[RISCV] Teach RISCVInsertVSETVLI to work without LiveIntervals (#94686)

Stacked on #94658.
    
We recently moved RISCVInsertVSETVLI from before vector register
allocation to after vector register allocation. When doing so, we added
an unconditional dependency on LiveIntervals - even at O0 where
LiveIntevals hadn't previously run. As reported in #93587, this was
apparently not safe to do.
    
This change makes LiveIntervals optional, and adjusts all the update
code to only run wen live intervals is present. The only real tricky
part of this change is the abstract state tracking in the dataflow. We
need to represent a "register w/unknown definition" state - but only
when we don't have LiveIntervals.
    
This adjust the abstract state definition so that the AVLIsReg state can
represent either a register + valno, or a register + unknown definition.
With LiveIntervals, we have an exact definition for each AVL use.
Without LiveIntervals, we treat the definition of a register AVL as
being unknown.
    
The key semantic change is that we now have a state in the lattice for
which something is known about the AVL value, but for which two
identical lattice elements do *not* necessarily represent the same AVL
value at runtime. Previously, the only case which could result in such
an unknown AVL was the fully unknown state (where VTYPE is also fully
unknown). This requires a small adjustment to hasSameAVL and lattice
state equality to draw this important distinction.
    
The net effect of this patch is that we remove the LiveIntervals
dependency at O0, and O0 code quality will regress for cases involving
register AVL values.
    
This patch is an alternative to
#93796 and
#94340. It is very directly
inspired by review conversation around them, and thus should be
considered coauthored by Luke.

Author: preames

llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp

@@ -48,10 +48,13 @@ static cl::opt<bool> DisableInsertVSETVLPHIOpt(
 namespace {
 
 /// Given a virtual register \p Reg, return the corresponding VNInfo for it.
-/// This will return nullptr if the virtual register is an implicit_def.
+/// This will return nullptr if the virtual register is an implicit_def or
+/// if LiveIntervals is not available.
 static VNInfo *getVNInfoFromReg(Register Reg, const MachineInstr &MI,
                                 const LiveIntervals *LIS) {
   assert(Reg.isVirtual());
+  if (!LIS)
+    return nullptr;
   auto &LI = LIS->getInterval(Reg);
   SlotIndex SI = LIS->getSlotIndexes()->getInstructionIndex(MI);
   return LI.getVNInfoBefore(SI);
@@ -512,7 +515,8 @@ DemandedFields getDemanded(const MachineInstr &MI, const RISCVSubtarget *ST) {
 /// values of the VL and VTYPE registers after insertion.
 class VSETVLIInfo {
   struct AVLDef {
-    // Every AVLDef should have a VNInfo.
+    // Every AVLDef should have a VNInfo, unless we're running without
+    // LiveIntervals in which case this will be nullptr.
     const VNInfo *ValNo;
     Register DefReg;
   };
@@ -526,7 +530,7 @@ class VSETVLIInfo {
     AVLIsReg,
     AVLIsImm,
     AVLIsVLMAX,
-    Unknown,
+    Unknown, // AVL and VTYPE are fully unknown
   } State = Uninitialized;
 
   // Fields from VTYPE.
@@ -552,7 +556,7 @@ class VSETVLIInfo {
   bool isUnknown() const { return State == Unknown; }
 
   void setAVLRegDef(const VNInfo *VNInfo, Register AVLReg) {
-    assert(VNInfo && AVLReg.isVirtual());
+    assert(AVLReg.isVirtual());
     AVLRegDef.ValNo = VNInfo;
     AVLRegDef.DefReg = AVLReg;
     State = AVLIsReg;
@@ -582,9 +586,11 @@ class VSETVLIInfo {
   }
   // Most AVLIsReg infos will have a single defining MachineInstr, unless it was
   // a PHI node. In that case getAVLVNInfo()->def will point to the block
-  // boundary slot.
+  // boundary slot.  If LiveIntervals isn't available, then nullptr is returned.
   const MachineInstr *getAVLDefMI(const LiveIntervals *LIS) const {
     assert(hasAVLReg());
+    if (!LIS)
+      return nullptr;
     auto *MI = LIS->getInstructionFromIndex(getAVLVNInfo()->def);
     assert(!(getAVLVNInfo()->isPHIDef() && MI));
     return MI;
@@ -628,10 +634,15 @@ class VSETVLIInfo {
     return (hasNonZeroAVL(LIS) && Other.hasNonZeroAVL(LIS));
   }
 
-  bool hasSameAVL(const VSETVLIInfo &Other) const {
-    if (hasAVLReg() && Other.hasAVLReg())
+  bool hasSameAVLLatticeValue(const VSETVLIInfo &Other) const {
+    if (hasAVLReg() && Other.hasAVLReg()) {
+      assert(!getAVLVNInfo() == !Other.getAVLVNInfo() &&
+             "we either have intervals or we don't");
+      if (!getAVLVNInfo())
+        return getAVLReg() == Other.getAVLReg();
       return getAVLVNInfo()->id == Other.getAVLVNInfo()->id &&
              getAVLReg() == Other.getAVLReg();
+    }
 
     if (hasAVLImm() && Other.hasAVLImm())
       return getAVLImm() == Other.getAVLImm();
@@ -642,6 +653,21 @@ class VSETVLIInfo {
     return false;
   }
 
+  // Return true if the two lattice values are guaranteed to have
+  // the same AVL value at runtime.
+  bool hasSameAVL(const VSETVLIInfo &Other) const {
+    // Without LiveIntervals, we don't know which instruction defines a
+    // register.  Since a register may be redefined, this means all AVLIsReg
+    // states must be treated as possibly distinct.
+    if (hasAVLReg() && Other.hasAVLReg()) {
+      assert(!getAVLVNInfo() == !Other.getAVLVNInfo() &&
+             "we either have intervals or we don't");
+      if (!getAVLVNInfo())
+        return false;
+    }
+    return hasSameAVLLatticeValue(Other);
+  }
+
   void setVTYPE(unsigned VType) {
     assert(isValid() && !isUnknown() &&
            "Can't set VTYPE for uninitialized or unknown");
@@ -741,7 +767,7 @@ class VSETVLIInfo {
     if (Other.isUnknown())
       return isUnknown();
 
-    if (!hasSameAVL(Other))
+    if (!hasSameAVLLatticeValue(Other))
       return false;
 
     // If the SEWLMULRatioOnly bits are different, then they aren't equal.
@@ -849,6 +875,7 @@ class RISCVInsertVSETVLI : public MachineFunctionPass {
   const RISCVSubtarget *ST;
   const TargetInstrInfo *TII;
   MachineRegisterInfo *MRI;
+  // Possibly null!
   LiveIntervals *LIS;
 
   std::vector<BlockData> BlockInfo;
@@ -863,9 +890,9 @@ class RISCVInsertVSETVLI : public MachineFunctionPass {
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.setPreservesCFG();
 
-    AU.addRequired<LiveIntervals>();
+    AU.addUsedIfAvailable<LiveIntervals>();
     AU.addPreserved<LiveIntervals>();
-    AU.addRequired<SlotIndexes>();
+    AU.addUsedIfAvailable<SlotIndexes>();
     AU.addPreserved<SlotIndexes>();
     AU.addPreserved<LiveDebugVariables>();
     AU.addPreserved<LiveStacks>();
@@ -1061,7 +1088,8 @@ void RISCVInsertVSETVLI::insertVSETVLI(MachineBasicBlock &MBB,
                     .addReg(RISCV::X0, RegState::Kill)
                     .addImm(Info.encodeVTYPE())
                     .addReg(RISCV::VL, RegState::Implicit);
-      LIS->InsertMachineInstrInMaps(*MI);
+      if (LIS)
+        LIS->InsertMachineInstrInMaps(*MI);
       return;
     }
 
@@ -1078,7 +1106,8 @@ void RISCVInsertVSETVLI::insertVSETVLI(MachineBasicBlock &MBB,
                         .addReg(RISCV::X0, RegState::Kill)
                         .addImm(Info.encodeVTYPE())
                         .addReg(RISCV::VL, RegState::Implicit);
-          LIS->InsertMachineInstrInMaps(*MI);
+          if (LIS)
+            LIS->InsertMachineInstrInMaps(*MI);
           return;
         }
       }
@@ -1090,7 +1119,8 @@ void RISCVInsertVSETVLI::insertVSETVLI(MachineBasicBlock &MBB,
                   .addReg(RISCV::X0, RegState::Define | RegState::Dead)
                   .addImm(Info.getAVLImm())
                   .addImm(Info.encodeVTYPE());
-    LIS->InsertMachineInstrInMaps(*MI);
+    if (LIS)
+      LIS->InsertMachineInstrInMaps(*MI);
     return;
   }
 
@@ -1100,8 +1130,10 @@ void RISCVInsertVSETVLI::insertVSETVLI(MachineBasicBlock &MBB,
                   .addReg(DestReg, RegState::Define | RegState::Dead)
                   .addReg(RISCV::X0, RegState::Kill)
                   .addImm(Info.encodeVTYPE());
-    LIS->InsertMachineInstrInMaps(*MI);
-    LIS->createAndComputeVirtRegInterval(DestReg);
+    if (LIS) {
+      LIS->InsertMachineInstrInMaps(*MI);
+      LIS->createAndComputeVirtRegInterval(DestReg);
+    }
     return;
   }
 
@@ -1111,12 +1143,14 @@ void RISCVInsertVSETVLI::insertVSETVLI(MachineBasicBlock &MBB,
                 .addReg(RISCV::X0, RegState::Define | RegState::Dead)
                 .addReg(AVLReg)
                 .addImm(Info.encodeVTYPE());
-  LIS->InsertMachineInstrInMaps(*MI);
-  // Normally the AVL's live range will already extend past the inserted vsetvli
-  // because the pseudos below will already use the AVL. But this isn't always
-  // the case, e.g. PseudoVMV_X_S doesn't have an AVL operand.
-  LIS->getInterval(AVLReg).extendInBlock(
-      LIS->getMBBStartIdx(&MBB), LIS->getInstructionIndex(*MI).getRegSlot());
+  if (LIS) {
+    LIS->InsertMachineInstrInMaps(*MI);
+    // Normally the AVL's live range will already extend past the inserted
+    // vsetvli because the pseudos below will already use the AVL. But this
+    // isn't always the case, e.g. PseudoVMV_X_S doesn't have an AVL operand.
+    LIS->getInterval(AVLReg).extendInBlock(
+        LIS->getMBBStartIdx(&MBB), LIS->getInstructionIndex(*MI).getRegSlot());
+  }
 }
 
 /// Return true if a VSETVLI is required to transition from CurInfo to Require
@@ -1230,10 +1264,14 @@ void RISCVInsertVSETVLI::transferAfter(VSETVLIInfo &Info,
   if (RISCV::isFaultFirstLoad(MI)) {
     // Update AVL to vl-output of the fault first load.
     assert(MI.getOperand(1).getReg().isVirtual());
-    auto &LI = LIS->getInterval(MI.getOperand(1).getReg());
-    SlotIndex SI = LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot();
-    VNInfo *VNI = LI.getVNInfoAt(SI);
-    Info.setAVLRegDef(VNI, MI.getOperand(1).getReg());
+    if (LIS) {
+      auto &LI = LIS->getInterval(MI.getOperand(1).getReg());
+      SlotIndex SI =
+          LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot();
+      VNInfo *VNI = LI.getVNInfoAt(SI);
+      Info.setAVLRegDef(VNI, MI.getOperand(1).getReg());
+    } else
+      Info.setAVLRegDef(nullptr, MI.getOperand(1).getReg());
     return;
   }
 
@@ -1327,6 +1365,9 @@ bool RISCVInsertVSETVLI::needVSETVLIPHI(const VSETVLIInfo &Require,
   if (!Require.hasAVLReg())
     return true;
 
+  if (!LIS)
+    return true;
+
   // We need the AVL to have been produced by a PHI node in this basic block.
   const VNInfo *Valno = Require.getAVLVNInfo();
   if (!Valno->isPHIDef() || LIS->getMBBFromIndex(Valno->def) != &MBB)
@@ -1402,27 +1443,29 @@ void RISCVInsertVSETVLI::emitVSETVLIs(MachineBasicBlock &MBB) {
         MachineOperand &VLOp = MI.getOperand(getVLOpNum(MI));
         if (VLOp.isReg()) {
           Register Reg = VLOp.getReg();
-          LiveInterval &LI = LIS->getInterval(Reg);
 
           // Erase the AVL operand from the instruction.
           VLOp.setReg(RISCV::NoRegister);
           VLOp.setIsKill(false);
-          SmallVector<MachineInstr *> DeadMIs;
-          LIS->shrinkToUses(&LI, &DeadMIs);
-          // We might have separate components that need split due to
-          // needVSETVLIPHI causing us to skip inserting a new VL def.
-          SmallVector<LiveInterval *> SplitLIs;
-          LIS->splitSeparateComponents(LI, SplitLIs);
-
-          // If the AVL was an immediate > 31, then it would have been emitted
-          // as an ADDI. However, the ADDI might not have been used in the
-          // vsetvli, or a vsetvli might not have been emitted, so it may be
-          // dead now.
-          for (MachineInstr *DeadMI : DeadMIs) {
-            if (!TII->isAddImmediate(*DeadMI, Reg))
-              continue;
-            LIS->RemoveMachineInstrFromMaps(*DeadMI);
-            DeadMI->eraseFromParent();
+          if (LIS) {
+            LiveInterval &LI = LIS->getInterval(Reg);
+            SmallVector<MachineInstr *> DeadMIs;
+            LIS->shrinkToUses(&LI, &DeadMIs);
+            // We might have separate components that need split due to
+            // needVSETVLIPHI causing us to skip inserting a new VL def.
+            SmallVector<LiveInterval *> SplitLIs;
+            LIS->splitSeparateComponents(LI, SplitLIs);
+
+            // If the AVL was an immediate > 31, then it would have been emitted
+            // as an ADDI. However, the ADDI might not have been used in the
+            // vsetvli, or a vsetvli might not have been emitted, so it may be
+            // dead now.
+            for (MachineInstr *DeadMI : DeadMIs) {
+              if (!TII->isAddImmediate(*DeadMI, Reg))
+                continue;
+              LIS->RemoveMachineInstrFromMaps(*DeadMI);
+              DeadMI->eraseFromParent();
+            }
           }
         }
         MI.addOperand(MachineOperand::CreateReg(RISCV::VL, /*isDef*/ false,
@@ -1479,6 +1522,9 @@ void RISCVInsertVSETVLI::doPRE(MachineBasicBlock &MBB) {
   if (!UnavailablePred || !AvailableInfo.isValid())
     return;
 
+  if (!LIS)
+    return;
+
   // If we don't know the exact VTYPE, we can't copy the vsetvli to the exit of
   // the unavailable pred.
   if (AvailableInfo.hasSEWLMULRatioOnly())
@@ -1625,7 +1671,7 @@ void RISCVInsertVSETVLI::coalesceVSETVLIs(MachineBasicBlock &MBB) const {
 
           // The def of DefReg moved to MI, so extend the LiveInterval up to
           // it.
-          if (DefReg.isVirtual()) {
+          if (DefReg.isVirtual() && LIS) {
             LiveInterval &DefLI = LIS->getInterval(DefReg);
             SlotIndex MISlot = LIS->getInstructionIndex(MI).getRegSlot();
             VNInfo *DefVNI = DefLI.getVNInfoAt(DefLI.beginIndex());
@@ -1654,13 +1700,15 @@ void RISCVInsertVSETVLI::coalesceVSETVLIs(MachineBasicBlock &MBB) const {
 
           if (OldVLReg && OldVLReg.isVirtual()) {
             // NextMI no longer uses OldVLReg so shrink its LiveInterval.
-            LIS->shrinkToUses(&LIS->getInterval(OldVLReg));
+            if (LIS)
+              LIS->shrinkToUses(&LIS->getInterval(OldVLReg));
 
             MachineInstr *VLOpDef = MRI->getUniqueVRegDef(OldVLReg);
             if (VLOpDef && TII->isAddImmediate(*VLOpDef, OldVLReg) &&
                 MRI->use_nodbg_empty(OldVLReg)) {
               VLOpDef->eraseFromParent();
-              LIS->removeInterval(OldVLReg);
+              if (LIS)
+                LIS->removeInterval(OldVLReg);
             }
           }
           MI.setDesc(NextMI->getDesc());
@@ -1676,7 +1724,8 @@ void RISCVInsertVSETVLI::coalesceVSETVLIs(MachineBasicBlock &MBB) const {
 
   NumCoalescedVSETVL += ToDelete.size();
   for (auto *MI : ToDelete) {
-    LIS->RemoveMachineInstrFromMaps(*MI);
+    if (LIS)
+      LIS->RemoveMachineInstrFromMaps(*MI);
     MI->eraseFromParent();
   }
 }
@@ -1691,12 +1740,14 @@ void RISCVInsertVSETVLI::insertReadVL(MachineBasicBlock &MBB) {
         auto ReadVLMI = BuildMI(MBB, I, MI.getDebugLoc(),
                                 TII->get(RISCV::PseudoReadVL), VLOutput);
         // Move the LiveInterval's definition down to PseudoReadVL.
-        SlotIndex NewDefSI =
-            LIS->InsertMachineInstrInMaps(*ReadVLMI).getRegSlot();
-        LiveInterval &DefLI = LIS->getInterval(VLOutput);
-        VNInfo *DefVNI = DefLI.getVNInfoAt(DefLI.beginIndex());
-        DefLI.removeSegment(DefLI.beginIndex(), NewDefSI);
-        DefVNI->def = NewDefSI;
+        if (LIS) {
+          SlotIndex NewDefSI =
+              LIS->InsertMachineInstrInMaps(*ReadVLMI).getRegSlot();
+          LiveInterval &DefLI = LIS->getInterval(VLOutput);
+          VNInfo *DefVNI = DefLI.getVNInfoAt(DefLI.beginIndex());
+          DefLI.removeSegment(DefLI.beginIndex(), NewDefSI);
+          DefVNI->def = NewDefSI;
+        }
       }
       // We don't use the vl output of the VLEFF/VLSEGFF anymore.
       MI.getOperand(1).setReg(RISCV::X0);
@@ -1714,7 +1765,7 @@ bool RISCVInsertVSETVLI::runOnMachineFunction(MachineFunction &MF) {
 
   TII = ST->getInstrInfo();
   MRI = &MF.getRegInfo();
-  LIS = &getAnalysis<LiveIntervals>();
+  LIS = getAnalysisIfAvailable<LiveIntervals>();
 
   assert(BlockInfo.empty() && "Expect empty block infos");
   BlockInfo.resize(MF.getNumBlockIDs());

llvm/test/CodeGen/RISCV/O0-pipeline.ll

@@ -47,9 +47,6 @@
 ; CHECK-NEXT:       Eliminate PHI nodes for register allocation
 ; CHECK-NEXT:       Two-Address instruction pass
 ; CHECK-NEXT:       Fast Register Allocator
-; CHECK-NEXT:       MachineDominator Tree Construction
-; CHECK-NEXT:       Slot index numbering
-; CHECK-NEXT:       Live Interval Analysis
 ; CHECK-NEXT:       RISC-V Insert VSETVLI pass
 ; CHECK-NEXT:       Fast Register Allocator
 ; CHECK-NEXT:       Remove Redundant DEBUG_VALUE analysis

llvm/test/CodeGen/RISCV/rvv/pr93587.ll

@@ -0,0 +1,37 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=riscv64 -O0 < %s | FileCheck %s
+
+; Make sure we don't run LiveIntervals at O0, otherwise it will crash when
+; running on this unreachable block.
+
+define i16 @f() {
+; CHECK-LABEL: f:
+; CHECK:       # %bb.0: # %BB
+; CHECK-NEXT:    addi sp, sp, -16
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    j .LBB0_1
+; CHECK-NEXT:  .LBB0_1: # %BB1
+; CHECK-NEXT:    # =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    li a0, 0
+; CHECK-NEXT:    sd a0, 8(sp) # 8-byte Folded Spill
+; CHECK-NEXT:    j .LBB0_1
+; CHECK-NEXT:  # %bb.2: # %BB1
+; CHECK-NEXT:    li a0, 0
+; CHECK-NEXT:    bnez a0, .LBB0_1
+; CHECK-NEXT:    j .LBB0_3
+; CHECK-NEXT:  .LBB0_3: # %BB2
+; CHECK-NEXT:    ld a0, 8(sp) # 8-byte Folded Reload
+; CHECK-NEXT:    addi sp, sp, 16
+; CHECK-NEXT:    ret
+BB:
+  br label %BB1
+
+BB1:
+  %A = or i16 0, 0
+  %B = fcmp true float 0.000000e+00, 0.000000e+00
+  %C = or i1 %B, false
+  br i1 %C, label %BB1, label %BB2
+
+BB2:
+  ret i16 %A
+}