1//===-- NVPTXPrologEpilogPass.cpp - NVPTX prolog/epilog inserter ----------===//
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 is a copy of the generic LLVM PrologEpilogInserter pass, modified
10// to remove unneeded functionality and to handle virtual registers. Most code
11// here is a copy of PrologEpilogInserter.cpp.
12//
13//===----------------------------------------------------------------------===//
14
15#include "NVPTX.h"
16#include "llvm/CodeGen/MachineFrameInfo.h"
17#include "llvm/CodeGen/MachineFunction.h"
18#include "llvm/CodeGen/MachineFunctionPass.h"
19#include "llvm/CodeGen/TargetFrameLowering.h"
20#include "llvm/CodeGen/TargetRegisterInfo.h"
21#include "llvm/CodeGen/TargetSubtargetInfo.h"
22#include "llvm/IR/DebugInfoMetadata.h"
23#include "llvm/Pass.h"
24#include "llvm/Support/Debug.h"
25#include "llvm/Support/raw_ostream.h"
26
27using namespace llvm;
28
29#define DEBUG_TYPE "nvptx-prolog-epilog"
30
31namespace {
32class NVPTXPrologEpilogPass : public MachineFunctionPass {
33public:
34 static char ID;
35 NVPTXPrologEpilogPass() : MachineFunctionPass(ID) {}
36
37 bool runOnMachineFunction(MachineFunction &MF) override;
38
39 StringRef getPassName() const override { return "NVPTX Prolog Epilog Pass"; }
40
41private:
42 void calculateFrameObjectOffsets(MachineFunction &Fn);
43};
44}
45
46MachineFunctionPass *llvm::createNVPTXPrologEpilogPass() {
47 return new NVPTXPrologEpilogPass();
48}
49
50char NVPTXPrologEpilogPass::ID = 0;
51
52bool NVPTXPrologEpilogPass::runOnMachineFunction(MachineFunction &MF) {
53 const TargetSubtargetInfo &STI = MF.getSubtarget();
54 const TargetFrameLowering &TFI = *STI.getFrameLowering();
55 const TargetRegisterInfo &TRI = *STI.getRegisterInfo();
56 bool Modified = false;
57
58 calculateFrameObjectOffsets(Fn&: MF);
59
60 for (MachineBasicBlock &MBB : MF) {
61 for (MachineInstr &MI : MBB) {
62 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
63 if (!MI.getOperand(i).isFI())
64 continue;
65
66 // Frame indices in debug values are encoded in a target independent
67 // way with simply the frame index and offset rather than any
68 // target-specific addressing mode.
69 if (MI.isDebugValue()) {
70 MachineOperand &Op = MI.getOperand(i);
71 assert(
72 MI.isDebugOperand(&Op) &&
73 "Frame indices can only appear as a debug operand in a DBG_VALUE*"
74 " machine instruction");
75 Register Reg;
76 auto Offset =
77 TFI.getFrameIndexReference(MF, FI: Op.getIndex(), FrameReg&: Reg);
78 Op.ChangeToRegister(Reg, /*isDef=*/false);
79 const DIExpression *DIExpr = MI.getDebugExpression();
80 if (MI.isNonListDebugValue()) {
81 DIExpr = TRI.prependOffsetExpression(Expr: MI.getDebugExpression(), PrependFlags: DIExpression::ApplyOffset, Offset);
82 } else {
83 SmallVector<uint64_t, 3> Ops;
84 TRI.getOffsetOpcodes(Offset, Ops);
85 unsigned OpIdx = MI.getDebugOperandIndex(Op: &Op);
86 DIExpr = DIExpression::appendOpsToArg(Expr: DIExpr, Ops, ArgNo: OpIdx);
87 }
88 MI.getDebugExpressionOp().setMetadata(DIExpr);
89 continue;
90 }
91
92 TRI.eliminateFrameIndex(MI, SPAdj: 0, FIOperandNum: i, RS: nullptr);
93 Modified = true;
94 }
95 }
96 }
97
98 // Add function prolog/epilog
99 TFI.emitPrologue(MF, MBB&: MF.front());
100
101 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
102 // If last instruction is a return instruction, add an epilogue
103 if (I->isReturnBlock())
104 TFI.emitEpilogue(MF, MBB&: *I);
105 }
106
107 return Modified;
108}
109
110/// AdjustStackOffset - Helper function used to adjust the stack frame offset.
111static inline void AdjustStackOffset(MachineFrameInfo &MFI, int FrameIdx,
112 bool StackGrowsDown, int64_t &Offset,
113 Align &MaxAlign) {
114 // If the stack grows down, add the object size to find the lowest address.
115 if (StackGrowsDown)
116 Offset += MFI.getObjectSize(ObjectIdx: FrameIdx);
117
118 Align Alignment = MFI.getObjectAlign(ObjectIdx: FrameIdx);
119
120 // If the alignment of this object is greater than that of the stack, then
121 // increase the stack alignment to match.
122 MaxAlign = std::max(a: MaxAlign, b: Alignment);
123
124 // Adjust to alignment boundary.
125 Offset = alignTo(Size: Offset, A: Alignment);
126
127 if (StackGrowsDown) {
128 LLVM_DEBUG(dbgs() << "alloc FI(" << FrameIdx << ") at SP[" << -Offset
129 << "]\n");
130 MFI.setObjectOffset(ObjectIdx: FrameIdx, SPOffset: -Offset); // Set the computed offset
131 } else {
132 LLVM_DEBUG(dbgs() << "alloc FI(" << FrameIdx << ") at SP[" << Offset
133 << "]\n");
134 MFI.setObjectOffset(ObjectIdx: FrameIdx, SPOffset: Offset);
135 Offset += MFI.getObjectSize(ObjectIdx: FrameIdx);
136 }
137}
138
139void
140NVPTXPrologEpilogPass::calculateFrameObjectOffsets(MachineFunction &Fn) {
141 const TargetFrameLowering &TFI = *Fn.getSubtarget().getFrameLowering();
142 const TargetRegisterInfo *RegInfo = Fn.getSubtarget().getRegisterInfo();
143
144 bool StackGrowsDown =
145 TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
146
147 // Loop over all of the stack objects, assigning sequential addresses...
148 MachineFrameInfo &MFI = Fn.getFrameInfo();
149
150 // Start at the beginning of the local area.
151 // The Offset is the distance from the stack top in the direction
152 // of stack growth -- so it's always nonnegative.
153 int LocalAreaOffset = TFI.getOffsetOfLocalArea();
154 if (StackGrowsDown)
155 LocalAreaOffset = -LocalAreaOffset;
156 assert(LocalAreaOffset >= 0
157 && "Local area offset should be in direction of stack growth");
158 int64_t Offset = LocalAreaOffset;
159
160 // If there are fixed sized objects that are preallocated in the local area,
161 // non-fixed objects can't be allocated right at the start of local area.
162 // We currently don't support filling in holes in between fixed sized
163 // objects, so we adjust 'Offset' to point to the end of last fixed sized
164 // preallocated object.
165 for (int i = MFI.getObjectIndexBegin(); i != 0; ++i) {
166 int64_t FixedOff;
167 if (StackGrowsDown) {
168 // The maximum distance from the stack pointer is at lower address of
169 // the object -- which is given by offset. For down growing stack
170 // the offset is negative, so we negate the offset to get the distance.
171 FixedOff = -MFI.getObjectOffset(ObjectIdx: i);
172 } else {
173 // The maximum distance from the start pointer is at the upper
174 // address of the object.
175 FixedOff = MFI.getObjectOffset(ObjectIdx: i) + MFI.getObjectSize(ObjectIdx: i);
176 }
177 if (FixedOff > Offset) Offset = FixedOff;
178 }
179
180 // NOTE: We do not have a call stack
181
182 Align MaxAlign = MFI.getMaxAlign();
183
184 // No scavenger
185
186 // FIXME: Once this is working, then enable flag will change to a target
187 // check for whether the frame is large enough to want to use virtual
188 // frame index registers. Functions which don't want/need this optimization
189 // will continue to use the existing code path.
190 if (MFI.getUseLocalStackAllocationBlock()) {
191 Align Alignment = MFI.getLocalFrameMaxAlign();
192
193 // Adjust to alignment boundary.
194 Offset = alignTo(Size: Offset, A: Alignment);
195
196 LLVM_DEBUG(dbgs() << "Local frame base offset: " << Offset << "\n");
197
198 // Resolve offsets for objects in the local block.
199 for (unsigned i = 0, e = MFI.getLocalFrameObjectCount(); i != e; ++i) {
200 std::pair<int, int64_t> Entry = MFI.getLocalFrameObjectMap(i);
201 int64_t FIOffset = (StackGrowsDown ? -Offset : Offset) + Entry.second;
202 LLVM_DEBUG(dbgs() << "alloc FI(" << Entry.first << ") at SP[" << FIOffset
203 << "]\n");
204 MFI.setObjectOffset(ObjectIdx: Entry.first, SPOffset: FIOffset);
205 }
206 // Allocate the local block
207 Offset += MFI.getLocalFrameSize();
208
209 MaxAlign = std::max(a: Alignment, b: MaxAlign);
210 }
211
212 // No stack protector
213
214 // Then assign frame offsets to stack objects that are not used to spill
215 // callee saved registers.
216 for (unsigned i = 0, e = MFI.getObjectIndexEnd(); i != e; ++i) {
217 if (MFI.isObjectPreAllocated(ObjectIdx: i) &&
218 MFI.getUseLocalStackAllocationBlock())
219 continue;
220 if (MFI.isDeadObjectIndex(ObjectIdx: i))
221 continue;
222
223 AdjustStackOffset(MFI, FrameIdx: i, StackGrowsDown, Offset, MaxAlign);
224 }
225
226 // No scavenger
227
228 if (!TFI.targetHandlesStackFrameRounding()) {
229 // If we have reserved argument space for call sites in the function
230 // immediately on entry to the current function, count it as part of the
231 // overall stack size.
232 if (MFI.adjustsStack() && TFI.hasReservedCallFrame(MF: Fn))
233 Offset += MFI.getMaxCallFrameSize();
234
235 // Round up the size to a multiple of the alignment. If the function has
236 // any calls or alloca's, align to the target's StackAlignment value to
237 // ensure that the callee's frame or the alloca data is suitably aligned;
238 // otherwise, for leaf functions, align to the TransientStackAlignment
239 // value.
240 Align StackAlign;
241 if (MFI.adjustsStack() || MFI.hasVarSizedObjects() ||
242 (RegInfo->hasStackRealignment(MF: Fn) && MFI.getObjectIndexEnd() != 0))
243 StackAlign = TFI.getStackAlign();
244 else
245 StackAlign = TFI.getTransientStackAlign();
246
247 // If the frame pointer is eliminated, all frame offsets will be relative to
248 // SP not FP. Align to MaxAlign so this works.
249 Offset = alignTo(Size: Offset, A: std::max(a: StackAlign, b: MaxAlign));
250 }
251
252 // Update frame info to pretend that this is part of the stack...
253 int64_t StackSize = Offset - LocalAreaOffset;
254 MFI.setStackSize(StackSize);
255}
256