[DAG] Update canCreateUndefOrPoison to handle ISD::VECTOR_COMPRESS

[Michael-Chen-NJU]

Updates canCreateUndefOrPoison to return a precise result for ISD::VECTOR_COMPRESS, correctly reporting that it generates undef (in the tail) but not poison.

Fixes #167710

[llvmbot]

@llvm/pr-subscribers-backend-x86

@llvm/pr-subscribers-llvm-selectiondag

Author: 陈子昂 (Michael-Chen-NJU)

Changes

Updates canCreateUndefOrPoison to return a precise result for ISD::VECTOR_COMPRESS, correctly reporting that it generates undef (in the tail) but not poison.

Fixes #167710

---

Full diff: https://github.com/llvm/llvm-project/pull/168010.diff

1 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp (+5)

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index c2b4c19846316..38fee7c83e1c0 100644
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -5702,6 +5702,11 @@ bool SelectionDAG::canCreateUndefOrPoison(SDValue Op, const APInt &DemandedElts,
     return false;
   }
 
+  case ISD::VECTOR_COMPRESS:
+    // Return true only if undef is checked and at least one element is
+    // demanded.
+    return !PoisonOnly && !DemandedElts.isZero();
+
   default:
     // Allow the target to implement this method for its nodes.
     if (Opcode >= ISD::BUILTIN_OP_END || Opcode == ISD::INTRINSIC_WO_CHAIN ||

[arsenm]

Needs tests

[RKSimon]

@Michael-Chen-NJU You'll probably have to do value tracking style tests:

declare <16 x i32> @llvm.experimental.vector.compress.v16i32(<16 x i32>, <16 x i1>, <16 x i32>)

define <16 x i32> @src(<16 x i32> %a0, <16 x i32> %a1, <16 x i8> %a3) {
  %cmp = icmp sgt <16 x i32> %a0, %a1
  %ext = zext <16 x i8> %a3 to <16 x i32>
  %cpr = call <16 x i32> @llvm.experimental.vector.compress.v16i32(<16 x i32> %ext, <16 x i1> %cmp, <16 x i32> splat(i32 15))
  %fr = freeze <16 x i32> %cpr
  %and = and <16 x i32> %fr, splat (i32 15)
  ret <16 x i32> %and
}

define <16 x i32> @tgt(<16 x i32> %a0, <16 x i32> %a1, <16 x i8> %a3) {
  %cmp = icmp sgt <16 x i32> %a0, %a1
  %ext = zext <16 x i8> %a3 to <16 x i32>
  %cpr = call <16 x i32> @llvm.experimental.vector.compress.v16i32(<16 x i32> %ext, <16 x i1> %cmp, <16 x i32> splat(i32 15))
  %and = and <16 x i32> %cpr, splat (i32 255)
  ret <16 x i32> %and
}

[github-actions]

✅ With the latest revision this PR passed the undef deprecator.

[github-actions]

🐧 Linux x64 Test Results

• 186293 tests passed
• 4853 tests skipped

[RKSimon]

LGTM