[llvm] [RISCV] Lower (vector_interleave X, undef) to (vzext_vl X). (PR #87283)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

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

Author: Craig Topper (topperc)

<details>
<summary>Changes</summary>

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
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>)

``````````

</details>


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