[RISCV][VLOPT] Add support for checkUsers when UserMI is an reduction

Reductions are weird because for some operands, they are vector registers but
only read the first lane. For these operands, we do not need to check to make
sure the EEW and EMUL ratios match. However, we need to make sure that when
the reduction instruction has a non-zero VL operand, we don't try and set the
CommonVL=0. Since this is an edge case, we decide just not to optimize anything
when this occurs.

Author: michaelmaitland

llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp

@@ -621,6 +621,23 @@ static OperandInfo getOperandInfo(const MachineOperand &MO,
     return OperandInfo(MIVLMul, MILog2SEW);
   }
 
+  // Vector Reduction Operations
+  // Vector Single-Width Integer Reduction Instructions
+  // The Dest and VS1 only read element 0 of the vector register. Return unknown
+  // for these. VS2 has EEW=SEW and EMUL=LMUL.
+  case RISCV::VREDAND_VS:
+  case RISCV::VREDMAX_VS:
+  case RISCV::VREDMAXU_VS:
+  case RISCV::VREDMIN_VS:
+  case RISCV::VREDMINU_VS:
+  case RISCV::VREDOR_VS:
+  case RISCV::VREDSUM_VS:
+  case RISCV::VREDXOR_VS: {
+    if (MO.getOperandNo() == 2)
+      return OperandInfo(MIVLMul, MILog2SEW);
+    return {};
+  }
+
   default:
     return {};
   }
@@ -943,11 +960,28 @@ bool RISCVVLOptimizer::checkUsers(const MachineOperand *&CommonVL,
 
     // Instructions like reductions may use a vector register as a scalar
     // register. In this case, we should treat it like a scalar register which
-    // does not impact the decision on whether to optimize VL.
-    // TODO: Treat it like a scalar register instead of bailing out.
+    // does not impact the decision on whether to optimize VL. But if there is
+    // another user of MI and it has VL=0, we need to be sure not to reduce the
+    // VL of MI to zero when the VLOp of UserOp is may be non-zero.
     if (isVectorOpUsedAsScalarOp(UserOp)) {
-      CanReduceVL = false;
-      break;
+      [[maybe_unused]] Register R = UserOp.getReg();
+      [[maybe_unused]] const TargetRegisterClass *RC = MRI->getRegClass(R);
+      assert(RISCV::VRRegClass.hasSubClassEq(RC) &&
+             "Expect LMUL 1 register class for vector as scalar operands!");
+      LLVM_DEBUG(dbgs() << "    Used this operand as a scalar operand\n");
+      const MCInstrDesc &Desc = UserMI.getDesc();
+      unsigned VLOpNum = RISCVII::getVLOpNum(Desc);
+      const MachineOperand &VLOp = UserMI.getOperand(VLOpNum);
+      if ((VLOp.isReg() && VLOp.getReg() != RISCV::X0) ||
+          (VLOp.isImm() && VLOp.getImm() != 0)) {
+        if (!CommonVL) {
+          CommonVL = &VLOp;
+          continue;
+        } else if (!CommonVL->isIdenticalTo(VLOp)) {
+          CanReduceVL = false;
+          break;
+        }
+      }
     }
 
     if (mayReadPastVL(UserMI)) {

llvm/test/CodeGen/RISCV/rvv/vl-opt-op-info.mir

@@ -1064,3 +1064,22 @@ body: |
     %x:vr = PseudoVMAND_MM_B1 $noreg, $noreg, -1, 0
     %y:vr = PseudoVIOTA_M_MF2 $noreg, %x, 1, 3 /* e8 */, 0
 ...
+name: vred_vs2
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: vred_vs2
+    ; CHECK: %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, 1, 3 /* e8 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %y:vr = PseudoVREDAND_VS_M1_E8 $noreg, %x, $noreg, 1, 3 /* e8 */, 0 /* tu, mu */
+    %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 3 /* e8 */, 0
+    %y:vr = PseudoVREDAND_VS_M1_E8 $noreg, %x, $noreg, 1, 3 /* e8 */, 0
+...
+---
+name: vred_vs1
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: vred_vs1
+    ; CHECK: %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, 1, 3 /* e8 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %y:vr = PseudoVREDAND_VS_M1_E8 $noreg, $noreg, %x, 1, 3 /* e8 */, 0 /* tu, mu */
+    %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 3 /* e8 */, 0
+    %y:vr = PseudoVREDAND_VS_M1_E8 $noreg, $noreg, %x, 1, 3 /* e8 */, 0
+...

llvm/test/CodeGen/RISCV/rvv/vl-opt.mir

@@ -38,7 +38,7 @@ name: use_largest_common_vl_imm_imm
 body: |
   bb.0:
     ; CHECK-LABEL: name: use_largest_common_vl_imm_imm
-    ; CHECK: %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, 2, 3 /* e8 */, 0 /* tu, mu */
+    ; CHECK: %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 3 /* e8 */, 0 /* tu, mu */
     ; CHECK-NEXT: %y:vr = PseudoVADD_VV_M1 $noreg, %x, $noreg, 1, 3 /* e8 */, 0 /* tu, mu */
     ; CHECK-NEXT: %z:vr = PseudoVADD_VV_M1 $noreg, %x, $noreg, 2, 3 /* e8 */, 0 /* tu, mu */
     %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 3 /* e8 */, 0
@@ -110,4 +110,15 @@ body: |
     %y:vr = PseudoVADD_VV_M1 $noreg, %x, $noreg, 1, 3 /* e8 */, 0
     %z:vr = PseudoVADD_VV_M1 $noreg, %x, $noreg, -1, 3 /* e8 */, 0
 ...
-
+---
+name: vred_other_user_is_vl0
+body: |
+  bb.0:
+    ; CHECK-LABEL: name: vred_other_user_is_vl0
+    ; CHECK: %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 3 /* e8 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %y:vr = PseudoVREDSUM_VS_M1_E8 $noreg, $noreg, %x, 1, 3 /* e8 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %z:vr = PseudoVADD_VV_M1 $noreg, %x, $noreg, 0, 3 /* e8 */, 0 /* tu, mu */
+    %x:vr = PseudoVADD_VV_M1 $noreg, $noreg, $noreg, -1, 3 /* e8 */, 0
+    %y:vr = PseudoVREDSUM_VS_M1_E8 $noreg, $noreg, %x, 1, 3 /* e8 */, 0
+    %z:vr = PseudoVADD_VV_M1 $noreg, %x, $noreg, 0, 3 /* e8 */, 0
+...