[Joel-Wwalker]

visitVECTOR_DEINTERLEAVE folded every result to undef when only operand 0 was undef. The operands of VECTOR_DEINTERLEAVE are consecutive pieces of the input (Result[I][J] = CONCAT_VECTORS(...)[I + N * J]), so every result takes lanes from all of them. In the reproducer from #238711 the low half of a deinterleave2 input is poison and the high half is a loaded vector, and the whole function was reduced to ret.

Check that all operands are undef, as visitVECTOR_INTERLEAVE does. The fold came from #228858, whose description already says "all undef inputs"; its deinterleave3_poison test is unchanged.

The first commit adds the tests with the current output, the second has the fix.

Fixes #238711.

Assisted by Claude (Anthropic).

[llvmorg-github-actions]

@llvm/pr-subscribers-llvm-selectiondag

@llvm/pr-subscribers-backend-aarch64

Author: Joel Walker (Joel-Wwalker)

Changes

visitVECTOR_DEINTERLEAVE folded every result to undef when only operand 0 was undef. The operands of VECTOR_DEINTERLEAVE are consecutive pieces of the input (Result[I][J] = CONCAT_VECTORS(...)[I + N * J]), so every result takes lanes from all of them. In the reproducer from #238711 the low half of a deinterleave2 input is poison and the high half is a loaded vector, and the whole function was reduced to ret.

Check that all operands are undef, as visitVECTOR_INTERLEAVE does. The fold came from #228858, whose description already says "all undef inputs"; its deinterleave3_poison test is unchanged.

The first commit adds the tests with the current output, the second has the fix.

Fixes #238711.

Assisted by Claude (Anthropic).

---

Full diff: https://github.com/llvm/llvm-project/pull/240057.diff

2 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+3-3)
• (modified) llvm/test/CodeGen/AArch64/fixed-vector-deinterleave.ll (+64)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 5b02ccb18ef23d..ff143ac71fcadf 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -28283,10 +28283,10 @@ SDValue DAGCombiner::visitVECTOR_INTERLEAVE(SDNode *N) {
 
 SDValue DAGCombiner::visitVECTOR_DEINTERLEAVE(SDNode *N) {
   EVT VT = N->getValueType(0);
-  SDValue Op0 = N->getOperand(0);
 
-  // Canonicalize shuffle undef -> {undef, undef, ..}
-  if (Op0.isUndef()) {
+  // Canonicalize shuffle undef, undef -> {undef, undef, ..}
+  // Every result takes lanes from every operand, so all operands must be undef.
+  if (all_of(N->op_values(), [](SDValue Op) { return Op.isUndef(); })) {
     SDLoc DL(N);
     SmallVector<SDValue> Ops(N->getNumValues(), DAG.getUNDEF(VT));
     return DAG.getMergeValues(Ops, DL);
diff --git a/llvm/test/CodeGen/AArch64/fixed-vector-deinterleave.ll b/llvm/test/CodeGen/AArch64/fixed-vector-deinterleave.ll
index ba8b2d4af20914..e24c16de6edd9f 100644
--- a/llvm/test/CodeGen/AArch64/fixed-vector-deinterleave.ll
+++ b/llvm/test/CodeGen/AArch64/fixed-vector-deinterleave.ll
@@ -873,3 +873,67 @@ define {<8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8
   %retval = call {<8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat>} @llvm.vector.deinterleave8.v64bf16(<64 x bfloat> %ins7)
   ret {<8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat>, <8 x bfloat>} %retval
 }
+
+; Only the low half of the deinterleaved vector is poison, so the results are
+; not poison: lanes 2 and 3 of each result come from %v.
+define void @deinterleave2_v8i8_lo_poison(ptr %p, ptr %o) {
+; CHECK-SD-LABEL: deinterleave2_v8i8_lo_poison:
+; CHECK-SD:       // %bb.0:
+; CHECK-SD-NEXT:    ldr s0, [x0]
+; CHECK-SD-NEXT:    ushll v0.8h, v0.8b, #0
+; CHECK-SD-NEXT:    uzp2 v1.4h, v0.4h, v0.4h
+; CHECK-SD-NEXT:    uzp1 v0.4h, v0.4h, v0.4h
+; CHECK-SD-NEXT:    zip2 v0.2s, v0.2s, v1.2s
+; CHECK-SD-NEXT:    uzp1 v0.8b, v0.8b, v0.8b
+; CHECK-SD-NEXT:    str s0, [x1]
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: deinterleave2_v8i8_lo_poison:
+; CHECK-GI:       // %bb.0:
+; CHECK-GI-NEXT:    ld1 { v0.s }[1], [x0]
+; CHECK-GI-NEXT:    adrp x8, .LCPI42_0
+; CHECK-GI-NEXT:    uzp1 v1.8b, v0.8b, v0.8b
+; CHECK-GI-NEXT:    uzp2 v0.8b, v0.8b, v0.8b
+; CHECK-GI-NEXT:    mov v1.d[1], v0.d[0]
+; CHECK-GI-NEXT:    ldr d0, [x8, :lo12:.LCPI42_0]
+; CHECK-GI-NEXT:    tbl v0.16b, { v1.16b }, v0.16b
+; CHECK-GI-NEXT:    str s0, [x1]
+; CHECK-GI-NEXT:    ret
+  %v = load <4 x i8>, ptr %p
+  %x = shufflevector <4 x i8> %v, <4 x i8> poison, <8 x i32> <i32 poison, i32 poison, i32 poison, i32 poison, i32 0, i32 1, i32 2, i32 3>
+  %d = call {<4 x i8>, <4 x i8>} @llvm.vector.deinterleave2.v8i8(<8 x i8> %x)
+  %even = extractvalue {<4 x i8>, <4 x i8>} %d, 0
+  %odd = extractvalue {<4 x i8>, <4 x i8>} %d, 1
+  %r = shufflevector <4 x i8> %even, <4 x i8> %odd, <4 x i32> <i32 2, i32 3, i32 6, i32 7>
+  store <4 x i8> %r, ptr %o
+  ret void
+}
+
+define <4 x i32> @deinterleave2_v8i32_lo_poison(<4 x i32> %v) {
+; CHECK-LABEL: deinterleave2_v8i32_lo_poison:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    uzp1 v0.4s, v0.4s, v0.4s
+; CHECK-NEXT:    ret
+  %x = shufflevector <4 x i32> %v, <4 x i32> poison, <8 x i32> <i32 poison, i32 poison, i32 poison, i32 poison, i32 0, i32 1, i32 2, i32 3>
+  %d = call {<4 x i32>, <4 x i32>} @llvm.vector.deinterleave2.v8i32(<8 x i32> %x)
+  %even = extractvalue {<4 x i32>, <4 x i32>} %d, 0
+  ret <4 x i32> %even
+}
+
+define <2 x i32> @deinterleave3_v6i32_first_poison(<2 x i32> %b, <2 x i32> %c) {
+; CHECK-LABEL: deinterleave3_v6i32_first_poison:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    sub sp, sp, #32
+; CHECK-NEXT:    .cfi_def_cfa_offset 32
+; CHECK-NEXT:    add x8, sp, #8
+; CHECK-NEXT:    stp d0, d1, [sp, #16]
+; CHECK-NEXT:    ld3 { v31.2s, v0.2s, v1.2s }, [x8]
+; CHECK-NEXT:    // kill: def $d0 killed $d0 killed $d31_d0_d1
+; CHECK-NEXT:    add sp, sp, #32
+; CHECK-NEXT:    ret
+  %bc = shufflevector <2 x i32> %b, <2 x i32> %c, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+  %x = shufflevector <4 x i32> %bc, <4 x i32> poison, <6 x i32> <i32 poison, i32 poison, i32 0, i32 1, i32 2, i32 3>
+  %d = call {<2 x i32>, <2 x i32>, <2 x i32>} @llvm.vector.deinterleave3.v6i32(<6 x i32> %x)
+  %r = extractvalue {<2 x i32>, <2 x i32>, <2 x i32>} %d, 1
+  ret <2 x i32> %r
+}