[PATCH] D124684: [RISCV] Fix incorrect codegen for masked vmsge{u}.vx with mask agnostic.

[Zakk Chen via Phabricator via llvm-commits]

khchen created this revision.
khchen added reviewers: craig.topper, rogfer01, frasercrmck, kito-cheng, arcbbb, monkchiang, eopXD.
Herald added subscribers: sunshaoce, VincentWu, luke957, StephenFan, vkmr, evandro, luismarques, apazos, sameer.abuasal, s.egerton, Jim, benna, psnobl, jocewei, PkmX, the_o, brucehoult, MartinMosbeck, edward-jones, zzheng, jrtc27, niosHD, sabuasal, simoncook, johnrusso, rbar, asb, hiraditya, arichardson.
Herald added a project: All.
khchen requested review of this revision.
Herald added subscribers: llvm-commits, pcwang-thead, MaskRay.
Herald added a project: LLVM.

The result was totally wrong.
We could use mask undisturbed result to emulate the mask agnostic result.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D124684

Files:
  llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
  llvm/test/CodeGen/RISCV/rvv/masked-tama.ll


Index: llvm/test/CodeGen/RISCV/rvv/masked-tama.ll
===================================================================
--- llvm/test/CodeGen/RISCV/rvv/masked-tama.ll
+++ llvm/test/CodeGen/RISCV/rvv/masked-tama.ll
@@ -1431,7 +1431,9 @@
 ; RV64-LABEL: intrinsic_vmsge_mask_vx_nxv1i64_i64:
 ; RV64:       # %bb.0: # %entry
 ; RV64-NEXT:    vsetvli zero, a1, e64, m1, ta, ma
-; RV64-NEXT:    vmslt.vx v0, v8, a0, v0.t
+; RV64-NEXT:    vmslt.vx v8, v8, a0, v0.t
+; RV64-NEXT:    vsetvli zero, zero, e8, mf8, ta, mu
+; RV64-NEXT:    vmxor.mm v0, v8, v0
 ; RV64-NEXT:    ret
 entry:
   %a = call <vscale x 1 x i1> @llvm.riscv.vmsge.mask.nxv1i64.i64(
Index: llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
===================================================================
--- llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
+++ llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
@@ -1121,16 +1121,15 @@
       SDValue V0 = CurDAG->getRegister(RISCV::V0, VT);
 
       // Otherwise use
-      // vmslt{u}.vx vd, va, x, v0.t; if mask policy is agnostic.
+      // vmslt{u}.vx vd, va, x, v0.t; vmxor.mm vd, vd, v0
+      // The result is mask undisturbed.
+      // We use the same instructions to emulate mask agnostic behavior, because
+      // the agnostic result can be either undisturbed or all 1.
       SDValue Cmp = SDValue(
           CurDAG->getMachineNode(VMSLTMaskOpcode, DL, VT,
                                  {MaskedOff, Src1, Src2, V0, VL, SEW, Glue}),
           0);
-      if (MaskedOff.isUndef()) {
-        ReplaceNode(Node, Cmp.getNode());
-        return;
-      }
-      // Need vmxor.mm vd, vd, v0 to assign inactive value.
+      // vmxor.mm vd, vd, v0 is used to update active value.
       ReplaceNode(Node, CurDAG->getMachineNode(VMXOROpcode, DL, VT,
                                                {Cmp, Mask, VL, MaskSEW}));
       return;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D124684.426100.patch
Type: text/x-patch
Size: 1848 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220429/7b199311/attachment.bin>