1 //===-- AArch64A53Fix835769.cpp -------------------------------------------===//
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
7 //===----------------------------------------------------------------------===//
8 // This pass changes code to work around Cortex-A53 erratum 835769.
9 // It works around it by inserting a nop instruction in code sequences that
10 // in some circumstances may trigger the erratum.
11 // It inserts a nop instruction between a sequence of the following 2 classes
13 // instr 1: mem-instr (including loads, stores and prefetches).
14 // instr 2: non-SIMD integer multiply-accumulate writing 64-bit X registers.
15 //===----------------------------------------------------------------------===//
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineFunctionPass.h"
21 #include "llvm/CodeGen/MachineInstr.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/TargetInstrInfo.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/raw_ostream.h"
30 #define DEBUG_TYPE "aarch64-fix-cortex-a53-835769"
32 STATISTIC(NumNopsAdded
, "Number of Nops added to work around erratum 835769");
34 //===----------------------------------------------------------------------===//
37 // Is the instruction a match for the instruction that comes first in the
38 // sequence of instructions that can trigger the erratum?
39 static bool isFirstInstructionInSequence(MachineInstr
*MI
) {
40 // Must return true if this instruction is a load, a store or a prefetch.
41 switch (MI
->getOpcode()) {
43 case AArch64::PRFMroW
:
44 case AArch64::PRFMroX
:
49 return MI
->mayLoadOrStore();
53 // Is the instruction a match for the instruction that comes second in the
54 // sequence that can trigger the erratum?
55 static bool isSecondInstructionInSequence(MachineInstr
*MI
) {
56 // Must return true for non-SIMD integer multiply-accumulates, writing
57 // to a 64-bit register.
58 switch (MI
->getOpcode()) {
59 // Erratum cannot be triggered when the destination register is 32 bits,
60 // therefore only include the following.
61 case AArch64::MSUBXrrr
:
62 case AArch64::MADDXrrr
:
63 case AArch64::SMADDLrrr
:
64 case AArch64::SMSUBLrrr
:
65 case AArch64::UMADDLrrr
:
66 case AArch64::UMSUBLrrr
:
67 // Erratum can only be triggered by multiply-adds, not by regular
68 // non-accumulating multiplies, i.e. when Ra=XZR='11111'
69 return MI
->getOperand(3).getReg() != AArch64::XZR
;
76 //===----------------------------------------------------------------------===//
79 class AArch64A53Fix835769
: public MachineFunctionPass
{
80 const TargetInstrInfo
*TII
;
84 explicit AArch64A53Fix835769() : MachineFunctionPass(ID
) {
85 initializeAArch64A53Fix835769Pass(*PassRegistry::getPassRegistry());
88 bool runOnMachineFunction(MachineFunction
&F
) override
;
90 MachineFunctionProperties
getRequiredProperties() const override
{
91 return MachineFunctionProperties().set(
92 MachineFunctionProperties::Property::NoVRegs
);
95 StringRef
getPassName() const override
{
96 return "Workaround A53 erratum 835769 pass";
99 void getAnalysisUsage(AnalysisUsage
&AU
) const override
{
100 AU
.setPreservesCFG();
101 MachineFunctionPass::getAnalysisUsage(AU
);
105 bool runOnBasicBlock(MachineBasicBlock
&MBB
);
107 char AArch64A53Fix835769::ID
= 0;
109 } // end anonymous namespace
111 INITIALIZE_PASS(AArch64A53Fix835769
, "aarch64-fix-cortex-a53-835769-pass",
112 "AArch64 fix for A53 erratum 835769", false, false)
114 //===----------------------------------------------------------------------===//
117 AArch64A53Fix835769::runOnMachineFunction(MachineFunction
&F
) {
118 LLVM_DEBUG(dbgs() << "***** AArch64A53Fix835769 *****\n");
119 bool Changed
= false;
120 TII
= F
.getSubtarget().getInstrInfo();
122 for (auto &MBB
: F
) {
123 Changed
|= runOnBasicBlock(MBB
);
128 // Return the block that was fallen through to get to MBB, if any,
129 // otherwise nullptr.
130 static MachineBasicBlock
*getBBFallenThrough(MachineBasicBlock
*MBB
,
131 const TargetInstrInfo
*TII
) {
132 // Get the previous machine basic block in the function.
133 MachineFunction::iterator
MBBI(MBB
);
135 // Can't go off top of function.
136 if (MBBI
== MBB
->getParent()->begin())
139 MachineBasicBlock
*TBB
= nullptr, *FBB
= nullptr;
140 SmallVector
<MachineOperand
, 2> Cond
;
142 MachineBasicBlock
*PrevBB
= &*std::prev(MBBI
);
143 for (MachineBasicBlock
*S
: MBB
->predecessors())
144 if (S
== PrevBB
&& !TII
->analyzeBranch(*PrevBB
, TBB
, FBB
, Cond
) && !TBB
&&
151 // Iterate through fallen through blocks trying to find a previous non-pseudo if
152 // there is one, otherwise return nullptr. Only look for instructions in
153 // previous blocks, not the current block, since we only use this to look at
155 static MachineInstr
*getLastNonPseudo(MachineBasicBlock
&MBB
,
156 const TargetInstrInfo
*TII
) {
157 MachineBasicBlock
*FMBB
= &MBB
;
159 // If there is no non-pseudo in the current block, loop back around and try
160 // the previous block (if there is one).
161 while ((FMBB
= getBBFallenThrough(FMBB
, TII
))) {
162 for (MachineInstr
&I
: make_range(FMBB
->rbegin(), FMBB
->rend()))
167 // There was no previous non-pseudo in the fallen through blocks
171 static void insertNopBeforeInstruction(MachineBasicBlock
&MBB
, MachineInstr
* MI
,
172 const TargetInstrInfo
*TII
) {
173 // If we are the first instruction of the block, put the NOP at the end of
174 // the previous fallthrough block
175 if (MI
== &MBB
.front()) {
176 MachineInstr
*I
= getLastNonPseudo(MBB
, TII
);
177 assert(I
&& "Expected instruction");
178 DebugLoc DL
= I
->getDebugLoc();
179 BuildMI(I
->getParent(), DL
, TII
->get(AArch64::HINT
)).addImm(0);
182 DebugLoc DL
= MI
->getDebugLoc();
183 BuildMI(MBB
, MI
, DL
, TII
->get(AArch64::HINT
)).addImm(0);
190 AArch64A53Fix835769::runOnBasicBlock(MachineBasicBlock
&MBB
) {
191 bool Changed
= false;
192 LLVM_DEBUG(dbgs() << "Running on MBB: " << MBB
193 << " - scanning instructions...\n");
195 // First, scan the basic block, looking for a sequence of 2 instructions
196 // that match the conditions under which the erratum may trigger.
198 // List of terminating instructions in matching sequences
199 std::vector
<MachineInstr
*> Sequences
;
201 MachineInstr
*PrevInstr
= nullptr;
203 // Try and find the last non-pseudo instruction in any fallen through blocks,
204 // if there isn't one, then we use nullptr to represent that.
205 PrevInstr
= getLastNonPseudo(MBB
, TII
);
207 for (auto &MI
: MBB
) {
208 MachineInstr
*CurrInstr
= &MI
;
209 LLVM_DEBUG(dbgs() << " Examining: " << MI
);
211 LLVM_DEBUG(dbgs() << " PrevInstr: " << *PrevInstr
212 << " CurrInstr: " << *CurrInstr
213 << " isFirstInstructionInSequence(PrevInstr): "
214 << isFirstInstructionInSequence(PrevInstr
) << "\n"
215 << " isSecondInstructionInSequence(CurrInstr): "
216 << isSecondInstructionInSequence(CurrInstr
) << "\n");
217 if (isFirstInstructionInSequence(PrevInstr
) &&
218 isSecondInstructionInSequence(CurrInstr
)) {
219 LLVM_DEBUG(dbgs() << " ** pattern found at Idx " << Idx
<< "!\n");
220 Sequences
.push_back(CurrInstr
);
223 if (!CurrInstr
->isPseudo())
224 PrevInstr
= CurrInstr
;
228 LLVM_DEBUG(dbgs() << "Scan complete, " << Sequences
.size()
229 << " occurrences of pattern found.\n");
231 // Then update the basic block, inserting nops between the detected sequences.
232 for (auto &MI
: Sequences
) {
234 insertNopBeforeInstruction(MBB
, MI
, TII
);
240 // Factory function used by AArch64TargetMachine to add the pass to
242 FunctionPass
*llvm::createAArch64A53Fix835769() {
243 return new AArch64A53Fix835769();