[RISCV] Remove redundant SDNode creation for same reg class value

[4vtomat]

This fixes #114329 (comment)

[llvmbot]

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

Author: Brandon Wu (4vtomat)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/114348.diff

1 Files Affected:

• (modified) llvm/lib/Target/RISCV/RISCVISelLowering.cpp (+3-4)

diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index af7a39b2580a37..43de3d65de986c 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -21345,6 +21345,7 @@ bool RISCVTargetLowering::splitValueIntoRegisterParts(
   }
 
   if (ValueVT.isRISCVVectorTuple() && PartVT.isRISCVVectorTuple()) {
+#ifndef NDEBUG
     unsigned ValNF = ValueVT.getRISCVVectorTupleNumFields();
     [[maybe_unused]] unsigned ValLMUL =
         divideCeil(ValueVT.getSizeInBits(), ValNF * RISCV::RVVBitsPerBlock);
@@ -21352,11 +21353,9 @@ bool RISCVTargetLowering::splitValueIntoRegisterParts(
     [[maybe_unused]] unsigned PartLMUL =
         divideCeil(PartVT.getSizeInBits(), PartNF * RISCV::RVVBitsPerBlock);
     assert(ValNF == PartNF && ValLMUL == PartLMUL &&
-           "RISC-V vector tuple type only accepts same register class type "
-           "TUPLE_INSERT");
+           "RISC-V vector tuple type only accepts same register class copy");
+#endif
 
-    Val = DAG.getNode(RISCVISD::TUPLE_INSERT, DL, PartVT, DAG.getUNDEF(PartVT),
-                      Val, DAG.getVectorIdxConstant(0, DL));
     Parts[0] = Val;
     return true;
   }

[wangpc-pp]

LGTM.
(Please add a description that this patch fixes comment)

[topperc]

Is there ever case where PartVT isn't exactly equal to ValueVT? Why do we need all the math? If they aren't equal you would need at least a bitcast.

[4vtomat]

It's because of fractional LMUL, I think LMUL = 1 and fractional LMUL use the same register class, do we still need a bitcast?

[topperc]

If the types don't match then I think we need some operation to represent the type change. So I guess the original code is correct.

[4vtomat]

I see, let me close this PR.

[4vtomat]

We still need this "TUPLE_INSERT" since it's possible that ValueVT and PartVT has different type but has same register class such as fractional LMUL and LMUL = 1.