Change allowsUnalignedMemoryAccesses to take type argument since some targets
[llvm/avr.git] / lib / Target / X86 / X86FloatingPointRegKill.cpp
blob3e0385c79c19e90924433c052799f7b47593395e
1 //===-- X86FloatingPoint.cpp - FP_REG_KILL inserter -----------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines the pass which inserts FP_REG_KILL instructions.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "x86-codegen"
15 #include "X86.h"
16 #include "X86InstrInfo.h"
17 #include "X86Subtarget.h"
18 #include "llvm/Instructions.h"
19 #include "llvm/CodeGen/MachineFunctionPass.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/MachineRegisterInfo.h"
22 #include "llvm/CodeGen/Passes.h"
23 #include "llvm/Target/TargetMachine.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/Compiler.h"
26 #include "llvm/Support/CFG.h"
27 #include "llvm/ADT/Statistic.h"
28 using namespace llvm;
30 STATISTIC(NumFPKill, "Number of FP_REG_KILL instructions added");
32 namespace {
33 struct VISIBILITY_HIDDEN FPRegKiller : public MachineFunctionPass {
34 static char ID;
35 FPRegKiller() : MachineFunctionPass(&ID) {}
37 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
38 AU.setPreservesCFG();
39 AU.addPreservedID(MachineLoopInfoID);
40 AU.addPreservedID(MachineDominatorsID);
41 MachineFunctionPass::getAnalysisUsage(AU);
44 virtual bool runOnMachineFunction(MachineFunction &MF);
46 virtual const char *getPassName() const { return "X86 FP_REG_KILL inserter"; }
48 char FPRegKiller::ID = 0;
51 FunctionPass *llvm::createX87FPRegKillInserterPass() { return new FPRegKiller(); }
53 bool FPRegKiller::runOnMachineFunction(MachineFunction &MF) {
54 // If we are emitting FP stack code, scan the basic block to determine if this
55 // block defines any FP values. If so, put an FP_REG_KILL instruction before
56 // the terminator of the block.
58 // Note that FP stack instructions are used in all modes for long double,
59 // so we always need to do this check.
60 // Also note that it's possible for an FP stack register to be live across
61 // an instruction that produces multiple basic blocks (SSE CMOV) so we
62 // must check all the generated basic blocks.
64 // Scan all of the machine instructions in these MBBs, checking for FP
65 // stores. (RFP32 and RFP64 will not exist in SSE mode, but RFP80 might.)
67 // Fast-path: If nothing is using the x87 registers, we don't need to do
68 // any scanning.
69 MachineRegisterInfo &MRI = MF.getRegInfo();
70 if (MRI.getRegClassVirtRegs(X86::RFP80RegisterClass).empty() &&
71 MRI.getRegClassVirtRegs(X86::RFP64RegisterClass).empty() &&
72 MRI.getRegClassVirtRegs(X86::RFP32RegisterClass).empty())
73 return false;
75 bool Changed = false;
76 const X86Subtarget &Subtarget = MF.getTarget().getSubtarget<X86Subtarget>();
77 MachineFunction::iterator MBBI = MF.begin();
78 MachineFunction::iterator EndMBB = MF.end();
79 for (; MBBI != EndMBB; ++MBBI) {
80 MachineBasicBlock *MBB = MBBI;
82 // If this block returns, ignore it. We don't want to insert an FP_REG_KILL
83 // before the return.
84 if (!MBB->empty()) {
85 MachineBasicBlock::iterator EndI = MBB->end();
86 --EndI;
87 if (EndI->getDesc().isReturn())
88 continue;
91 bool ContainsFPCode = false;
92 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
93 !ContainsFPCode && I != E; ++I) {
94 if (I->getNumOperands() != 0 && I->getOperand(0).isReg()) {
95 const TargetRegisterClass *clas;
96 for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) {
97 if (I->getOperand(op).isReg() && I->getOperand(op).isDef() &&
98 TargetRegisterInfo::isVirtualRegister(I->getOperand(op).getReg()) &&
99 ((clas = MRI.getRegClass(I->getOperand(op).getReg())) ==
100 X86::RFP32RegisterClass ||
101 clas == X86::RFP64RegisterClass ||
102 clas == X86::RFP80RegisterClass)) {
103 ContainsFPCode = true;
104 break;
109 // Check PHI nodes in successor blocks. These PHI's will be lowered to have
110 // a copy of the input value in this block. In SSE mode, we only care about
111 // 80-bit values.
112 if (!ContainsFPCode) {
113 // Final check, check LLVM BB's that are successors to the LLVM BB
114 // corresponding to BB for FP PHI nodes.
115 const BasicBlock *LLVMBB = MBB->getBasicBlock();
116 const PHINode *PN;
117 for (succ_const_iterator SI = succ_begin(LLVMBB), E = succ_end(LLVMBB);
118 !ContainsFPCode && SI != E; ++SI) {
119 for (BasicBlock::const_iterator II = SI->begin();
120 (PN = dyn_cast<PHINode>(II)); ++II) {
121 if (PN->getType()==Type::getX86_FP80Ty(LLVMBB->getContext()) ||
122 (!Subtarget.hasSSE1() && PN->getType()->isFloatingPoint()) ||
123 (!Subtarget.hasSSE2() &&
124 PN->getType()==Type::getDoubleTy(LLVMBB->getContext()))) {
125 ContainsFPCode = true;
126 break;
131 // Finally, if we found any FP code, emit the FP_REG_KILL instruction.
132 if (ContainsFPCode) {
133 BuildMI(*MBB, MBBI->getFirstTerminator(), DebugLoc::getUnknownLoc(),
134 MF.getTarget().getInstrInfo()->get(X86::FP_REG_KILL));
135 ++NumFPKill;
136 Changed = true;
140 return Changed;