[llvm] [SPARC] Lower dynamic v2i32 vector extracts on 32-bit SPARC (PR #222348)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-sparc

Author: Daniel Cederman (doac)

<details>
<summary>Changes</summary>

Lower dynamically indexed extracts after vector legalization so that instruction selection never receives a variable index.

Constant indexes use the existing TableGen patterns.

Fixes the following error:

LLVM ERROR: Cannot select: 0x5e7e87e5b590: i32 = extract_vector_elt 0x5e7e87e5b910, 0x5e7e87e5b0c0
  0x5e7e87e5b910: v2i32 = BUILD_VECTOR 0x5e7e87e5af70, 0x5e7e87e5b6e0
    0x5e7e87e5af70: i32,ch = CopyFromReg 0x5e7e87e24840, Register:i32 %0
      0x5e7e87e5b750: i32 = Register %0
    0x5e7e87e5b6e0: i32,ch = CopyFromReg 0x5e7e87e24840, Register:i32 %1
      0x5e7e87e5b7c0: i32 = Register %1
  0x5e7e87e5b0c0: i32,ch = CopyFromReg 0x5e7e87e24840, Register:i32 %2
    0x5e7e87e5afe0: i32 = Register %2
In function: extract_v2i8

Assisted-by: Codex

---
Full diff: https://github.com/llvm/llvm-project/pull/222348.diff


2 Files Affected:

- (modified) llvm/lib/Target/Sparc/SparcISelLowering.cpp (+19-1) 
- (modified) llvm/test/CodeGen/SPARC/vector-extract-elt.ll (+15) 


``````````diff
diff --git a/llvm/lib/Target/Sparc/SparcISelLowering.cpp b/llvm/lib/Target/Sparc/SparcISelLowering.cpp
index 66638f599559a..a1e97fe9ba99f 100644
--- a/llvm/lib/Target/Sparc/SparcISelLowering.cpp
+++ b/llvm/lib/Target/Sparc/SparcISelLowering.cpp
@@ -1687,7 +1687,7 @@ SparcTargetLowering::SparcTargetLowering(const TargetMachine &TM,
     // However, load and store *are* legal.
     setOperationAction(ISD::LOAD, MVT::v2i32, Legal);
     setOperationAction(ISD::STORE, MVT::v2i32, Legal);
-    setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i32, Legal);
+    setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i32, Custom);
     setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Legal);
 
     // And we need to promote i64 loads/stores into vector load/store
@@ -3169,6 +3169,22 @@ static SDValue LowerATOMIC_LOAD_STORE(SDValue Op, SelectionDAG &DAG) {
   return Op;
 }
 
+static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
+  SDLoc DL(Op);
+  SDValue Vec = Op.getOperand(0);
+  SDValue Idx = Op.getOperand(1);
+  assert(Vec.getSimpleValueType() == MVT::v2i32 && "Unexpected vector type");
+
+  // Constant indices are handled by the TableGen patterns.
+  if (isa<ConstantSDNode>(Idx))
+    return Op;
+
+  SDValue Even = DAG.getTargetExtractSubreg(SP::sub_even, DL, MVT::i32, Vec);
+  SDValue Odd = DAG.getTargetExtractSubreg(SP::sub_odd, DL, MVT::i32, Vec);
+  return DAG.getSelectCC(DL, Idx, DAG.getConstant(0, DL, Idx.getValueType()),
+                         Even, Odd, ISD::SETEQ);
+}
+
 SDValue SparcTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
                                                      SelectionDAG &DAG) const {
   unsigned IntNo = Op.getConstantOperandVal(0);
@@ -3239,6 +3255,8 @@ LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::FP_ROUND:           return LowerF128_FPROUND(Op, DAG, *this);
   case ISD::ATOMIC_LOAD:
   case ISD::ATOMIC_STORE:       return LowerATOMIC_LOAD_STORE(Op, DAG);
+  case ISD::EXTRACT_VECTOR_ELT:
+    return LowerEXTRACT_VECTOR_ELT(Op, DAG);
   case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
   }
 }
diff --git a/llvm/test/CodeGen/SPARC/vector-extract-elt.ll b/llvm/test/CodeGen/SPARC/vector-extract-elt.ll
index ef2a4323accce..e238e7a8f20a6 100644
--- a/llvm/test/CodeGen/SPARC/vector-extract-elt.ll
+++ b/llvm/test/CodeGen/SPARC/vector-extract-elt.ll
@@ -17,3 +17,18 @@ define i1 @test1(ptr %in) {
   %bool = icmp slt i32 %sum, 0
   ret i1 %bool
 }
+
+; A legal v2i32 is used internally to represent 64-bit integer register pairs.
+; Make sure a dynamic extract produced while legalizing a smaller vector does
+; not reach instruction selection, which only handles constant indices.
+define i8 @extract_v2i8(<2 x i8> %v, i32 %idx) {
+; CHECK-LABEL: extract_v2i8:
+; CHECK:       cmp %o2, 0
+; CHECK:       be
+; CHECK:       mov %o0, %o2
+; CHECK:       mov %o1, %o2
+; CHECK:       retl
+; CHECK:       mov %o2, %o0
+  %elt = extractelement <2 x i8> %v, i32 %idx
+  ret i8 %elt
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/222348