1//=== ARMCallingConv.h - ARM Custom Calling Convention Routines -*- C++ -*-===//
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 contains the custom routines for the ARM Calling Convention that
11// aren't done by tablegen.
12//
13//===----------------------------------------------------------------------===//
14
15#ifndef LLVM_LIB_TARGET_ARM_ARMCALLINGCONV_H
16#define LLVM_LIB_TARGET_ARM_ARMCALLINGCONV_H
17
18#include "ARM.h"
19#include "ARMBaseInstrInfo.h"
20#include "ARMSubtarget.h"
21#include "llvm/CodeGen/CallingConvLower.h"
22#include "llvm/IR/CallingConv.h"
23#include "llvm/Target/TargetInstrInfo.h"
24
25namespace llvm {
26
27// APCS f64 is in register pairs, possibly split to stack
28static bool f64AssignAPCS(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
29                          CCValAssign::LocInfo &LocInfo,
30                          CCState &State, bool CanFail) {
31  static const MCPhysReg RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
32
33  // Try to get the first register.
34  if (unsigned Reg = State.AllocateReg(RegList))
35    State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
36  else {
37    // For the 2nd half of a v2f64, do not fail.
38    if (CanFail)
39      return false;
40
41    // Put the whole thing on the stack.
42    State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
43                                           State.AllocateStack(8, 4),
44                                           LocVT, LocInfo));
45    return true;
46  }
47
48  // Try to get the second register.
49  if (unsigned Reg = State.AllocateReg(RegList))
50    State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
51  else
52    State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
53                                           State.AllocateStack(4, 4),
54                                           LocVT, LocInfo));
55  return true;
56}
57
58static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
59                                   CCValAssign::LocInfo &LocInfo,
60                                   ISD::ArgFlagsTy &ArgFlags,
61                                   CCState &State) {
62  if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
63    return false;
64  if (LocVT == MVT::v2f64 &&
65      !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
66    return false;
67  return true;  // we handled it
68}
69
70// AAPCS f64 is in aligned register pairs
71static bool f64AssignAAPCS(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
72                           CCValAssign::LocInfo &LocInfo,
73                           CCState &State, bool CanFail) {
74  static const MCPhysReg HiRegList[] = { ARM::R0, ARM::R2 };
75  static const MCPhysReg LoRegList[] = { ARM::R1, ARM::R3 };
76  static const MCPhysReg ShadowRegList[] = { ARM::R0, ARM::R1 };
77  static const MCPhysReg GPRArgRegs[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
78
79  unsigned Reg = State.AllocateReg(HiRegList, ShadowRegList);
80  if (Reg == 0) {
81
82    // If we had R3 unallocated only, now we still must to waste it.
83    Reg = State.AllocateReg(GPRArgRegs);
84    assert((!Reg || Reg == ARM::R3) && "Wrong GPRs usage for f64");
85
86    // For the 2nd half of a v2f64, do not just fail.
87    if (CanFail)
88      return false;
89
90    // Put the whole thing on the stack.
91    State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
92                                           State.AllocateStack(8, 8),
93                                           LocVT, LocInfo));
94    return true;
95  }
96
97  unsigned i;
98  for (i = 0; i < 2; ++i)
99    if (HiRegList[i] == Reg)
100      break;
101
102  unsigned T = State.AllocateReg(LoRegList[i]);
103  (void)T;
104  assert(T == LoRegList[i] && "Could not allocate register");
105
106  State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
107  State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
108                                         LocVT, LocInfo));
109  return true;
110}
111
112static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
113                                    CCValAssign::LocInfo &LocInfo,
114                                    ISD::ArgFlagsTy &ArgFlags,
115                                    CCState &State) {
116  if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true))
117    return false;
118  if (LocVT == MVT::v2f64 &&
119      !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false))
120    return false;
121  return true;  // we handled it
122}
123
124static bool f64RetAssign(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
125                         CCValAssign::LocInfo &LocInfo, CCState &State) {
126  static const MCPhysReg HiRegList[] = { ARM::R0, ARM::R2 };
127  static const MCPhysReg LoRegList[] = { ARM::R1, ARM::R3 };
128
129  unsigned Reg = State.AllocateReg(HiRegList, LoRegList);
130  if (Reg == 0)
131    return false; // we didn't handle it
132
133  unsigned i;
134  for (i = 0; i < 2; ++i)
135    if (HiRegList[i] == Reg)
136      break;
137
138  State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo));
139  State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i],
140                                         LocVT, LocInfo));
141  return true;
142}
143
144static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
145                                      CCValAssign::LocInfo &LocInfo,
146                                      ISD::ArgFlagsTy &ArgFlags,
147                                      CCState &State) {
148  if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
149    return false;
150  if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
151    return false;
152  return true;  // we handled it
153}
154
155static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
156                                       CCValAssign::LocInfo &LocInfo,
157                                       ISD::ArgFlagsTy &ArgFlags,
158                                       CCState &State) {
159  return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
160                                   State);
161}
162
163static const MCPhysReg RRegList[] = { ARM::R0,  ARM::R1,  ARM::R2,  ARM::R3 };
164
165static const MCPhysReg SRegList[] = { ARM::S0,  ARM::S1,  ARM::S2,  ARM::S3,
166                                      ARM::S4,  ARM::S5,  ARM::S6,  ARM::S7,
167                                      ARM::S8,  ARM::S9,  ARM::S10, ARM::S11,
168                                      ARM::S12, ARM::S13, ARM::S14,  ARM::S15 };
169static const MCPhysReg DRegList[] = { ARM::D0, ARM::D1, ARM::D2, ARM::D3,
170                                      ARM::D4, ARM::D5, ARM::D6, ARM::D7 };
171static const MCPhysReg QRegList[] = { ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3 };
172
173
174// Allocate part of an AAPCS HFA or HVA. We assume that each member of the HA
175// has InConsecutiveRegs set, and that the last member also has
176// InConsecutiveRegsLast set. We must process all members of the HA before
177// we can allocate it, as we need to know the total number of registers that
178// will be needed in order to (attempt to) allocate a contiguous block.
179static bool CC_ARM_AAPCS_Custom_Aggregate(unsigned &ValNo, MVT &ValVT,
180                                          MVT &LocVT,
181                                          CCValAssign::LocInfo &LocInfo,
182                                          ISD::ArgFlagsTy &ArgFlags,
183                                          CCState &State) {
184  SmallVectorImpl<CCValAssign> &PendingMembers = State.getPendingLocs();
185
186  // AAPCS HFAs must have 1-4 elements, all of the same type
187  if (PendingMembers.size() > 0)
188    assert(PendingMembers[0].getLocVT() == LocVT);
189
190  // Add the argument to the list to be allocated once we know the size of the
191  // aggregate. Store the type's required alignmnent as extra info for later: in
192  // the [N x i64] case all trace has been removed by the time we actually get
193  // to do allocation.
194  PendingMembers.push_back(CCValAssign::getPending(ValNo, ValVT, LocVT, LocInfo,
195                                                   ArgFlags.getOrigAlign()));
196
197  if (!ArgFlags.isInConsecutiveRegsLast())
198    return true;
199
200  // Try to allocate a contiguous block of registers, each of the correct
201  // size to hold one member.
202  auto &DL = State.getMachineFunction().getDataLayout();
203  unsigned StackAlign = DL.getStackAlignment();
204  unsigned Align = std::min(PendingMembers[0].getExtraInfo(), StackAlign);
205
206  ArrayRef<MCPhysReg> RegList;
207  switch (LocVT.SimpleTy) {
208  case MVT::i32: {
209    RegList = RRegList;
210    unsigned RegIdx = State.getFirstUnallocated(RegList);
211
212    // First consume all registers that would give an unaligned object. Whether
213    // we go on stack or in regs, no-one will be using them in future.
214    unsigned RegAlign = alignTo(Align, 4) / 4;
215    while (RegIdx % RegAlign != 0 && RegIdx < RegList.size())
216      State.AllocateReg(RegList[RegIdx++]);
217
218    break;
219  }
220  case MVT::f32:
221    RegList = SRegList;
222    break;
223  case MVT::f64:
224    RegList = DRegList;
225    break;
226  case MVT::v2f64:
227    RegList = QRegList;
228    break;
229  default:
230    llvm_unreachable("Unexpected member type for block aggregate");
231    break;
232  }
233
234  unsigned RegResult = State.AllocateRegBlock(RegList, PendingMembers.size());
235  if (RegResult) {
236    for (SmallVectorImpl<CCValAssign>::iterator It = PendingMembers.begin();
237         It != PendingMembers.end(); ++It) {
238      It->convertToReg(RegResult);
239      State.addLoc(*It);
240      ++RegResult;
241    }
242    PendingMembers.clear();
243    return true;
244  }
245
246  // Register allocation failed, we'll be needing the stack
247  unsigned Size = LocVT.getSizeInBits() / 8;
248  if (LocVT == MVT::i32 && State.getNextStackOffset() == 0) {
249    // If nothing else has used the stack until this point, a non-HFA aggregate
250    // can be split between regs and stack.
251    unsigned RegIdx = State.getFirstUnallocated(RegList);
252    for (auto &It : PendingMembers) {
253      if (RegIdx >= RegList.size())
254        It.convertToMem(State.AllocateStack(Size, Size));
255      else
256        It.convertToReg(State.AllocateReg(RegList[RegIdx++]));
257
258      State.addLoc(It);
259    }
260    PendingMembers.clear();
261    return true;
262  } else if (LocVT != MVT::i32)
263    RegList = SRegList;
264
265  // Mark all regs as unavailable (AAPCS rule C.2.vfp for VFP, C.6 for core)
266  for (auto Reg : RegList)
267    State.AllocateReg(Reg);
268
269  for (auto &It : PendingMembers) {
270    It.convertToMem(State.AllocateStack(Size, Align));
271    State.addLoc(It);
272
273    // After the first item has been allocated, the rest are packed as tightly
274    // as possible. (E.g. an incoming i64 would have starting Align of 8, but
275    // we'll be allocating a bunch of i32 slots).
276    Align = Size;
277  }
278
279  // All pending members have now been allocated
280  PendingMembers.clear();
281
282  // This will be allocated by the last member of the aggregate
283  return true;
284}
285
286} // End llvm namespace
287
288#endif
289