From b6516aeef12a05aa47515f76e18fc426d85babbd Mon Sep 17 00:00:00 2001
From: Benjamin Kramer <benny.kra@googlemail.com>
Date: Sat, 15 Jan 2011 20:30:30 +0000
Subject: [PATCH] Reimplement CTPOP legalization with the "best" algorithm from
http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
In a silly microbenchmark on a 65 nm core2 this is 1.5x faster than the old
code in 32 bit mode and about 2x faster in 64 bit mode. It's also a lot shorter,
especially when counting 64 bit population on a 32 bit target.
I hope this is fast enough to replace Kernighan-style counting loops even when
the input is rather sparse.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@123547 91177308-0d34-0410-b5e6-96231b3b80d8
---
lib/CodeGen/SelectionDAG/LegalizeDAG.cpp | 63 +++++++++++++++++-------
1 file changed, 45 insertions(+), 18 deletions(-)
diff --git a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index cbe38450f98..3a29424d070 100644
--- a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -2388,6 +2388,17 @@ SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, DebugLoc dl) {
}
}
+/// SplatByte - Distribute ByteVal over NumBits bits.
+static APInt SplatByte(unsigned NumBits, uint8_t ByteVal) {
+ APInt Val = APInt(NumBits, ByteVal);
+ unsigned Shift = 8;
+ for (unsigned i = NumBits; i > 8; i >>= 1) {
+ Val = (Val << Shift) | Val;
+ Shift <<= 1;
+ }
+ return Val;
+}
+
/// ExpandBitCount - Expand the specified bitcount instruction into operations.
///
SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
@@ -2395,26 +2406,42 @@ SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
switch (Opc) {
default: assert(0 && "Cannot expand this yet!");
case ISD::CTPOP: {
- static const uint64_t mask[6] = {
- 0x5555555555555555ULL, 0x3333333333333333ULL,
- 0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
- 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
- };
EVT VT = Op.getValueType();
EVT ShVT = TLI.getShiftAmountTy();
- unsigned len = VT.getSizeInBits();
- for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
- //x = (x & mask[i][len/8]) + (x >> (1 << i) & mask[i][len/8])
- unsigned EltSize = VT.isVector() ?
- VT.getVectorElementType().getSizeInBits() : len;
- SDValue Tmp2 = DAG.getConstant(APInt(EltSize, mask[i]), VT);
- SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
- Op = DAG.getNode(ISD::ADD, dl, VT,
- DAG.getNode(ISD::AND, dl, VT, Op, Tmp2),
- DAG.getNode(ISD::AND, dl, VT,
- DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3),
- Tmp2));
- }
+ unsigned Len = VT.getSizeInBits();
+
+ // This is the "best" algorithm from
+ // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
+
+ SDValue Mask55 = DAG.getConstant(SplatByte(Len, 0x55), VT);
+ SDValue Mask33 = DAG.getConstant(SplatByte(Len, 0x33), VT);
+ SDValue Mask0F = DAG.getConstant(SplatByte(Len, 0x0F), VT);
+ SDValue Mask01 = DAG.getConstant(SplatByte(Len, 0x01), VT);
+
+ // v = v - ((v >> 1) & 0x55555555...)
+ Op = DAG.getNode(ISD::SUB, dl, VT, Op,
+ DAG.getNode(ISD::AND, dl, VT,
+ DAG.getNode(ISD::SRL, dl, VT, Op,
+ DAG.getConstant(1, ShVT)),
+ Mask55));
+ // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
+ Op = DAG.getNode(ISD::ADD, dl, VT,
+ DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
+ DAG.getNode(ISD::AND, dl, VT,
+ DAG.getNode(ISD::SRL, dl, VT, Op,
+ DAG.getConstant(2, ShVT)),
+ Mask33));
+ // v = (v + (v >> 4)) & 0x0F0F0F0F...
+ Op = DAG.getNode(ISD::AND, dl, VT,
+ DAG.getNode(ISD::ADD, dl, VT, Op,
+ DAG.getNode(ISD::SRL, dl, VT, Op,
+ DAG.getConstant(4, ShVT))),
+ Mask0F);
+ // v = (v * 0x01010101...) >> (Len - 8)
+ Op = DAG.getNode(ISD::SRL, dl, VT,
+ DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
+ DAG.getConstant(Len - 8, ShVT));
+
return Op;
}
case ISD::CTLZ: {