1 //===- RISCVMatInt.cpp - Immediate materialisation -------------*- C++ -*--===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 9 #include "RISCVMatInt.h" 10 #include "MCTargetDesc/RISCVMCTargetDesc.h" 11 #include "llvm/ADT/APInt.h" 12 #include "llvm/Support/MathExtras.h" 13 using namespace llvm; 14 15 // Recursively generate a sequence for materializing an integer. 16 static void generateInstSeqImpl(int64_t Val, 17 const FeatureBitset &ActiveFeatures, 18 RISCVMatInt::InstSeq &Res) { 19 bool IsRV64 = ActiveFeatures[RISCV::Feature64Bit]; 20 21 if (isInt<32>(Val)) { 22 // Depending on the active bits in the immediate Value v, the following 23 // instruction sequences are emitted: 24 // 25 // v == 0 : ADDI 26 // v[0,12) != 0 && v[12,32) == 0 : ADDI 27 // v[0,12) == 0 && v[12,32) != 0 : LUI 28 // v[0,32) != 0 : LUI+ADDI(W) 29 int64_t Hi20 = ((Val + 0x800) >> 12) & 0xFFFFF; 30 int64_t Lo12 = SignExtend64<12>(Val); 31 32 if (Hi20) 33 Res.push_back(RISCVMatInt::Inst(RISCV::LUI, Hi20)); 34 35 if (Lo12 || Hi20 == 0) { 36 unsigned AddiOpc = (IsRV64 && Hi20) ? RISCV::ADDIW : RISCV::ADDI; 37 Res.push_back(RISCVMatInt::Inst(AddiOpc, Lo12)); 38 } 39 return; 40 } 41 42 assert(IsRV64 && "Can't emit >32-bit imm for non-RV64 target"); 43 44 // In the worst case, for a full 64-bit constant, a sequence of 8 instructions 45 // (i.e., LUI+ADDIW+SLLI+ADDI+SLLI+ADDI+SLLI+ADDI) has to be emmitted. Note 46 // that the first two instructions (LUI+ADDIW) can contribute up to 32 bits 47 // while the following ADDI instructions contribute up to 12 bits each. 48 // 49 // On the first glance, implementing this seems to be possible by simply 50 // emitting the most significant 32 bits (LUI+ADDIW) followed by as many left 51 // shift (SLLI) and immediate additions (ADDI) as needed. However, due to the 52 // fact that ADDI performs a sign extended addition, doing it like that would 53 // only be possible when at most 11 bits of the ADDI instructions are used. 54 // Using all 12 bits of the ADDI instructions, like done by GAS, actually 55 // requires that the constant is processed starting with the least significant 56 // bit. 57 // 58 // In the following, constants are processed from LSB to MSB but instruction 59 // emission is performed from MSB to LSB by recursively calling 60 // generateInstSeq. In each recursion, first the lowest 12 bits are removed 61 // from the constant and the optimal shift amount, which can be greater than 62 // 12 bits if the constant is sparse, is determined. Then, the shifted 63 // remaining constant is processed recursively and gets emitted as soon as it 64 // fits into 32 bits. The emission of the shifts and additions is subsequently 65 // performed when the recursion returns. 66 67 int64_t Lo12 = SignExtend64<12>(Val); 68 int64_t Hi52 = ((uint64_t)Val + 0x800ull) >> 12; 69 int ShiftAmount = 12 + findFirstSet((uint64_t)Hi52); 70 Hi52 = SignExtend64(Hi52 >> (ShiftAmount - 12), 64 - ShiftAmount); 71 72 generateInstSeqImpl(Hi52, ActiveFeatures, Res); 73 74 Res.push_back(RISCVMatInt::Inst(RISCV::SLLI, ShiftAmount)); 75 if (Lo12) 76 Res.push_back(RISCVMatInt::Inst(RISCV::ADDI, Lo12)); 77 } 78 79 namespace llvm { 80 namespace RISCVMatInt { 81 InstSeq generateInstSeq(int64_t Val, const FeatureBitset &ActiveFeatures) { 82 RISCVMatInt::InstSeq Res; 83 generateInstSeqImpl(Val, ActiveFeatures, Res); 84 85 // If the constant is positive we might be able to generate a shifted constant 86 // with no leading zeros and use a final SRLI to restore them. 87 if (Val > 0 && Res.size() > 2) { 88 assert(ActiveFeatures[RISCV::Feature64Bit] && 89 "Expected RV32 to only need 2 instructions"); 90 unsigned LeadingZeros = countLeadingZeros((uint64_t)Val); 91 uint64_t ShiftedVal = (uint64_t)Val << LeadingZeros; 92 // Fill in the bits that will be shifted out with 1s. An example where this 93 // helps is trailing one masks with 32 or more ones. This will generate 94 // ADDI -1 and an SRLI. 95 ShiftedVal |= maskTrailingOnes<uint64_t>(LeadingZeros); 96 97 RISCVMatInt::InstSeq TmpSeq; 98 generateInstSeqImpl(ShiftedVal, ActiveFeatures, TmpSeq); 99 TmpSeq.push_back(RISCVMatInt::Inst(RISCV::SRLI, LeadingZeros)); 100 101 // Keep the new sequence if it is an improvement. 102 if (TmpSeq.size() < Res.size()) { 103 Res = TmpSeq; 104 // A 2 instruction sequence is the best we can do. 105 if (Res.size() <= 2) 106 return Res; 107 } 108 109 // Some cases can benefit from filling the lower bits with zeros instead. 110 ShiftedVal &= maskTrailingZeros<uint64_t>(LeadingZeros); 111 TmpSeq.clear(); 112 generateInstSeqImpl(ShiftedVal, ActiveFeatures, TmpSeq); 113 TmpSeq.push_back(RISCVMatInt::Inst(RISCV::SRLI, LeadingZeros)); 114 115 // Keep the new sequence if it is an improvement. 116 if (TmpSeq.size() < Res.size()) { 117 Res = TmpSeq; 118 // A 2 instruction sequence is the best we can do. 119 if (Res.size() <= 2) 120 return Res; 121 } 122 123 // If we have exactly 32 leading zeros and Zba, we can try using zext.w at 124 // the end of the sequence. 125 if (LeadingZeros == 32 && ActiveFeatures[RISCV::FeatureExtZba]) { 126 // Try replacing upper bits with 1. 127 uint64_t LeadingOnesVal = Val | maskLeadingOnes<uint64_t>(LeadingZeros); 128 TmpSeq.clear(); 129 generateInstSeqImpl(LeadingOnesVal, ActiveFeatures, TmpSeq); 130 TmpSeq.push_back(RISCVMatInt::Inst(RISCV::ADDUW, 0)); 131 132 // Keep the new sequence if it is an improvement. 133 if (TmpSeq.size() < Res.size()) { 134 Res = TmpSeq; 135 // A 2 instruction sequence is the best we can do. 136 if (Res.size() <= 2) 137 return Res; 138 } 139 } 140 } 141 142 return Res; 143 } 144 145 int getIntMatCost(const APInt &Val, unsigned Size, 146 const FeatureBitset &ActiveFeatures) { 147 bool IsRV64 = ActiveFeatures[RISCV::Feature64Bit]; 148 int PlatRegSize = IsRV64 ? 64 : 32; 149 150 // Split the constant into platform register sized chunks, and calculate cost 151 // of each chunk. 152 int Cost = 0; 153 for (unsigned ShiftVal = 0; ShiftVal < Size; ShiftVal += PlatRegSize) { 154 APInt Chunk = Val.ashr(ShiftVal).sextOrTrunc(PlatRegSize); 155 InstSeq MatSeq = generateInstSeq(Chunk.getSExtValue(), ActiveFeatures); 156 Cost += MatSeq.size(); 157 } 158 return std::max(1, Cost); 159 } 160 } // namespace RISCVMatInt 161 } // namespace llvm 162