[RISCV] Move exact VLEN VLMAX encoding to RISCVInsertVSETVLI. NFC-ish

In llvm#75412 and llvm#75509 we started to use the vsetvli a0, zero encoding to reduce
register pressure when we know the exact VLEN.

We do this by canonicalizing VL ops to X0 in RISCVISelLowering.

This is only needed for cases where the AVL doesn't fit inside an immediate
though, since we already have code in RISCVInsertVSETVLI that detects if we
know the exact VLEN and uses the immediate encoding if possible.

This patch removes the need to do the canonicalization in RISCVISelLowering by
handling said case in RISCVInsertVSETVLI itself.

There are a handful of reasons I'm proposing to do this:
- It keeps the vsetvli logic in the one pass
- We get to move code out of SelectionDAG so we don't forget about it in
  GlobalISEL when the time comes
- Canonicalizing the VL in RISCVISelLowering means that we lose the original VL
  for fixed length vectors, which might be useful information for passes that
  sit between lowering and RISCVInsertVSETVLI. (I discovered this when working
  on a patch for RISCVFoldMasks.cpp that worked on fixed length vectors)

Author: lukel97

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -2592,16 +2592,8 @@ static SDValue getAllOnesMask(MVT VecVT, SDValue VL, const SDLoc &DL,
   return DAG.getNode(RISCVISD::VMSET_VL, DL, MaskVT, VL);
 }
 
-static SDValue getVLOp(uint64_t NumElts, MVT ContainerVT, const SDLoc &DL,
-                       SelectionDAG &DAG, const RISCVSubtarget &Subtarget) {
-  // If we know the exact VLEN, and our VL is exactly equal to VLMAX,
-  // canonicalize the representation.  InsertVSETVLI will pick the immediate
-  // encoding later if profitable.
-  const auto [MinVLMAX, MaxVLMAX] =
-      RISCVTargetLowering::computeVLMAXBounds(ContainerVT, Subtarget);
-  if (MinVLMAX == MaxVLMAX && NumElts == MinVLMAX)
-    return DAG.getRegister(RISCV::X0, Subtarget.getXLenVT());
-
+static SDValue getVLOp(uint64_t NumElts, const SDLoc &DL, SelectionDAG &DAG,
+                       const RISCVSubtarget &Subtarget) {
   return DAG.getConstant(NumElts, DL, Subtarget.getXLenVT());
 }
 
@@ -2618,7 +2610,7 @@ static std::pair<SDValue, SDValue>
 getDefaultVLOps(uint64_t NumElts, MVT ContainerVT, const SDLoc &DL,
                 SelectionDAG &DAG, const RISCVSubtarget &Subtarget) {
   assert(ContainerVT.isScalableVector() && "Expecting scalable container type");
-  SDValue VL = getVLOp(NumElts, ContainerVT, DL, DAG, Subtarget);
+  SDValue VL = getVLOp(NumElts, DL, DAG, Subtarget);
   SDValue Mask = getAllOnesMask(ContainerVT, VL, DL, DAG);
   return {Mask, VL};
 }
@@ -8872,8 +8864,7 @@ SDValue RISCVTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
     MVT VT = Op->getSimpleValueType(0);
     MVT ContainerVT = getContainerForFixedLengthVector(VT);
 
-    SDValue VL = getVLOp(VT.getVectorNumElements(), ContainerVT, DL, DAG,
-                         Subtarget);
+    SDValue VL = getVLOp(VT.getVectorNumElements(), DL, DAG, Subtarget);
     SDValue IntID = DAG.getTargetConstant(VlsegInts[NF - 2], DL, XLenVT);
     auto *Load = cast<MemIntrinsicSDNode>(Op);
     SmallVector<EVT, 9> ContainerVTs(NF, ContainerVT);
@@ -8993,8 +8984,7 @@ SDValue RISCVTargetLowering::LowerINTRINSIC_VOID(SDValue Op,
     MVT VT = Op->getOperand(2).getSimpleValueType();
     MVT ContainerVT = getContainerForFixedLengthVector(VT);
 
-    SDValue VL = getVLOp(VT.getVectorNumElements(), ContainerVT, DL, DAG,
-                         Subtarget);
+    SDValue VL = getVLOp(VT.getVectorNumElements(), DL, DAG, Subtarget);
     SDValue IntID = DAG.getTargetConstant(VssegInts[NF - 2], DL, XLenVT);
     SDValue Ptr = Op->getOperand(NF + 2);
 
@@ -9498,7 +9488,7 @@ SDValue RISCVTargetLowering::lowerINSERT_SUBVECTOR(SDValue Op,
     // Set the vector length to only the number of elements we care about. Note
     // that for slideup this includes the offset.
     unsigned EndIndex = OrigIdx + SubVecVT.getVectorNumElements();
-    SDValue VL = getVLOp(EndIndex, ContainerVT, DL, DAG, Subtarget);
+    SDValue VL = getVLOp(EndIndex, DL, DAG, Subtarget);
 
     // Use tail agnostic policy if we're inserting over Vec's tail.
     unsigned Policy = RISCVII::TAIL_UNDISTURBED_MASK_UNDISTURBED;
@@ -9675,8 +9665,7 @@ SDValue RISCVTargetLowering::lowerEXTRACT_SUBVECTOR(SDValue Op,
         getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget).first;
     // Set the vector length to only the number of elements we care about. This
     // avoids sliding down elements we're going to discard straight away.
-    SDValue VL = getVLOp(SubVecVT.getVectorNumElements(), ContainerVT, DL, DAG,
-                         Subtarget);
+    SDValue VL = getVLOp(SubVecVT.getVectorNumElements(), DL, DAG, Subtarget);
     SDValue SlidedownAmt = DAG.getConstant(OrigIdx, DL, XLenVT);
     SDValue Slidedown =
         getVSlidedown(DAG, Subtarget, DL, ContainerVT,
@@ -10096,7 +10085,7 @@ RISCVTargetLowering::lowerFixedLengthVectorLoadToRVV(SDValue Op,
     return DAG.getMergeValues({Result, NewLoad.getValue(1)}, DL);
   }
 
-  SDValue VL = getVLOp(VT.getVectorNumElements(), ContainerVT, DL, DAG, Subtarget);
+  SDValue VL = getVLOp(VT.getVectorNumElements(), DL, DAG, Subtarget);
 
   bool IsMaskOp = VT.getVectorElementType() == MVT::i1;
   SDValue IntID = DAG.getTargetConstant(
@@ -10153,8 +10142,7 @@ RISCVTargetLowering::lowerFixedLengthVectorStoreToRVV(SDValue Op,
     return DAG.getStore(Store->getChain(), DL, NewValue, Store->getBasePtr(),
                         Store->getMemOperand());
 
-  SDValue VL = getVLOp(VT.getVectorNumElements(), ContainerVT, DL, DAG,
-                       Subtarget);
+  SDValue VL = getVLOp(VT.getVectorNumElements(), DL, DAG, Subtarget);
 
   bool IsMaskOp = VT.getVectorElementType() == MVT::i1;
   SDValue IntID = DAG.getTargetConstant(

llvm/lib/Target/RISCV/RISCVInsertVSETVLI.cpp

@@ -858,13 +858,14 @@ static VSETVLIInfo computeInfoForInstr(const MachineInstr &MI, uint64_t TSFlags,
 
   if (RISCVII::hasVLOp(TSFlags)) {
     const MachineOperand &VLOp = MI.getOperand(getVLOpNum(MI));
+    const unsigned VLMAX = computeVLMAX(ST.getRealMaxVLen(), SEW, VLMul);
+
     if (VLOp.isImm()) {
       int64_t Imm = VLOp.getImm();
       // Conver the VLMax sentintel to X0 register.
       if (Imm == RISCV::VLMaxSentinel) {
         // If we know the exact VLEN, see if we can use the constant encoding
         // for the VLMAX instead.  This reduces register pressure slightly.
-        const unsigned VLMAX = computeVLMAX(ST.getRealMaxVLen(), SEW, VLMul);
         if (ST.getRealMinVLen() == ST.getRealMaxVLen() && VLMAX <= 31)
           InstrInfo.setAVLImm(VLMAX);
         else
@@ -873,7 +874,18 @@ static VSETVLIInfo computeInfoForInstr(const MachineInstr &MI, uint64_t TSFlags,
       else
         InstrInfo.setAVLImm(Imm);
     } else {
-      InstrInfo.setAVLReg(VLOp.getReg());
+      Register Reg = VLOp.getReg();
+      // If VL is a register, it may be a materialized constant that didn't fit
+      // into an uimm5. If we also know the exact VLEN, and the VL is equal to
+      // the exact VLEN, use the X0 encoding so we don't need the ADDI.
+      // doLocalPostpass will remove the ADDI if it's dead.
+      if (ST.getRealMinVLen() == ST.getRealMaxVLen() &&
+          VLOp.getReg().isVirtual())
+        if (auto *VLDef = MRI->getVRegDef(VLOp.getReg());
+            VLDef && isNonZeroLoadImmediate(*VLDef) &&
+            VLDef->getOperand(2).getImm() == VLMAX)
+          Reg = RISCV::X0;
+      InstrInfo.setAVLReg(Reg);
     }
   } else {
     assert(isScalarExtractInstr(MI));