SDValue AMDGPUTargetLowering::performSrlCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
- if (N->getValueType(0) != MVT::i64)
- return SDValue();
-
- const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
+ auto *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
if (!RHS)
return SDValue();
+ EVT VT = N->getValueType(0);
+ SDValue LHS = N->getOperand(0);
unsigned ShiftAmt = RHS->getZExtValue();
+ SelectionDAG &DAG = DCI.DAG;
+ SDLoc SL(N);
+
+ // fold (srl (and x, c1 << c2), c2) -> (and (srl(x, c2), c1)
+ // this improves the ability to match BFE patterns in isel.
+ if (LHS.getOpcode() == ISD::AND) {
+ if (auto *Mask = dyn_cast<ConstantSDNode>(LHS.getOperand(1))) {
+ if (Mask->getAPIntValue().isShiftedMask() &&
+ Mask->getAPIntValue().countTrailingZeros() == ShiftAmt) {
+ return DAG.getNode(
+ ISD::AND, SL, VT,
+ DAG.getNode(ISD::SRL, SL, VT, LHS.getOperand(0), N->getOperand(1)),
+ DAG.getNode(ISD::SRL, SL, VT, LHS.getOperand(1), N->getOperand(1)));
+ }
+ }
+ }
+
+ if (VT != MVT::i64)
+ return SDValue();
+
if (ShiftAmt < 32)
return SDValue();
// srl i64:x, C for C >= 32
// =>
// build_pair (srl hi_32(x), C - 32), 0
-
- SelectionDAG &DAG = DCI.DAG;
- SDLoc SL(N);
-
SDValue One = DAG.getConstant(1, SL, MVT::i32);
SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
- SDValue VecOp = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, N->getOperand(0));
- SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32,
- VecOp, One);
+ SDValue VecOp = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, LHS);
+ SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecOp, One);
SDValue NewConst = DAG.getConstant(ShiftAmt - 32, SL, MVT::i32);
SDValue NewShift = DAG.getNode(ISD::SRL, SL, MVT::i32, Hi, NewConst);
; GFX9-NEXT: v_mov_b32_e32 v2, s2
; GFX9-NEXT: ds_write_b16 v1, v2 offset:4
; GFX9-NEXT: s_waitcnt vmcnt(0)
-; GFX9-NEXT: v_and_b32_e32 v0, 0x7f0000, v0
-; GFX9-NEXT: ds_write_b8_d16_hi v1, v0 offset:6
+; GFX9-NEXT: v_bfe_u32 v0, v0, 16, 7
+; GFX9-NEXT: ds_write_b8 v1, v0 offset:6
; GFX9-NEXT: ds_write_b32 v1, v3
; GFX9-NEXT: s_endpgm
store i55 %arg, i55 addrspace(3)* %ptr, align 8