[RISCV] Fix incorrect optimization for masked vmsgeu.vi with 0 immediate.

[khchen]

Summary

vmsgeu.vi with 0 is always true, but in the masked with mask undisturbed
policy, we still need to keep inactive elelemt which come from maskedoff.

We could return mask directly if it's mask agnostic policy in the future.

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

1077

vor -> vmor

1079–1087

vor -> vmor

llvm/test/CodeGen/RISCV/rvv/vmsgeu-rv32.ll

2114

samm-> same?

llvm/test/CodeGen/RISCV/rvv/vmsgeu-rv64.ll

2096

samm -> same?

[khchen]

Fix typos. Thanks Craig.

[craig.topper]

LGTM

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

case Intrinsic::riscv_vmsge_mask: {
         if (CVal >= -15 && CVal <= 16) {
           if (!IsUnsigned || CVal != 0)
             break;
           IsCmpUnsignedZero = true;
         }
       }
       MVT Src1VT = Src1.getSimpleValueType();
       unsigned VMSLTOpcode, VMSLTMaskOpcode, VMXOROpcode, VMANDNOpcode,
-          VMSetOpcode, VMANDOpcode;
+          VMOROpcode;
       switch (RISCVTargetLowering::getLMUL(Src1VT)) {
       default:
         llvm_unreachable("Unexpected LMUL!");
-#define CASE_VMSLT_VMSET_OPCODES(lmulenum, suffix, suffix_b)                   \
+#define CASE_VMSLT_OPCODES(lmulenum, suffix, suffix_b)                         \
   case RISCVII::VLMUL::lmulenum:                                               \
     VMSLTOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix                 \
                              : RISCV::PseudoVMSLT_VX_##suffix;                 \
     VMSLTMaskOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix##_MASK      \
                                  : RISCV::PseudoVMSLT_VX_##suffix##_MASK;      \
-    VMSetOpcode = RISCV::PseudoVMSET_M_##suffix_b;                             \
     break;
-        CASE_VMSLT_VMSET_OPCODES(LMUL_F8, MF8, B1)
-        CASE_VMSLT_VMSET_OPCODES(LMUL_F4, MF4, B2)
-        CASE_VMSLT_VMSET_OPCODES(LMUL_F2, MF2, B4)
-        CASE_VMSLT_VMSET_OPCODES(LMUL_1, M1, B8)
-        CASE_VMSLT_VMSET_OPCODES(LMUL_2, M2, B16)
-        CASE_VMSLT_VMSET_OPCODES(LMUL_4, M4, B32)
-        CASE_VMSLT_VMSET_OPCODES(LMUL_8, M8, B64)
-#undef CASE_VMSLT_VMSET_OPCODES
+        CASE_VMSLT_OPCODES(LMUL_F8, MF8, B1)
+        CASE_VMSLT_OPCODES(LMUL_F4, MF4, B2)
+        CASE_VMSLT_OPCODES(LMUL_F2, MF2, B4)
+        CASE_VMSLT_OPCODES(LMUL_1, M1, B8)
+        CASE_VMSLT_OPCODES(LMUL_2, M2, B16)
+        CASE_VMSLT_OPCODES(LMUL_4, M4, B32)
+        CASE_VMSLT_OPCODES(LMUL_8, M8, B64)
+#undef CASE_VMSLT_OPCODES
       }
       // Mask operations use the LMUL from the mask type.
       switch (RISCVTargetLowering::getLMUL(VT)) {
       default:
         llvm_unreachable("Unexpected LMUL!");
-#define CASE_VMXOR_VMANDN_VMAND_OPCODES(lmulenum, suffix)                       \
+#define CASE_VMXOR_VMANDN_VMOR_OPCODES(lmulenum, suffix)                       \
   case RISCVII::VLMUL::lmulenum:                                               \
     VMXOROpcode = RISCV::PseudoVMXOR_MM_##suffix;                              \
     VMANDNOpcode = RISCV::PseudoVMANDN_MM_##suffix;                            \
-    VMANDOpcode = RISCV::PseudoVMAND_MM_##suffix;                              \
+    VMOROpcode = RISCV::PseudoVMOR_MM_##suffix;                                \
     break;
-        CASE_VMXOR_VMANDN_VMAND_OPCODES(LMUL_F8, MF8)
-        CASE_VMXOR_VMANDN_VMAND_OPCODES(LMUL_F4, MF4)
-        CASE_VMXOR_VMANDN_VMAND_OPCODES(LMUL_F2, MF2)
-        CASE_VMXOR_VMANDN_VMAND_OPCODES(LMUL_1, M1)
-        CASE_VMXOR_VMANDN_VMAND_OPCODES(LMUL_2, M2)
-        CASE_VMXOR_VMANDN_VMAND_OPCODES(LMUL_4, M4)
-        CASE_VMXOR_VMANDN_VMAND_OPCODES(LMUL_8, M8)
-#undef CASE_VMXOR_VMANDN_VMAND_OPCODES
+        CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_F8, MF8)
+        CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_F4, MF4)
+        CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_F2, MF2)
+        CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_1, M1)
+        CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_2, M2)
+        CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_4, M4)
+        CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_8, M8)
+#undef CASE_VMXOR_VMANDN_VMOR_OPCODES
       }
       SDValue SEW = CurDAG->getTargetConstant(
           Log2_32(Src1VT.getScalarSizeInBits()), DL, XLenVT);
       SDValue MaskSEW = CurDAG->getTargetConstant(0, DL, XLenVT);
       SDValue VL;
       selectVLOp(Node->getOperand(5), VL);
       SDValue MaskedOff = Node->getOperand(1);
       SDValue Mask = Node->getOperand(4);
 
-      // If vmsgeu_mask with 0 immediate, expand it to {vmset, vmand}.
+      // If vmsgeu_mask with 0 immediate, expand it to vmor mask, maskedoff.
       if (IsCmpUnsignedZero) {
-        SDValue VMSet =
-            SDValue(CurDAG->getMachineNode(VMSetOpcode, DL, VT, VL, SEW), 0);
-        ReplaceNode(Node, CurDAG->getMachineNode(VMANDOpcode, DL, VT,
-                                                 {Mask, VMSet, VL, MaskSEW}));
+        // We don't need vmor if the MaskedOff and the Mask are the same
+        // value.
+        if (Mask == MaskedOff) {
+          ReplaceUses(Node, Mask.getNode());
+          return;
+        }
+        ReplaceNode(Node,
+                    CurDAG->getMachineNode(VMOROpcode, DL, VT,
+                                           {Mask, MaskedOff, VL, MaskSEW}));
         return;
       }
 
       // If the MaskedOff value and the Mask are the same value use
       // vmslt{u}.vx vt, va, x;  vmandn.mm vd, vd, vt
       // This avoids needing to copy v0 to vd before starting the next sequence.
       if (Mask == MaskedOff) {
         SDValue Cmp = SDValue(

llvm/test/CodeGen/RISCV/rvv/vmsgeu-rv32.ll

%a = call <vscale x 2 x i1> @llvm.riscv.vmsgeu.nxv2i16.i16(
     i32 %1)
 
   ret <vscale x 2 x i1> %a
 }
 
 define <vscale x 2 x i1> @intrinsic_vmsgeu_mask_vi_nxv2i16_i16(<vscale x 2 x i1> %0, <vscale x 2 x i16> %1, <vscale x 2 x i1> %2, i32 %3) nounwind {
 ; CHECK-LABEL: intrinsic_vmsgeu_mask_vi_nxv2i16_i16:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vsetvli zero, a0, e16, mf2, ta, mu
-; CHECK-NEXT:    vmset.m v8
-; CHECK-NEXT:    vmand.mm v0, v9, v8
+; CHECK-NEXT:    vsetvli zero, a0, e8, mf4, ta, mu
+; CHECK-NEXT:    vmor.mm v0, v9, v0
 ; CHECK-NEXT:    ret
 entry:
   %a = call <vscale x 2 x i1> @llvm.riscv.vmsgeu.mask.nxv2i16.i16(
     <vscale x 2 x i1> %0,
     <vscale x 2 x i16> %1,
     i16 0,
     <vscale x 2 x i1> %2,
     i32 %3)
 
   ret <vscale x 2 x i1> %a
 }
 
+define <vscale x 2 x i1> @intrinsic_vmsgeu_mask_vi_nxv2i16_i16_same_mask_maskedoff(<vscale x 2 x i1> %0, <vscale x 2 x i16> %1, i32 %2) nounwind {
+; CHECK-LABEL: intrinsic_vmsgeu_mask_vi_nxv2i16_i16_same_mask_maskedoff:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 2 x i1> @llvm.riscv.vmsgeu.mask.nxv2i16.i16(
+    <vscale x 2 x i1> %0,
+    <vscale x 2 x i16> %1,
+    i16 0,
+    <vscale x 2 x i1> %0,
+    i32 %2)
+
+  ret <vscale x 2 x i1> %a
+}
+
 define <vscale x 4 x i1> @intrinsic_vmsgeu_vi_nxv4i16_i16(<vscale x 4 x i16> %0, i32 %1) nounwind {
 ; CHECK-LABEL: intrinsic_vmsgeu_vi_nxv4i16_i16:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    vsetvli zero, a0, e16, m1, ta, mu
 ; CHECK-NEXT:    vmset.m v0
 ; CHECK-NEXT:    ret
 entry:
   %a = call <vscale x 4 x i1> @llvm.riscv.vmsgeu.nxv4i16.i16(

llvm/test/CodeGen/RISCV/rvv/vmsgeu-rv64.ll

%a = call <vscale x 2 x i1> @llvm.riscv.vmsgeu.nxv2i16.i16(
     i64 %1)
 
   ret <vscale x 2 x i1> %a
 }
 
 define <vscale x 2 x i1> @intrinsic_vmsgeu_mask_vi_nxv2i16_i16(<vscale x 2 x i1> %0, <vscale x 2 x i16> %1, <vscale x 2 x i1> %2, i64 %3) nounwind {
 ; CHECK-LABEL: intrinsic_vmsgeu_mask_vi_nxv2i16_i16:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vsetvli zero, a0, e16, mf2, ta, mu
-; CHECK-NEXT:    vmset.m v8
-; CHECK-NEXT:    vmand.mm v0, v9, v8
+; CHECK-NEXT:    vsetvli zero, a0, e8, mf4, ta, mu
+; CHECK-NEXT:    vmor.mm v0, v9, v0
 ; CHECK-NEXT:    ret
 entry:
   %a = call <vscale x 2 x i1> @llvm.riscv.vmsgeu.mask.nxv2i16.i16(
     <vscale x 2 x i1> %0,
     <vscale x 2 x i16> %1,
     i16 0,
     <vscale x 2 x i1> %2,
     i64 %3)
entry:
   %a = call <vscale x 4 x i1> @llvm.riscv.vmsgeu.nxv4i16.i16(
     <vscale x 4 x i16> %0,
     i16 0,
     i64 %1)
 
   ret <vscale x 4 x i1> %a
 }
 
+define <vscale x 2 x i1> @intrinsic_vmsgeu_mask_vi_nxv2i16_i16_same_mask_maskedoff(<vscale x 2 x i1> %0, <vscale x 2 x i16> %1, i64 %2) nounwind {
+; CHECK-LABEL: intrinsic_vmsgeu_mask_vi_nxv2i16_i16_same_mask_maskedoff:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    ret
+entry:
+  %a = call <vscale x 2 x i1> @llvm.riscv.vmsgeu.mask.nxv2i16.i16(
+    <vscale x 2 x i1> %0,
+    <vscale x 2 x i16> %1,
+    i16 0,
+    <vscale x 2 x i1> %0,
+    i64 %2)
+
+  ret <vscale x 2 x i1> %a
+}
+
 define <vscale x 4 x i1> @intrinsic_vmsgeu_mask_vi_nxv4i16_i16(<vscale x 4 x i1> %0, <vscale x 4 x i16> %1, <vscale x 4 x i1> %2, i64 %3) nounwind {
 ; CHECK-LABEL: intrinsic_vmsgeu_mask_vi_nxv4i16_i16:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    vmv1r.v v10, v0
 ; CHECK-NEXT:    vsetvli zero, a0, e16, m1, ta, mu
 ; CHECK-NEXT:    vmv1r.v v0, v9
 ; CHECK-NEXT:    vmsgtu.vi v10, v8, 0, v0.t
 ; CHECK-NEXT:    vmv.v.v v0, v10