lc0
+    J2_jump %bb.2, implicit-def dead $pc
+
+  bb.2:
+
+...
diff --git a/llvm/test/CodeGen/Hexagon/swp-carried-dep5.mir b/llvm/test/CodeGen/Hexagon/swp-carried-dep5.mir
diff --git a/llvm/test/CodeGen/Hexagon/swp-epilog-phi8.ll b/llvm/test/CodeGen/Hexagon/swp-epilog-phi8.ll
