[llvm] [RISCV] Eagerly optimize scalar packing for buildvector lowering (PR #156062)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

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

Author: Philip Reames (preames)

<details>
<summary>Changes</summary>

Instead of relying on DAG to cleanup the redundant ANDs, go ahead actively try to prove them redundant at construction.

I noticed this because SimplifyDemandedBits was dropping the disjoint on the OR, but there's at least one other codegen diff in the tests as well.

---
Full diff: https://github.com/llvm/llvm-project/pull/156062.diff


2 Files Affected:

- (modified) llvm/lib/Target/RISCV/RISCVISelLowering.cpp (+5-2) 
- (modified) llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll (+1-1) 


``````````diff
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index 4c39bcf8494a4..80083f843f9af 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -4257,6 +4257,7 @@ static SDValue lowerBuildVectorViaPacking(SDValue Op, SelectionDAG &DAG,
   MVT XLenVT = Subtarget.getXLenVT();
   SDValue Mask = DAG.getConstant(
       APInt::getLowBitsSet(XLenVT.getSizeInBits(), ElemSizeInBits), DL, XLenVT);
+  unsigned ZeroPrefix = XLenVT.getSizeInBits() - ElemSizeInBits;
   auto pack = [&](SDValue A, SDValue B) {
     // Bias the scheduling of the inserted operations to near the
     // definition of the element - this tends to reduce register
@@ -4270,8 +4271,10 @@ static SDValue lowerBuildVectorViaPacking(SDValue Op, SelectionDAG &DAG,
                              ElemDL, XLenVT, A, B),
           0);
 
-    A = DAG.getNode(ISD::AND, SDLoc(A), XLenVT, A, Mask);
-    B = DAG.getNode(ISD::AND, SDLoc(B), XLenVT, B, Mask);
+    if (DAG.computeKnownBits(A).countMinLeadingZeros() < ZeroPrefix)
+      A = DAG.getNode(ISD::AND, SDLoc(A), XLenVT, A, Mask);
+    if (DAG.computeKnownBits(B).countMinLeadingZeros() < ZeroPrefix)
+      B = DAG.getNode(ISD::AND, SDLoc(B), XLenVT, B, Mask);
     SDValue ShtAmt = DAG.getConstant(ElemSizeInBits, ElemDL, XLenVT);
     return DAG.getNode(ISD::OR, ElemDL, XLenVT, A,
                        DAG.getNode(ISD::SHL, ElemDL, XLenVT, B, ShtAmt),
diff --git a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll
index d9bb007a10f71..8aaffd41cbd16 100644
--- a/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/fixed-vectors-int-buildvec.ll
@@ -2629,7 +2629,7 @@ define <16 x i8> @buildvec_v16i8_undef_edges(ptr %p) {
 ; RVA22U64-NEXT:    or a0, a0, a4
 ; RVA22U64-NEXT:    slli a6, a6, 24
 ; RVA22U64-NEXT:    or a1, a1, a2
-; RVA22U64-NEXT:    add.uw a1, a6, a1
+; RVA22U64-NEXT:    or a1, a6, a1
 ; RVA22U64-NEXT:    or a0, a0, a3
 ; RVA22U64-NEXT:    vsetivli zero, 2, e64, m1, ta, ma
 ; RVA22U64-NEXT:    vmv.v.x v8, a1

``````````

</details>


https://github.com/llvm/llvm-project/pull/156062