PPC: Combine duplicate (offset) lvsl Altivec intrinsics
The lvsl permutation control instruction is a function only of the alignment of
the pointer operand (relative to the 16-byte natural alignment of Altivec
vectors). As a result, multiple lvsl intrinsics where the operands differ by a
multiple of 16 can be combined.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@182708 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/PowerPC/PPCISelLowering.cpp b/lib/Target/PowerPC/PPCISelLowering.cpp
index 6c8af0c..f47376c 100644
--- a/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -540,6 +540,7 @@
setTargetDAGCombine(ISD::STORE);
setTargetDAGCombine(ISD::BR_CC);
setTargetDAGCombine(ISD::BSWAP);
+ setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
// Use reciprocal estimates.
if (TM.Options.UnsafeFPMath) {
@@ -6988,8 +6989,10 @@
// cause the last vector in the sequence to be (re)loaded. Otherwise,
// the next vector will be fetched as you might suspect was necessary.
- // FIXME: We might be able to reuse the permutation generation from
+ // We might be able to reuse the permutation generation from
// a different base address offset from this one by an aligned amount.
+ // The INTRINSIC_WO_CHAIN DAG combine will attempt to perform this
+ // optimization later.
SDValue PermCntl = BuildIntrinsicOp(Intrinsic::ppc_altivec_lvsl, Ptr,
DAG, dl, MVT::v16i8);
@@ -7074,6 +7077,30 @@
}
}
break;
+ case ISD::INTRINSIC_WO_CHAIN:
+ if (cast<ConstantSDNode>(N->getOperand(0))->getZExtValue() ==
+ Intrinsic::ppc_altivec_lvsl &&
+ N->getOperand(1)->getOpcode() == ISD::ADD) {
+ SDValue Add = N->getOperand(1);
+
+ if (DAG.MaskedValueIsZero(Add->getOperand(1),
+ APInt::getAllOnesValue(4 /* 16 byte alignment */).zext(
+ Add.getValueType().getScalarType().getSizeInBits()))) {
+ SDNode *BasePtr = Add->getOperand(0).getNode();
+ for (SDNode::use_iterator UI = BasePtr->use_begin(),
+ UE = BasePtr->use_end(); UI != UE; ++UI) {
+ if (UI->getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
+ cast<ConstantSDNode>(UI->getOperand(0))->getZExtValue() ==
+ Intrinsic::ppc_altivec_lvsl) {
+ // We've found another LVSL, and this address if an aligned
+ // multiple of that one. The results will be the same, so use the
+ // one we've just found instead.
+
+ return SDValue(*UI, 0);
+ }
+ }
+ }
+ }
case ISD::BSWAP:
// Turn BSWAP (LOAD) -> lhbrx/lwbrx.
if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) &&