[llvm] X86: make VBMI2 funnel shifts use VSHLD/VSHRD for const splats (PR #169401)

[via llvm-commits]

https://github.com/ArnavM3434 updated https://github.com/llvm/llvm-project/pull/169401

>From d78dccc8c1f5b9d043582c4f60b5ab1546a87db5 Mon Sep 17 00:00:00 2001
From: Arnav Mehta <arnavnmehta1 at gmail.com>
Date: Mon, 24 Nov 2025 12:39:37 -0500
Subject: [PATCH] X86: make VBMI2 funnel shifts use VSHLD/VSHRD for const
 splats

Move constant splat handling for vector funnel shifts into a DAG combiner
so that VBMI2 legal widths emit VSHLD/VSHRD directly (fixes #166949).

Signed-off-by: Arnav Mehta <arnavnmehta1 at gmail.com>
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 62 +++++++++++++++++++------
 1 file changed, 48 insertions(+), 14 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index dc84025c166a3..ed193cc224bcf 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -2073,8 +2073,8 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
 
     if (Subtarget.hasVBMI2()) {
       for (auto VT : {MVT::v32i16, MVT::v16i32, MVT::v8i64}) {
-        setOperationAction(ISD::FSHL, VT, Custom);
-        setOperationAction(ISD::FSHR, VT, Custom);
+        setOperationAction(ISD::FSHL, VT, Legal);
+        setOperationAction(ISD::FSHR, VT, Legal);
       }
 
       setOperationAction(ISD::ROTL, MVT::v32i16, Custom);
@@ -2089,8 +2089,8 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
   if (!Subtarget.useSoftFloat() && Subtarget.hasVBMI2()) {
     for (auto VT : {MVT::v8i16, MVT::v4i32, MVT::v2i64, MVT::v16i16, MVT::v8i32,
                     MVT::v4i64}) {
-      setOperationAction(ISD::FSHL, VT, Custom);
-      setOperationAction(ISD::FSHR, VT, Custom);
+      setOperationAction(ISD::FSHL, VT, Subtarget.hasVLX() ? Legal : Custom);
+      setOperationAction(ISD::FSHR, VT, Subtarget.hasVLX() ? Legal : Custom);
     }
   }
 
@@ -2703,6 +2703,8 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
                        ISD::STRICT_FP_EXTEND,
                        ISD::FP_ROUND,
                        ISD::STRICT_FP_ROUND,
+                       ISD::FSHL,
+                       ISD::FSHR,
                        ISD::INTRINSIC_VOID,
                        ISD::INTRINSIC_WO_CHAIN,
                        ISD::INTRINSIC_W_CHAIN});
@@ -31314,19 +31316,15 @@ static SDValue LowerFunnelShift(SDValue Op, const X86Subtarget &Subtarget,
     bool IsCstSplat = X86::isConstantSplat(Amt, APIntShiftAmt);
     unsigned NumElts = VT.getVectorNumElements();
 
-    if (Subtarget.hasVBMI2() && EltSizeInBits > 8) {
-
-      if (IsCstSplat) {
-        if (IsFSHR)
-          std::swap(Op0, Op1);
-        uint64_t ShiftAmt = APIntShiftAmt.urem(EltSizeInBits);
-        SDValue Imm = DAG.getTargetConstant(ShiftAmt, DL, MVT::i8);
-        return getAVX512Node(IsFSHR ? X86ISD::VSHRD : X86ISD::VSHLD, DL, VT,
-                             {Op0, Op1, Imm}, DAG, Subtarget);
-      }
+    // For non-VLX VBMI2 targets, widen 128/256-bit to 512-bit so
+    // the rest of the lowering/isel can select the VBMI2 forms.
+    // Only Custom types (v8i16, v4i32, v2i64, v16i16, v8i32, v4i64) can
+    // reach LowerFunnelShift with VBMI2 but no VLX, so no type check needed.
+    if (Subtarget.hasVBMI2() && !Subtarget.hasVLX() && EltSizeInBits > 8) {
       return getAVX512Node(IsFSHR ? ISD::FSHR : ISD::FSHL, DL, VT,
                            {Op0, Op1, Amt}, DAG, Subtarget);
     }
+
     assert((VT == MVT::v16i8 || VT == MVT::v32i8 || VT == MVT::v64i8 ||
             VT == MVT::v8i16 || VT == MVT::v16i16 || VT == MVT::v32i16 ||
             VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32) &&
@@ -57624,6 +57622,40 @@ static SDValue combineFP_TO_xINT_SAT(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+// Combiner: turn uniform-constant splat funnel shifts into VSHLD/VSHRD
+static SDValue combineFunnelShift(SDNode *N, SelectionDAG &DAG,
+                                  TargetLowering::DAGCombinerInfo &DCI,
+                                  const X86Subtarget &Subtarget) {
+  SDLoc DL(N);
+  SDValue Op0 = N->getOperand(0);
+  SDValue Op1 = N->getOperand(1);
+  SDValue Amt = N->getOperand(2);
+  EVT VT = Op0.getValueType();
+
+  if (!VT.isVector())
+    return SDValue();
+
+  // Only combine if the operation is legal for this type.
+  // This ensures we don't try to convert types that need to be
+  // widened/promoted.
+  if (!DAG.getTargetLoweringInfo().isOperationLegal(N->getOpcode(), VT))
+    return SDValue();
+
+  unsigned EltSize = VT.getScalarSizeInBits();
+  APInt ShiftVal;
+  if (!X86::isConstantSplat(Amt, ShiftVal))
+    return SDValue();
+
+  uint64_t ModAmt = ShiftVal.urem(EltSize);
+  SDValue Imm = DAG.getTargetConstant(ModAmt, DL, MVT::i8);
+  bool IsFSHR = N->getOpcode() == ISD::FSHR;
+
+  if (IsFSHR)
+    std::swap(Op0, Op1);
+  unsigned Opcode = IsFSHR ? X86ISD::VSHRD : X86ISD::VSHLD;
+  return DAG.getNode(Opcode, DL, VT, {Op0, Op1, Imm});
+}
+
 static bool needCarryOrOverflowFlag(SDValue Flags) {
   assert(Flags.getValueType() == MVT::i32 && "Unexpected VT!");
 
@@ -61228,6 +61260,8 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::INTRINSIC_VOID:  return combineINTRINSIC_VOID(N, DAG, DCI);
   case ISD::FP_TO_SINT_SAT:
   case ISD::FP_TO_UINT_SAT: return combineFP_TO_xINT_SAT(N, DAG, Subtarget);
+  case ISD::FSHL:
+  case ISD::FSHR: return combineFunnelShift(N, DAG, DCI, Subtarget);
     // clang-format on
   }