[RISCV][InsertVSETVLI] Rework code structure to make reasoning about undefined lanes explicit [NFC]

[reames]

Summary

We already have several places in this code which reason about whether the inactive lanes are defined, and are about to add one more in D151653. Let's go ahead and common the code so that we don't have the same concept repeating in multiply places.

Asking for review mostly because of the change to computeInfoForInstr. I want a second pair of eyes confirming I'm not missing something there. (All of the removed asserts are covered by the verifier; I checked.)

[luke]

LGTM, nice to see the clarification between undefined and agnostic

[frasercrmck]

LGTM with a comment question. Agreed, this is nice to see.

llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp

905

Does this comment hold? Could it now be improved?

llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp

 static bool isMaskRegOp(const MachineInstr &MI) {
   if (!RISCVII::hasSEWOp(MI.getDesc().TSFlags))
     return false;
   const unsigned Log2SEW = MI.getOperand(getSEWOpNum(MI)).getImm();
   // A Log2SEW of 0 is an operation on mask registers only.
   return Log2SEW == 0;
 }
 
+/// Return true if the inactive elements in the result are entirely undefined.
+/// Note that this is different from "agnostic" as defined by the vector
+/// specification.  Agnostic requires each lane to either be undisturbed, or
+/// take the value -1; no other value is allowed.
+static bool hasUndefinedMergeOp(const MachineInstr &MI,
+                                const MachineRegisterInfo &MRI) {
+
+  unsigned UseOpIdx;
+  if (!MI.isRegTiedToUseOperand(0, &UseOpIdx))
+    // If there is no passthrough operand, then the pass through
+    // lanes are undefined.
+    return true;
+
+  // If the tied operand is an IMPLICIT_DEF, the pass through lanes
+  // are undefined.
+  const MachineOperand &UseMO = MI.getOperand(UseOpIdx);
+  MachineInstr *UseMI = MRI.getVRegDef(UseMO.getReg());
+  return UseMI && UseMI->isImplicitDef();
+}
+
 /// Which subfields of VL or VTYPE have values we need to preserve?
 struct DemandedFields {
   // Some unknown property of VL is used.  If demanded, must preserve entire
   // value.
   bool VLAny = false;
   // Only zero vs non-zero is used. If demanded, can change non-zero values.
   bool VLZeroness = false;
   // What properties of SEW we need to preserve.
if (isMaskRegOp(MI)) {
     Res.LMUL = false;
   }
 
   // For vmv.s.x and vfmv.s.f, there are only two behaviors, VL = 0 and VL > 0.
   if (isScalarMoveInstr(MI)) {
     Res.LMUL = false;
     Res.SEWLMULRatio = false;
     Res.VLAny = false;
-    // For vmv.s.x and vfmv.s.f, if writing to an implicit_def operand, we don't
+    // For vmv.s.x and vfmv.s.f, if the merge operand is *undefined*, we don't
     // need to preserve any other bits and are thus compatible with any larger,
     // etype and can disregard policy bits.  Warning: It's tempting to try doing
     // this for any tail agnostic operation, but we can't as TA requires
     // tail lanes to either be the original value or -1.  We are writing
     // unknown bits to the lanes here.
-    auto *VRegDef = MRI->getVRegDef(MI.getOperand(1).getReg());
-    if (VRegDef && VRegDef->isImplicitDef()) {
+    if (hasUndefinedMergeOp(MI, *MRI)) {
       Res.SEW = DemandedFields::SEWGreaterThanOrEqual;
       Res.TailPolicy = false;
     }
   }
 
   return Res;
 }
 

 
 INITIALIZE_PASS(RISCVInsertVSETVLI, DEBUG_TYPE, RISCV_INSERT_VSETVLI_NAME,
                 false, false)
 
 static VSETVLIInfo computeInfoForInstr(const MachineInstr &MI, uint64_t TSFlags,
                                        const MachineRegisterInfo *MRI) {
   VSETVLIInfo InstrInfo;
 
-  bool TailAgnostic, MaskAgnostic;
-  unsigned UseOpIdx;
-  if (MI.isRegTiedToUseOperand(0, &UseOpIdx)) {
+  bool TailAgnostic = true;
+  bool MaskAgnostic = true;
+  if (!hasUndefinedMergeOp(MI, *MRI)) {
     // Start with undisturbed.
     TailAgnostic = false;
     MaskAgnostic = false;
 
     // If there is a policy operand, use it.
     if (RISCVII::hasVecPolicyOp(TSFlags)) {
       const MachineOperand &Op = MI.getOperand(MI.getNumExplicitOperands() - 1);
       uint64_t Policy = Op.getImm();
       assert(Policy <= (RISCVII::TAIL_AGNOSTIC | RISCVII::MASK_AGNOSTIC) &&
              "Invalid Policy Value");
       TailAgnostic = Policy & RISCVII::TAIL_AGNOSTIC;
       MaskAgnostic = Policy & RISCVII::MASK_AGNOSTIC;
     }
 
-    // If the tied operand is an IMPLICIT_DEF we can use TailAgnostic and
-    // MaskAgnostic.
-    const MachineOperand &UseMO = MI.getOperand(UseOpIdx);
-    MachineInstr *UseMI = MRI->getVRegDef(UseMO.getReg());
-    if (UseMI && UseMI->isImplicitDef()) {
-      TailAgnostic = true;
-      MaskAgnostic = true;
-    }
     // Some pseudo instructions force a tail agnostic policy despite having a
     // tied def.
     if (RISCVII::doesForceTailAgnostic(TSFlags))
       TailAgnostic = true;
 
     if (!RISCVII::usesMaskPolicy(TSFlags))
       MaskAgnostic = true;
-  } else {
-    // If there is no tied operand,, there shouldn't be a policy operand.
-    assert(!RISCVII::hasVecPolicyOp(TSFlags) && "Unexpected policy operand");
-    // No tied operand use agnostic policies.
-    TailAgnostic = true;
-    MaskAgnostic = true;
   }
 
   RISCVII::VLMUL VLMul = RISCVII::getLMul(TSFlags);
 
   unsigned Log2SEW = MI.getOperand(getSEWOpNum(MI)).getImm();
   // A Log2SEW of 0 is an operation on mask registers only.
   unsigned SEW = Log2SEW ? 1 << Log2SEW : 8;
   assert(RISCVVType::isValidSEW(SEW) && "Unexpected SEW");
bool RISCVInsertVSETVLI::needVSETVLI(const MachineInstr &MI,
                                      const VSETVLIInfo &CurInfo) const {
   assert(Require == computeInfoForInstr(MI, MI.getDesc().TSFlags, MRI));
 
   if (!CurInfo.isValid() || CurInfo.isUnknown() || CurInfo.hasSEWLMULRatioOnly())
     return true;
 
   DemandedFields Used = getDemanded(MI, MRI);
 
-  // A slidedown/slideup with an IMPLICIT_DEF merge op can freely clobber
+  // A slidedown/slideup with an *undefined* merge op can freely clobber
   // elements not copied from the source vector (e.g. masked off, tail, or
   // slideup's prefix). Notes:
   // * We can't modify SEW here since the slide amount is in units of SEW.
   // * VL=1 is special only because we have existing support for zero vs
   //   non-zero VL.  We could generalize this if we had a VL > C predicate.
   // * The LMUL1 restriction is for machines whose latency may depend on VL.
-  // * As above, this is only legal for IMPLICIT_DEF, not TA.
+  // * As above, this is only legal for tail "undefined" not "agnostic".
   if (isVSlideInstr(MI) && Require.hasAVLImm() && Require.getAVLImm() == 1 &&
-      isLMUL1OrSmaller(CurInfo.getVLMUL())) {
-    auto *VRegDef = MRI->getVRegDef(MI.getOperand(1).getReg());
-    if (VRegDef && VRegDef->isImplicitDef()) {
-      Used.VLAny = false;
-      Used.VLZeroness = true;
-      Used.LMUL = false;
-      Used.TailPolicy = false;
-    }
-  }
+      isLMUL1OrSmaller(CurInfo.getVLMUL()) && hasUndefinedMergeOp(MI, *MRI)) {
 
   if (CurInfo.isCompatible(Used, Require, *MRI))
     return false;
 
   // We didn't find a compatible value. If our AVL is a virtual register,
   // it might be defined by a VSET(I)VLI. If it has the same VLMAX we need
   // and the last VL/VTYPE we observed is the same, we don't need a
   // VSETVLI here.