| 1 | //===-- MVETailPredUtils.h - Tail predication utility functions -*- 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 | // This file contains utility functions for low overhead and tail predicated |
| 10 | // loops, shared between the ARMLowOverheadLoops pass and anywhere else that |
| 11 | // needs them. |
| 12 | // |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | |
| 15 | #ifndef LLVM_LIB_TARGET_ARM_MVETAILPREDUTILS_H |
| 16 | #define LLVM_LIB_TARGET_ARM_MVETAILPREDUTILS_H |
| 17 | |
| 18 | #include "llvm/CodeGen/MachineInstr.h" |
| 19 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 20 | #include "llvm/CodeGen/MachineOperand.h" |
| 21 | #include "llvm/CodeGen/TargetInstrInfo.h" |
| 22 | |
| 23 | namespace llvm { |
| 24 | |
| 25 | static inline unsigned VCTPOpcodeToLSTP(unsigned Opcode, bool IsDoLoop) { |
| 26 | switch (Opcode) { |
| 27 | default: |
| 28 | llvm_unreachable("unhandled vctp opcode" ); |
| 29 | break; |
| 30 | case ARM::MVE_VCTP8: |
| 31 | return IsDoLoop ? ARM::MVE_DLSTP_8 : ARM::MVE_WLSTP_8; |
| 32 | case ARM::MVE_VCTP16: |
| 33 | return IsDoLoop ? ARM::MVE_DLSTP_16 : ARM::MVE_WLSTP_16; |
| 34 | case ARM::MVE_VCTP32: |
| 35 | return IsDoLoop ? ARM::MVE_DLSTP_32 : ARM::MVE_WLSTP_32; |
| 36 | case ARM::MVE_VCTP64: |
| 37 | return IsDoLoop ? ARM::MVE_DLSTP_64 : ARM::MVE_WLSTP_64; |
| 38 | } |
| 39 | return 0; |
| 40 | } |
| 41 | |
| 42 | static inline unsigned getTailPredVectorWidth(unsigned Opcode) { |
| 43 | switch (Opcode) { |
| 44 | default: |
| 45 | llvm_unreachable("unhandled vctp opcode" ); |
| 46 | case ARM::MVE_VCTP8: |
| 47 | return 16; |
| 48 | case ARM::MVE_VCTP16: |
| 49 | return 8; |
| 50 | case ARM::MVE_VCTP32: |
| 51 | return 4; |
| 52 | case ARM::MVE_VCTP64: |
| 53 | return 2; |
| 54 | } |
| 55 | return 0; |
| 56 | } |
| 57 | |
| 58 | static inline bool isVCTP(const MachineInstr *MI) { |
| 59 | switch (MI->getOpcode()) { |
| 60 | default: |
| 61 | break; |
| 62 | case ARM::MVE_VCTP8: |
| 63 | case ARM::MVE_VCTP16: |
| 64 | case ARM::MVE_VCTP32: |
| 65 | case ARM::MVE_VCTP64: |
| 66 | return true; |
| 67 | } |
| 68 | return false; |
| 69 | } |
| 70 | |
| 71 | static inline bool isDoLoopStart(const MachineInstr &MI) { |
| 72 | return MI.getOpcode() == ARM::t2DoLoopStart || |
| 73 | MI.getOpcode() == ARM::t2DoLoopStartTP; |
| 74 | } |
| 75 | |
| 76 | static inline bool isWhileLoopStart(const MachineInstr &MI) { |
| 77 | return MI.getOpcode() == ARM::t2WhileLoopStart || |
| 78 | MI.getOpcode() == ARM::t2WhileLoopStartLR || |
| 79 | MI.getOpcode() == ARM::t2WhileLoopStartTP; |
| 80 | } |
| 81 | |
| 82 | static inline bool isLoopStart(const MachineInstr &MI) { |
| 83 | return isDoLoopStart(MI) || isWhileLoopStart(MI); |
| 84 | } |
| 85 | |
| 86 | // Return the TargetBB stored in a t2WhileLoopStartLR/t2WhileLoopStartTP. |
| 87 | inline MachineBasicBlock *getWhileLoopStartTargetBB(const MachineInstr &MI) { |
| 88 | assert(isWhileLoopStart(MI) && "Expected WhileLoopStart!" ); |
| 89 | unsigned Op = MI.getOpcode() == ARM::t2WhileLoopStartTP ? 3 : 2; |
| 90 | return MI.getOperand(i: Op).getMBB(); |
| 91 | } |
| 92 | |
| 93 | // WhileLoopStart holds the exit block, so produce a subs Op0, Op1, 0 and then a |
| 94 | // beq that branches to the exit branch. |
| 95 | // If UseCmp is true, this will create a t2CMP instead of a t2SUBri, meaning the |
| 96 | // value of LR into the loop will not be setup. This is used if the LR setup is |
| 97 | // done via another means (via a t2DoLoopStart, for example). |
| 98 | inline void RevertWhileLoopStartLR(MachineInstr *MI, const TargetInstrInfo *TII, |
| 99 | unsigned BrOpc = ARM::t2Bcc, |
| 100 | bool UseCmp = false) { |
| 101 | MachineBasicBlock *MBB = MI->getParent(); |
| 102 | assert((MI->getOpcode() == ARM::t2WhileLoopStartLR || |
| 103 | MI->getOpcode() == ARM::t2WhileLoopStartTP) && |
| 104 | "Only expected a t2WhileLoopStartLR/TP in RevertWhileLoopStartLR!" ); |
| 105 | |
| 106 | // Subs/Cmp |
| 107 | if (UseCmp) { |
| 108 | MachineInstrBuilder MIB = |
| 109 | BuildMI(BB&: *MBB, I: MI, MIMD: MI->getDebugLoc(), MCID: TII->get(Opcode: ARM::t2CMPri)); |
| 110 | MIB.add(MO: MI->getOperand(i: 1)); |
| 111 | MIB.addImm(Val: 0); |
| 112 | MIB.addImm(Val: ARMCC::AL); |
| 113 | MIB.addReg(RegNo: ARM::NoRegister); |
| 114 | } else { |
| 115 | MachineInstrBuilder MIB = |
| 116 | BuildMI(BB&: *MBB, I: MI, MIMD: MI->getDebugLoc(), MCID: TII->get(Opcode: ARM::t2SUBri)); |
| 117 | MIB.add(MO: MI->getOperand(i: 0)); |
| 118 | MIB.add(MO: MI->getOperand(i: 1)); |
| 119 | MIB.addImm(Val: 0); |
| 120 | MIB.addImm(Val: ARMCC::AL); |
| 121 | MIB.addReg(RegNo: ARM::NoRegister); |
| 122 | MIB.addReg(RegNo: ARM::CPSR, flags: RegState::Define); |
| 123 | } |
| 124 | |
| 125 | // Branch |
| 126 | MachineInstrBuilder MIB = |
| 127 | BuildMI(BB&: *MBB, I: MI, MIMD: MI->getDebugLoc(), MCID: TII->get(Opcode: BrOpc)); |
| 128 | MIB.addMBB(MBB: getWhileLoopStartTargetBB(MI: *MI)); // branch target |
| 129 | MIB.addImm(Val: ARMCC::EQ); // condition code |
| 130 | MIB.addReg(RegNo: ARM::CPSR); |
| 131 | |
| 132 | MI->eraseFromParent(); |
| 133 | } |
| 134 | |
| 135 | inline void RevertDoLoopStart(MachineInstr *MI, const TargetInstrInfo *TII) { |
| 136 | MachineBasicBlock *MBB = MI->getParent(); |
| 137 | BuildMI(BB&: *MBB, I: MI, MIMD: MI->getDebugLoc(), MCID: TII->get(Opcode: ARM::tMOVr)) |
| 138 | .add(MO: MI->getOperand(i: 0)) |
| 139 | .add(MO: MI->getOperand(i: 1)) |
| 140 | .add(MOs: predOps(Pred: ARMCC::AL)); |
| 141 | |
| 142 | MI->eraseFromParent(); |
| 143 | } |
| 144 | |
| 145 | inline void RevertLoopDec(MachineInstr *MI, const TargetInstrInfo *TII, |
| 146 | bool SetFlags = false) { |
| 147 | MachineBasicBlock *MBB = MI->getParent(); |
| 148 | |
| 149 | MachineInstrBuilder MIB = |
| 150 | BuildMI(BB&: *MBB, I: MI, MIMD: MI->getDebugLoc(), MCID: TII->get(Opcode: ARM::t2SUBri)); |
| 151 | MIB.add(MO: MI->getOperand(i: 0)); |
| 152 | MIB.add(MO: MI->getOperand(i: 1)); |
| 153 | MIB.add(MO: MI->getOperand(i: 2)); |
| 154 | MIB.addImm(Val: ARMCC::AL); |
| 155 | MIB.addReg(RegNo: 0); |
| 156 | |
| 157 | if (SetFlags) { |
| 158 | MIB.addReg(RegNo: ARM::CPSR); |
| 159 | MIB->getOperand(i: 5).setIsDef(true); |
| 160 | } else |
| 161 | MIB.addReg(RegNo: 0); |
| 162 | |
| 163 | MI->eraseFromParent(); |
| 164 | } |
| 165 | |
| 166 | // Generate a subs, or sub and cmp, and a branch instead of an LE. |
| 167 | inline void RevertLoopEnd(MachineInstr *MI, const TargetInstrInfo *TII, |
| 168 | unsigned BrOpc = ARM::t2Bcc, bool SkipCmp = false) { |
| 169 | MachineBasicBlock *MBB = MI->getParent(); |
| 170 | |
| 171 | // Create cmp |
| 172 | if (!SkipCmp) { |
| 173 | MachineInstrBuilder MIB = |
| 174 | BuildMI(BB&: *MBB, I: MI, MIMD: MI->getDebugLoc(), MCID: TII->get(Opcode: ARM::t2CMPri)); |
| 175 | MIB.add(MO: MI->getOperand(i: 0)); |
| 176 | MIB.addImm(Val: 0); |
| 177 | MIB.addImm(Val: ARMCC::AL); |
| 178 | MIB.addReg(RegNo: ARM::NoRegister); |
| 179 | } |
| 180 | |
| 181 | // Create bne |
| 182 | MachineInstrBuilder MIB = |
| 183 | BuildMI(BB&: *MBB, I: MI, MIMD: MI->getDebugLoc(), MCID: TII->get(Opcode: BrOpc)); |
| 184 | MIB.add(MO: MI->getOperand(i: 1)); // branch target |
| 185 | MIB.addImm(Val: ARMCC::NE); // condition code |
| 186 | MIB.addReg(RegNo: ARM::CPSR); |
| 187 | MI->eraseFromParent(); |
| 188 | } |
| 189 | |
| 190 | } // end namespace llvm |
| 191 | |
| 192 | #endif // LLVM_LIB_TARGET_ARM_MVETAILPREDUTILS_H |
| 193 | |