1 //===- CodeGenCommonISel.h - Common code between ISels ---------*- 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 declares common utilities that are shared between SelectionDAG and 10 // GlobalISel frameworks. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef LLVM_CODEGEN_CODEGENCOMMONISEL_H 15 #define LLVM_CODEGEN_CODEGENCOMMONISEL_H 16 17 #include "llvm/CodeGen/MachineBasicBlock.h" 18 #include <cassert> 19 namespace llvm { 20 21 class BasicBlock; 22 class MachineBasicBlock; 23 /// Encapsulates all of the information needed to generate a stack protector 24 /// check, and signals to isel when initialized that one needs to be generated. 25 /// 26 /// *NOTE* The following is a high level documentation of SelectionDAG Stack 27 /// Protector Generation. This is now also ported be shared with GlobalISel, 28 /// but without any significant changes. 29 /// 30 /// High Level Overview of ISel Stack Protector Generation: 31 /// 32 /// Previously, the "stack protector" IR pass handled stack protector 33 /// generation. This necessitated splitting basic blocks at the IR level to 34 /// create the success/failure basic blocks in the tail of the basic block in 35 /// question. As a result of this, calls that would have qualified for the 36 /// sibling call optimization were no longer eligible for optimization since 37 /// said calls were no longer right in the "tail position" (i.e. the immediate 38 /// predecessor of a ReturnInst instruction). 39 /// 40 /// Since the sibling call optimization causes the callee to reuse the caller's 41 /// stack, if we could delay the generation of the stack protector check until 42 /// later in CodeGen after the sibling call decision was made, we get both the 43 /// tail call optimization and the stack protector check! 44 /// 45 /// A few goals in solving this problem were: 46 /// 47 /// 1. Preserve the architecture independence of stack protector generation. 48 /// 49 /// 2. Preserve the normal IR level stack protector check for platforms like 50 /// OpenBSD for which we support platform-specific stack protector 51 /// generation. 52 /// 53 /// The main problem that guided the present solution is that one can not 54 /// solve this problem in an architecture independent manner at the IR level 55 /// only. This is because: 56 /// 57 /// 1. The decision on whether or not to perform a sibling call on certain 58 /// platforms (for instance i386) requires lower level information 59 /// related to available registers that can not be known at the IR level. 60 /// 61 /// 2. Even if the previous point were not true, the decision on whether to 62 /// perform a tail call is done in LowerCallTo in SelectionDAG (or 63 /// CallLowering in GlobalISel) which occurs after the Stack Protector 64 /// Pass. As a result, one would need to put the relevant callinst into the 65 /// stack protector check success basic block (where the return inst is 66 /// placed) and then move it back later at ISel/MI time before the 67 /// stack protector check if the tail call optimization failed. The MI 68 /// level option was nixed immediately since it would require 69 /// platform-specific pattern matching. The ISel level option was 70 /// nixed because SelectionDAG only processes one IR level basic block at a 71 /// time implying one could not create a DAG Combine to move the callinst. 72 /// 73 /// To get around this problem: 74 /// 75 /// 1. SelectionDAG can only process one block at a time, we can generate 76 /// multiple machine basic blocks for one IR level basic block. 77 /// This is how we handle bit tests and switches. 78 /// 79 /// 2. At the MI level, tail calls are represented via a special return 80 /// MIInst called "tcreturn". Thus if we know the basic block in which we 81 /// wish to insert the stack protector check, we get the correct behavior 82 /// by always inserting the stack protector check right before the return 83 /// statement. This is a "magical transformation" since no matter where 84 /// the stack protector check intrinsic is, we always insert the stack 85 /// protector check code at the end of the BB. 86 /// 87 /// Given the aforementioned constraints, the following solution was devised: 88 /// 89 /// 1. On platforms that do not support ISel stack protector check 90 /// generation, allow for the normal IR level stack protector check 91 /// generation to continue. 92 /// 93 /// 2. On platforms that do support ISel stack protector check 94 /// generation: 95 /// 96 /// a. Use the IR level stack protector pass to decide if a stack 97 /// protector is required/which BB we insert the stack protector check 98 /// in by reusing the logic already therein. 99 /// 100 /// b. After we finish selecting the basic block, we produce the validation 101 /// code with one of these techniques: 102 /// 1) with a call to a guard check function 103 /// 2) with inlined instrumentation 104 /// 105 /// 1) We insert a call to the check function before the terminator. 106 /// 107 /// 2) We first find a splice point in the parent basic block 108 /// before the terminator and then splice the terminator of said basic 109 /// block into the success basic block. Then we code-gen a new tail for 110 /// the parent basic block consisting of the two loads, the comparison, 111 /// and finally two branches to the success/failure basic blocks. We 112 /// conclude by code-gening the failure basic block if we have not 113 /// code-gened it already (all stack protector checks we generate in 114 /// the same function, use the same failure basic block). 115 class StackProtectorDescriptor { 116 public: 117 StackProtectorDescriptor() = default; 118 119 /// Returns true if all fields of the stack protector descriptor are 120 /// initialized implying that we should/are ready to emit a stack protector. 121 bool shouldEmitStackProtector() const { 122 return ParentMBB && SuccessMBB && FailureMBB; 123 } 124 125 bool shouldEmitFunctionBasedCheckStackProtector() const { 126 return ParentMBB && !SuccessMBB && !FailureMBB; 127 } 128 129 /// Initialize the stack protector descriptor structure for a new basic 130 /// block. 131 void initialize(const BasicBlock *BB, MachineBasicBlock *MBB, 132 bool FunctionBasedInstrumentation) { 133 // Make sure we are not initialized yet. 134 assert(!shouldEmitStackProtector() && "Stack Protector Descriptor is " 135 "already initialized!"); 136 ParentMBB = MBB; 137 if (!FunctionBasedInstrumentation) { 138 SuccessMBB = addSuccessorMBB(BB, MBB, /* IsLikely */ true); 139 FailureMBB = addSuccessorMBB(BB, MBB, /* IsLikely */ false, FailureMBB); 140 } 141 } 142 143 /// Reset state that changes when we handle different basic blocks. 144 /// 145 /// This currently includes: 146 /// 147 /// 1. The specific basic block we are generating a 148 /// stack protector for (ParentMBB). 149 /// 150 /// 2. The successor machine basic block that will contain the tail of 151 /// parent mbb after we create the stack protector check (SuccessMBB). This 152 /// BB is visited only on stack protector check success. 153 void resetPerBBState() { 154 ParentMBB = nullptr; 155 SuccessMBB = nullptr; 156 } 157 158 /// Reset state that only changes when we switch functions. 159 /// 160 /// This currently includes: 161 /// 162 /// 1. FailureMBB since we reuse the failure code path for all stack 163 /// protector checks created in an individual function. 164 /// 165 /// 2.The guard variable since the guard variable we are checking against is 166 /// always the same. 167 void resetPerFunctionState() { FailureMBB = nullptr; } 168 169 MachineBasicBlock *getParentMBB() { return ParentMBB; } 170 MachineBasicBlock *getSuccessMBB() { return SuccessMBB; } 171 MachineBasicBlock *getFailureMBB() { return FailureMBB; } 172 173 private: 174 /// The basic block for which we are generating the stack protector. 175 /// 176 /// As a result of stack protector generation, we will splice the 177 /// terminators of this basic block into the successor mbb SuccessMBB and 178 /// replace it with a compare/branch to the successor mbbs 179 /// SuccessMBB/FailureMBB depending on whether or not the stack protector 180 /// was violated. 181 MachineBasicBlock *ParentMBB = nullptr; 182 183 /// A basic block visited on stack protector check success that contains the 184 /// terminators of ParentMBB. 185 MachineBasicBlock *SuccessMBB = nullptr; 186 187 /// This basic block visited on stack protector check failure that will 188 /// contain a call to __stack_chk_fail(). 189 MachineBasicBlock *FailureMBB = nullptr; 190 191 /// Add a successor machine basic block to ParentMBB. If the successor mbb 192 /// has not been created yet (i.e. if SuccMBB = 0), then the machine basic 193 /// block will be created. Assign a large weight if IsLikely is true. 194 MachineBasicBlock *addSuccessorMBB(const BasicBlock *BB, 195 MachineBasicBlock *ParentMBB, 196 bool IsLikely, 197 MachineBasicBlock *SuccMBB = nullptr); 198 }; 199 200 /// Find the split point at which to splice the end of BB into its success stack 201 /// protector check machine basic block. 202 /// 203 /// On many platforms, due to ABI constraints, terminators, even before register 204 /// allocation, use physical registers. This creates an issue for us since 205 /// physical registers at this point can not travel across basic 206 /// blocks. Luckily, selectiondag always moves physical registers into vregs 207 /// when they enter functions and moves them through a sequence of copies back 208 /// into the physical registers right before the terminator creating a 209 /// ``Terminator Sequence''. This function is searching for the beginning of the 210 /// terminator sequence so that we can ensure that we splice off not just the 211 /// terminator, but additionally the copies that move the vregs into the 212 /// physical registers. 213 MachineBasicBlock::iterator 214 findSplitPointForStackProtector(MachineBasicBlock *BB, 215 const TargetInstrInfo &TII); 216 217 } // namespace llvm 218 219 #endif // LLVM_CODEGEN_CODEGENCOMMONISEL_H 220