1 //===-- NVPTXPrologEpilogPass.cpp - NVPTX prolog/epilog inserter ----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file is a copy of the generic LLVM PrologEpilogInserter pass, modified
11 // to remove unneeded functionality and to handle virtual registers. Most code
12 // here is a copy of PrologEpilogInserter.cpp.
14 //===----------------------------------------------------------------------===//
17 #include "llvm/CodeGen/MachineFrameInfo.h"
18 #include "llvm/CodeGen/MachineFunction.h"
19 #include "llvm/CodeGen/MachineFunctionPass.h"
20 #include "llvm/Pass.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Support/raw_ostream.h"
23 #include "llvm/Target/TargetFrameLowering.h"
24 #include "llvm/Target/TargetRegisterInfo.h"
25 #include "llvm/Target/TargetSubtargetInfo.h"
29 #define DEBUG_TYPE "nvptx-prolog-epilog"
32 class NVPTXPrologEpilogPass : public MachineFunctionPass {
35 NVPTXPrologEpilogPass() : MachineFunctionPass(ID) {}
37 bool runOnMachineFunction(MachineFunction &MF) override;
40 void calculateFrameObjectOffsets(MachineFunction &Fn);
44 MachineFunctionPass *llvm::createNVPTXPrologEpilogPass() {
45 return new NVPTXPrologEpilogPass();
48 char NVPTXPrologEpilogPass::ID = 0;
50 bool NVPTXPrologEpilogPass::runOnMachineFunction(MachineFunction &MF) {
51 const TargetSubtargetInfo &STI = MF.getSubtarget();
52 const TargetFrameLowering &TFI = *STI.getFrameLowering();
53 const TargetRegisterInfo &TRI = *STI.getRegisterInfo();
54 bool Modified = false;
56 calculateFrameObjectOffsets(MF);
58 for (MachineBasicBlock &MBB : MF) {
59 for (MachineInstr &MI : MBB) {
60 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
61 if (!MI.getOperand(i).isFI())
63 TRI.eliminateFrameIndex(MI, 0, i, nullptr);
69 // Add function prolog/epilog
70 TFI.emitPrologue(MF, MF.front());
72 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
73 // If last instruction is a return instruction, add an epilogue
74 if (I->isReturnBlock())
75 TFI.emitEpilogue(MF, *I);
81 /// AdjustStackOffset - Helper function used to adjust the stack frame offset.
83 AdjustStackOffset(MachineFrameInfo &MFI, int FrameIdx,
84 bool StackGrowsDown, int64_t &Offset,
86 // If the stack grows down, add the object size to find the lowest address.
88 Offset += MFI.getObjectSize(FrameIdx);
90 unsigned Align = MFI.getObjectAlignment(FrameIdx);
92 // If the alignment of this object is greater than that of the stack, then
93 // increase the stack alignment to match.
94 MaxAlign = std::max(MaxAlign, Align);
96 // Adjust to alignment boundary.
97 Offset = (Offset + Align - 1) / Align * Align;
100 DEBUG(dbgs() << "alloc FI(" << FrameIdx << ") at SP[" << -Offset << "]\n");
101 MFI.setObjectOffset(FrameIdx, -Offset); // Set the computed offset
103 DEBUG(dbgs() << "alloc FI(" << FrameIdx << ") at SP[" << Offset << "]\n");
104 MFI.setObjectOffset(FrameIdx, Offset);
105 Offset += MFI.getObjectSize(FrameIdx);
110 NVPTXPrologEpilogPass::calculateFrameObjectOffsets(MachineFunction &Fn) {
111 const TargetFrameLowering &TFI = *Fn.getSubtarget().getFrameLowering();
112 const TargetRegisterInfo *RegInfo = Fn.getSubtarget().getRegisterInfo();
114 bool StackGrowsDown =
115 TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
117 // Loop over all of the stack objects, assigning sequential addresses...
118 MachineFrameInfo &MFI = Fn.getFrameInfo();
120 // Start at the beginning of the local area.
121 // The Offset is the distance from the stack top in the direction
122 // of stack growth -- so it's always nonnegative.
123 int LocalAreaOffset = TFI.getOffsetOfLocalArea();
125 LocalAreaOffset = -LocalAreaOffset;
126 assert(LocalAreaOffset >= 0
127 && "Local area offset should be in direction of stack growth");
128 int64_t Offset = LocalAreaOffset;
130 // If there are fixed sized objects that are preallocated in the local area,
131 // non-fixed objects can't be allocated right at the start of local area.
132 // We currently don't support filling in holes in between fixed sized
133 // objects, so we adjust 'Offset' to point to the end of last fixed sized
134 // preallocated object.
135 for (int i = MFI.getObjectIndexBegin(); i != 0; ++i) {
137 if (StackGrowsDown) {
138 // The maximum distance from the stack pointer is at lower address of
139 // the object -- which is given by offset. For down growing stack
140 // the offset is negative, so we negate the offset to get the distance.
141 FixedOff = -MFI.getObjectOffset(i);
143 // The maximum distance from the start pointer is at the upper
144 // address of the object.
145 FixedOff = MFI.getObjectOffset(i) + MFI.getObjectSize(i);
147 if (FixedOff > Offset) Offset = FixedOff;
150 // NOTE: We do not have a call stack
152 unsigned MaxAlign = MFI.getMaxAlignment();
156 // FIXME: Once this is working, then enable flag will change to a target
157 // check for whether the frame is large enough to want to use virtual
158 // frame index registers. Functions which don't want/need this optimization
159 // will continue to use the existing code path.
160 if (MFI.getUseLocalStackAllocationBlock()) {
161 unsigned Align = MFI.getLocalFrameMaxAlign();
163 // Adjust to alignment boundary.
164 Offset = (Offset + Align - 1) / Align * Align;
166 DEBUG(dbgs() << "Local frame base offset: " << Offset << "\n");
168 // Resolve offsets for objects in the local block.
169 for (unsigned i = 0, e = MFI.getLocalFrameObjectCount(); i != e; ++i) {
170 std::pair<int, int64_t> Entry = MFI.getLocalFrameObjectMap(i);
171 int64_t FIOffset = (StackGrowsDown ? -Offset : Offset) + Entry.second;
172 DEBUG(dbgs() << "alloc FI(" << Entry.first << ") at SP[" <<
174 MFI.setObjectOffset(Entry.first, FIOffset);
176 // Allocate the local block
177 Offset += MFI.getLocalFrameSize();
179 MaxAlign = std::max(Align, MaxAlign);
182 // No stack protector
184 // Then assign frame offsets to stack objects that are not used to spill
185 // callee saved registers.
186 for (unsigned i = 0, e = MFI.getObjectIndexEnd(); i != e; ++i) {
187 if (MFI.isObjectPreAllocated(i) &&
188 MFI.getUseLocalStackAllocationBlock())
190 if (MFI.isDeadObjectIndex(i))
193 AdjustStackOffset(MFI, i, StackGrowsDown, Offset, MaxAlign);
198 if (!TFI.targetHandlesStackFrameRounding()) {
199 // If we have reserved argument space for call sites in the function
200 // immediately on entry to the current function, count it as part of the
201 // overall stack size.
202 if (MFI.adjustsStack() && TFI.hasReservedCallFrame(Fn))
203 Offset += MFI.getMaxCallFrameSize();
205 // Round up the size to a multiple of the alignment. If the function has
206 // any calls or alloca's, align to the target's StackAlignment value to
207 // ensure that the callee's frame or the alloca data is suitably aligned;
208 // otherwise, for leaf functions, align to the TransientStackAlignment
211 if (MFI.adjustsStack() || MFI.hasVarSizedObjects() ||
212 (RegInfo->needsStackRealignment(Fn) && MFI.getObjectIndexEnd() != 0))
213 StackAlign = TFI.getStackAlignment();
215 StackAlign = TFI.getTransientStackAlignment();
217 // If the frame pointer is eliminated, all frame offsets will be relative to
218 // SP not FP. Align to MaxAlign so this works.
219 StackAlign = std::max(StackAlign, MaxAlign);
220 unsigned AlignMask = StackAlign - 1;
221 Offset = (Offset + AlignMask) & ~uint64_t(AlignMask);
224 // Update frame info to pretend that this is part of the stack...
225 int64_t StackSize = Offset - LocalAreaOffset;
226 MFI.setStackSize(StackSize);