| 1 | //===-- PPCTOCRegDeps.cpp - Add Extra TOC Register Dependencies -----------===// |
| 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 | // When resolving an address using the ELF ABI TOC pointer, two relocations are |
| 10 | // generally required: one for the high part and one for the low part. Only |
| 11 | // the high part generally explicitly depends on r2 (the TOC pointer). And, so, |
| 12 | // we might produce code like this: |
| 13 | // |
| 14 | // .Ltmp526: |
| 15 | // addis 3, 2, .LC12@toc@ha |
| 16 | // .Ltmp1628: |
| 17 | // std 2, 40(1) |
| 18 | // ld 5, 0(27) |
| 19 | // ld 2, 8(27) |
| 20 | // ld 11, 16(27) |
| 21 | // ld 3, .LC12@toc@l(3) |
| 22 | // rldicl 4, 4, 0, 32 |
| 23 | // mtctr 5 |
| 24 | // bctrl |
| 25 | // ld 2, 40(1) |
| 26 | // |
| 27 | // And there is nothing wrong with this code, as such, but there is a linker bug |
| 28 | // in binutils (https://sourceware.org/bugzilla/show_bug.cgi?id=18414) that will |
| 29 | // misoptimize this code sequence to this: |
| 30 | // nop |
| 31 | // std r2,40(r1) |
| 32 | // ld r5,0(r27) |
| 33 | // ld r2,8(r27) |
| 34 | // ld r11,16(r27) |
| 35 | // ld r3,-32472(r2) |
| 36 | // clrldi r4,r4,32 |
| 37 | // mtctr r5 |
| 38 | // bctrl |
| 39 | // ld r2,40(r1) |
| 40 | // because the linker does not know (and does not check) that the value in r2 |
| 41 | // changed in between the instruction using the .LC12@toc@ha (TOC-relative) |
| 42 | // relocation and the instruction using the .LC12@toc@l(3) relocation. |
| 43 | // Because it finds these instructions using the relocations (and not by |
| 44 | // scanning the instructions), it has been asserted that there is no good way |
| 45 | // to detect the change of r2 in between. As a result, this bug may never be |
| 46 | // fixed (i.e. it may become part of the definition of the ABI). GCC was |
| 47 | // updated to add extra dependencies on r2 to instructions using the @toc@l |
| 48 | // relocations to avoid this problem, and we'll do the same here. |
| 49 | // |
| 50 | // This is done as a separate pass because: |
| 51 | // 1. These extra r2 dependencies are not really properties of the |
| 52 | // instructions, but rather due to a linker bug, and maybe one day we'll be |
| 53 | // able to get rid of them when targeting linkers without this bug (and, |
| 54 | // thus, keeping the logic centralized here will make that |
| 55 | // straightforward). |
| 56 | // 2. There are ISel-level peephole optimizations that propagate the @toc@l |
| 57 | // relocations to some user instructions, and so the exta dependencies do |
| 58 | // not apply only to a fixed set of instructions (without undesirable |
| 59 | // definition replication). |
| 60 | // |
| 61 | //===----------------------------------------------------------------------===// |
| 62 | |
| 63 | #include "PPC.h" |
| 64 | #include "PPCInstrInfo.h" |
| 65 | #include "PPCTargetMachine.h" |
| 66 | #include "llvm/ADT/STLExtras.h" |
| 67 | #include "llvm/ADT/Statistic.h" |
| 68 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 69 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 70 | #include "llvm/CodeGen/MachineInstr.h" |
| 71 | #include "llvm/Support/ErrorHandling.h" |
| 72 | |
| 73 | using namespace llvm; |
| 74 | |
| 75 | #define DEBUG_TYPE "ppc-toc-reg-deps" |
| 76 | |
| 77 | namespace { |
| 78 | // PPCTOCRegDeps pass - For simple functions without epilogue code, move |
| 79 | // returns up, and create conditional returns, to avoid unnecessary |
| 80 | // branch-to-blr sequences. |
| 81 | struct PPCTOCRegDeps : public MachineFunctionPass { |
| 82 | static char ID; |
| 83 | PPCTOCRegDeps() : MachineFunctionPass(ID) {} |
| 84 | |
| 85 | protected: |
| 86 | bool hasTOCLoReloc(const MachineInstr &MI) { |
| 87 | if (MI.getOpcode() == PPC::LDtocL || MI.getOpcode() == PPC::ADDItocL8 || |
| 88 | MI.getOpcode() == PPC::LWZtocL) |
| 89 | return true; |
| 90 | |
| 91 | for (const MachineOperand &MO : MI.operands()) { |
| 92 | if (MO.getTargetFlags() == PPCII::MO_TOC_LO) |
| 93 | return true; |
| 94 | } |
| 95 | |
| 96 | return false; |
| 97 | } |
| 98 | |
| 99 | bool processBlock(MachineBasicBlock &MBB) { |
| 100 | bool Changed = false; |
| 101 | |
| 102 | const bool isPPC64 = |
| 103 | MBB.getParent()->getSubtarget<PPCSubtarget>().isPPC64(); |
| 104 | const unsigned TOCReg = isPPC64 ? PPC::X2 : PPC::R2; |
| 105 | |
| 106 | for (auto &MI : MBB) { |
| 107 | if (!hasTOCLoReloc(MI)) |
| 108 | continue; |
| 109 | |
| 110 | MI.addOperand(Op: MachineOperand::CreateReg(Reg: TOCReg, |
| 111 | isDef: false /*IsDef*/, |
| 112 | isImp: true /*IsImp*/)); |
| 113 | Changed = true; |
| 114 | } |
| 115 | |
| 116 | return Changed; |
| 117 | } |
| 118 | |
| 119 | public: |
| 120 | bool runOnMachineFunction(MachineFunction &MF) override { |
| 121 | bool Changed = false; |
| 122 | |
| 123 | for (MachineBasicBlock &B : llvm::make_early_inc_range(Range&: MF)) |
| 124 | if (processBlock(MBB&: B)) |
| 125 | Changed = true; |
| 126 | |
| 127 | return Changed; |
| 128 | } |
| 129 | |
| 130 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 131 | MachineFunctionPass::getAnalysisUsage(AU); |
| 132 | } |
| 133 | }; |
| 134 | } |
| 135 | |
| 136 | INITIALIZE_PASS(PPCTOCRegDeps, DEBUG_TYPE, |
| 137 | "PowerPC TOC Register Dependencies" , false, false) |
| 138 | |
| 139 | char PPCTOCRegDeps::ID = 0; |
| 140 | FunctionPass* |
| 141 | llvm::createPPCTOCRegDepsPass() { return new PPCTOCRegDeps(); } |
| 142 | |
| 143 | |