[RISCV] Lower (vector_interleave X, undef) to (vzext_vl X).

[topperc]

If the odd vector is undef or poison, the widening add and multiply trick
doesn't work unless we freeze the odd vector.

Unfortunately, freezing doesn't work when the operand is provably
undef/poison. MIR doesn't have a representation for freeze so it
just becomes a COPY from IMPLICIT_DEF which freely propagates undef
to each operand independently.

To work around this, check for undef explicitly and lower to a VZEXT_VL
of the even vector. This produces better code than we'd get from a
freeze anyway.

I've left a FIXME for adding a freeze. I'll do that as a separate patch
as it affects other tests and doesn't help with the new test.

[llvmbot]

@llvm/pr-subscribers-backend-risc-v

Author: Craig Topper (topperc)

Changes

If the odd vector is undef or poison, the widening add and multiply trick
doesn't work unless we freeze the odd vector.

Unfortunately, freezing doesn't work when the operand is provably
undef/poison. MIR doesn't have a representation for freeze so it
just becomes a COPY from IMPLICIT_DEF which freely propagates undef
to each operand independently.

To work around this, check for undef explicitly and lower to a VZEXT_VL
of the even vector. This produces better code than we'd get from a
freeze anyway.

I've left a FIXME for adding a freeze. I'll do that as a separate patch
as it affects other tests and doesn't help with the new test.

---

Full diff: https://github.com/llvm/llvm-project/pull/87283.diff

2 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVISelLowering.cpp (+10-1)
• (modified) llvm/test/CodeGen/RISCV/rvv/vector-interleave.ll (+21)

diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index f693cbd3bea51e..c37938cd9559f0 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -4624,7 +4624,13 @@ static SDValue getWideningInterleave(SDValue EvenV, SDValue OddV,
   SDValue Passthru = DAG.getUNDEF(WideContainerVT);
 
   SDValue Interleaved;
-  if (Subtarget.hasStdExtZvbb()) {
+  if (OddV.isUndef()) {
+    // If OddV is undef, this is a zero extend.
+    // FIXME: Not only does this optimize the code, it fixes some correctness
+    // issues because MIR does not have freeze.
+    Interleaved = DAG.getNode(RISCVISD::VZEXT_VL, DL, WideContainerVT, EvenV,
+                              Mask, VL);
+  } else if (Subtarget.hasStdExtZvbb()) {
     // Interleaved = (OddV << VecVT.getScalarSizeInBits()) + EvenV.
     SDValue OffsetVec =
         DAG.getSplatVector(VecContainerVT, DL,
@@ -4635,6 +4641,9 @@ static SDValue getWideningInterleave(SDValue EvenV, SDValue OddV,
     Interleaved = DAG.getNode(RISCVISD::VWADDU_W_VL, DL, WideContainerVT,
                               Interleaved, EvenV, Passthru, Mask, VL);
   } else {
+    // FIXME: We should freeze the odd vector here. We already handled the case
+    // of provably undef/poison above.
+
     // Widen EvenV and OddV with 0s and add one copy of OddV to EvenV with
     // vwaddu.vv
     Interleaved = DAG.getNode(RISCVISD::VWADDU_VL, DL, WideContainerVT, EvenV,
diff --git a/llvm/test/CodeGen/RISCV/rvv/vector-interleave.ll b/llvm/test/CodeGen/RISCV/rvv/vector-interleave.ll
index 1acc0fec8fe586..7068e044dfc4f4 100644
--- a/llvm/test/CodeGen/RISCV/rvv/vector-interleave.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/vector-interleave.ll
@@ -656,6 +656,27 @@ define <vscale x 16 x double> @vector_interleave_nxv16f64_nxv8f64(<vscale x 8 x
   ret <vscale x 16 x double> %res
 }
 
+; FIXME: The last operand to the vwaddu.vv and vwmaccu.vx are both undef. They
+; need to be the same register with the same contents. Otherwise, the even
+; elements will not contain just the values from %a.
+define <vscale x 8 x i32> @vector_interleave_nxv8i32_nxv4i32_poison(<vscale x 4 x i32> %a) {
+; CHECK-LABEL: vector_interleave_nxv8i32_nxv4i32_poison:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e64, m4, ta, ma
+; CHECK-NEXT:    vzext.vf2 v12, v8
+; CHECK-NEXT:    vmv.v.v v8, v12
+; CHECK-NEXT:    ret
+;
+; ZVBB-LABEL: vector_interleave_nxv8i32_nxv4i32_poison:
+; ZVBB:       # %bb.0:
+; ZVBB-NEXT:    vsetvli a0, zero, e64, m4, ta, ma
+; ZVBB-NEXT:    vzext.vf2 v12, v8
+; ZVBB-NEXT:    vmv.v.v v8, v12
+; ZVBB-NEXT:    ret
+  %res = call <vscale x 8 x i32> @llvm.experimental.vector.interleave2.nxv8i32(<vscale x 4 x i32> %a, <vscale x 4 x i32> poison)
+  ret <vscale x 8 x i32> %res
+}
+
 declare <vscale x 64 x half> @llvm.experimental.vector.interleave2.nxv64f16(<vscale x 32 x half>, <vscale x 32 x half>)
 declare <vscale x 32 x float> @llvm.experimental.vector.interleave2.nxv32f32(<vscale x 16 x float>, <vscale x 16 x float>)
 declare <vscale x 16 x double> @llvm.experimental.vector.interleave2.nxv16f64(<vscale x 8 x double>, <vscale x 8 x double>)

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[lukel97]

LGTM

[kito-cheng]

And maybe (vector_interleave undef, X) to (vsll (vzext_vl X) sew_of_x) for a separate patch?

[topperc]

Committed as 8c1dc5d and a9af66a

[preames]

And maybe (vector_interleave undef, X) to (vsll (vzext_vl X) sew_of_x) for a separate patch?

I happened to be looking at this code today, and had a similar realization. If anyone has interest, I think this generalizes significantly.

This tactic can be generalized for an entire family of shuffles of the form:

• <a0, zero, a1, zero> -- this case
• <zero, a0, zero, a1> -- the vwsll case (also the one kito points out)
• <a0, zero, zero, zero, a1, zero, zero, zero> -- the zext.vf4 case

And the VWADD is just a specialization of the ISD::SELECT case when alternating lanes are zero in the two inputs.

Interestingly, the recursive reasoning would nearly make the special case two argument form redundant. The only bit left is the decision to "commit" undef to be zero.