1//=== ARMCallingConv.cpp - ARM Custom CC Routines ---------------*- C++ -*-===//
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 contains the custom routines for the ARM Calling Convention that
10// aren't done by tablegen, and includes the table generated implementations.
11//
12//===----------------------------------------------------------------------===//
13
14#include "ARM.h"
15#include "ARMCallingConv.h"
16#include "ARMSubtarget.h"
17#include "ARMRegisterInfo.h"
18using namespace llvm;
19
20// APCS f64 is in register pairs, possibly split to stack
21static bool f64AssignAPCS(unsigned ValNo, MVT ValVT, MVT LocVT,
22 CCValAssign::LocInfo LocInfo,
23 CCState &State, bool CanFail) {
24 static const MCPhysReg RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
25
26 // Try to get the first register.
27 if (unsigned Reg = State.AllocateReg(Regs: RegList))
28 State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, RegNo: Reg, LocVT, HTP: LocInfo));
29 else {
30 // For the 2nd half of a v2f64, do not fail.
31 if (CanFail)
32 return false;
33
34 // Put the whole thing on the stack.
35 State.addLoc(V: CCValAssign::getCustomMem(
36 ValNo, ValVT, Offset: State.AllocateStack(Size: 8, Alignment: Align(4)), LocVT, HTP: LocInfo));
37 return true;
38 }
39
40 // Try to get the second register.
41 if (unsigned Reg = State.AllocateReg(Regs: RegList))
42 State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, RegNo: Reg, LocVT, HTP: LocInfo));
43 else
44 State.addLoc(V: CCValAssign::getCustomMem(
45 ValNo, ValVT, Offset: State.AllocateStack(Size: 4, Alignment: Align(4)), LocVT, HTP: LocInfo));
46 return true;
47}
48
49static bool CC_ARM_APCS_Custom_f64(unsigned ValNo, MVT ValVT, MVT LocVT,
50 CCValAssign::LocInfo LocInfo,
51 ISD::ArgFlagsTy ArgFlags,
52 CCState &State) {
53 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, CanFail: true))
54 return false;
55 if (LocVT == MVT::v2f64 &&
56 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, CanFail: false))
57 return false;
58 return true; // we handled it
59}
60
61// AAPCS f64 is in aligned register pairs
62static bool f64AssignAAPCS(unsigned ValNo, MVT ValVT, MVT LocVT,
63 CCValAssign::LocInfo LocInfo,
64 CCState &State, bool CanFail) {
65 static const MCPhysReg HiRegList[] = { ARM::R0, ARM::R2 };
66 static const MCPhysReg LoRegList[] = { ARM::R1, ARM::R3 };
67 static const MCPhysReg ShadowRegList[] = { ARM::R0, ARM::R1 };
68 static const MCPhysReg GPRArgRegs[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
69
70 unsigned Reg = State.AllocateReg(Regs: HiRegList, ShadowRegs: ShadowRegList);
71 if (Reg == 0) {
72
73 // If we had R3 unallocated only, now we still must to waste it.
74 Reg = State.AllocateReg(Regs: GPRArgRegs);
75 assert((!Reg || Reg == ARM::R3) && "Wrong GPRs usage for f64");
76
77 // For the 2nd half of a v2f64, do not just fail.
78 if (CanFail)
79 return false;
80
81 // Put the whole thing on the stack.
82 State.addLoc(V: CCValAssign::getCustomMem(
83 ValNo, ValVT, Offset: State.AllocateStack(Size: 8, Alignment: Align(8)), LocVT, HTP: LocInfo));
84 return true;
85 }
86
87 unsigned i;
88 for (i = 0; i < 2; ++i)
89 if (HiRegList[i] == Reg)
90 break;
91
92 unsigned T = State.AllocateReg(Reg: LoRegList[i]);
93 (void)T;
94 assert(T == LoRegList[i] && "Could not allocate register");
95
96 State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, RegNo: Reg, LocVT, HTP: LocInfo));
97 State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, RegNo: LoRegList[i],
98 LocVT, HTP: LocInfo));
99 return true;
100}
101
102static bool CC_ARM_AAPCS_Custom_f64(unsigned ValNo, MVT ValVT, MVT LocVT,
103 CCValAssign::LocInfo LocInfo,
104 ISD::ArgFlagsTy ArgFlags,
105 CCState &State) {
106 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, CanFail: true))
107 return false;
108 if (LocVT == MVT::v2f64 &&
109 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, CanFail: false))
110 return false;
111 return true; // we handled it
112}
113
114static bool f64RetAssign(unsigned ValNo, MVT ValVT, MVT LocVT,
115 CCValAssign::LocInfo LocInfo, CCState &State) {
116 static const MCPhysReg HiRegList[] = { ARM::R0, ARM::R2 };
117 static const MCPhysReg LoRegList[] = { ARM::R1, ARM::R3 };
118
119 unsigned Reg = State.AllocateReg(Regs: HiRegList, ShadowRegs: LoRegList);
120 if (Reg == 0)
121 return false; // we didn't handle it
122
123 unsigned i;
124 for (i = 0; i < 2; ++i)
125 if (HiRegList[i] == Reg)
126 break;
127
128 State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, RegNo: Reg, LocVT, HTP: LocInfo));
129 State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, RegNo: LoRegList[i],
130 LocVT, HTP: LocInfo));
131 return true;
132}
133
134static bool RetCC_ARM_APCS_Custom_f64(unsigned ValNo, MVT ValVT, MVT LocVT,
135 CCValAssign::LocInfo LocInfo,
136 ISD::ArgFlagsTy ArgFlags,
137 CCState &State) {
138 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
139 return false;
140 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State))
141 return false;
142 return true; // we handled it
143}
144
145static bool RetCC_ARM_AAPCS_Custom_f64(unsigned ValNo, MVT ValVT, MVT LocVT,
146 CCValAssign::LocInfo LocInfo,
147 ISD::ArgFlagsTy ArgFlags,
148 CCState &State) {
149 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags,
150 State);
151}
152
153static const MCPhysReg RRegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 };
154
155static const MCPhysReg SRegList[] = { ARM::S0, ARM::S1, ARM::S2, ARM::S3,
156 ARM::S4, ARM::S5, ARM::S6, ARM::S7,
157 ARM::S8, ARM::S9, ARM::S10, ARM::S11,
158 ARM::S12, ARM::S13, ARM::S14, ARM::S15 };
159static const MCPhysReg DRegList[] = { ARM::D0, ARM::D1, ARM::D2, ARM::D3,
160 ARM::D4, ARM::D5, ARM::D6, ARM::D7 };
161static const MCPhysReg QRegList[] = { ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3 };
162
163
164// Allocate part of an AAPCS HFA or HVA. We assume that each member of the HA
165// has InConsecutiveRegs set, and that the last member also has
166// InConsecutiveRegsLast set. We must process all members of the HA before
167// we can allocate it, as we need to know the total number of registers that
168// will be needed in order to (attempt to) allocate a contiguous block.
169static bool CC_ARM_AAPCS_Custom_Aggregate(unsigned ValNo, MVT ValVT,
170 MVT LocVT,
171 CCValAssign::LocInfo LocInfo,
172 ISD::ArgFlagsTy ArgFlags,
173 CCState &State) {
174 SmallVectorImpl<CCValAssign> &PendingMembers = State.getPendingLocs();
175
176 // AAPCS HFAs must have 1-4 elements, all of the same type
177 if (PendingMembers.size() > 0)
178 assert(PendingMembers[0].getLocVT() == LocVT);
179
180 // Add the argument to the list to be allocated once we know the size of the
181 // aggregate. Store the type's required alignment as extra info for later: in
182 // the [N x i64] case all trace has been removed by the time we actually get
183 // to do allocation.
184 PendingMembers.push_back(Elt: CCValAssign::getPending(
185 ValNo, ValVT, LocVT, HTP: LocInfo, ExtraInfo: ArgFlags.getNonZeroOrigAlign().value()));
186
187 if (!ArgFlags.isInConsecutiveRegsLast())
188 return true;
189
190 // Try to allocate a contiguous block of registers, each of the correct
191 // size to hold one member.
192 auto &DL = State.getMachineFunction().getDataLayout();
193 const Align StackAlign = DL.getStackAlignment();
194 const Align FirstMemberAlign(PendingMembers[0].getExtraInfo());
195 Align Alignment = std::min(a: FirstMemberAlign, b: StackAlign);
196
197 ArrayRef<MCPhysReg> RegList;
198 switch (LocVT.SimpleTy) {
199 case MVT::i32: {
200 RegList = RRegList;
201 unsigned RegIdx = State.getFirstUnallocated(Regs: RegList);
202
203 // First consume all registers that would give an unaligned object. Whether
204 // we go on stack or in regs, no-one will be using them in future.
205 unsigned RegAlign = alignTo(Value: Alignment.value(), Align: 4) / 4;
206 while (RegIdx % RegAlign != 0 && RegIdx < RegList.size())
207 State.AllocateReg(Reg: RegList[RegIdx++]);
208
209 break;
210 }
211 case MVT::f16:
212 case MVT::bf16:
213 case MVT::f32:
214 RegList = SRegList;
215 break;
216 case MVT::v4f16:
217 case MVT::v4bf16:
218 case MVT::f64:
219 RegList = DRegList;
220 break;
221 case MVT::v8f16:
222 case MVT::v8bf16:
223 case MVT::v2f64:
224 RegList = QRegList;
225 break;
226 default:
227 llvm_unreachable("Unexpected member type for block aggregate");
228 break;
229 }
230
231 unsigned RegResult = State.AllocateRegBlock(Regs: RegList, RegsRequired: PendingMembers.size());
232 if (RegResult) {
233 for (CCValAssign &PendingMember : PendingMembers) {
234 PendingMember.convertToReg(RegNo: RegResult);
235 State.addLoc(V: PendingMember);
236 ++RegResult;
237 }
238 PendingMembers.clear();
239 return true;
240 }
241
242 // Register allocation failed, we'll be needing the stack
243 unsigned Size = LocVT.getSizeInBits() / 8;
244 if (LocVT == MVT::i32 && State.getStackSize() == 0) {
245 // If nothing else has used the stack until this point, a non-HFA aggregate
246 // can be split between regs and stack.
247 unsigned RegIdx = State.getFirstUnallocated(Regs: RegList);
248 for (auto &It : PendingMembers) {
249 if (RegIdx >= RegList.size())
250 It.convertToMem(Offset: State.AllocateStack(Size, Alignment: Align(Size)));
251 else
252 It.convertToReg(RegNo: State.AllocateReg(Reg: RegList[RegIdx++]));
253
254 State.addLoc(V: It);
255 }
256 PendingMembers.clear();
257 return true;
258 }
259
260 if (LocVT != MVT::i32)
261 RegList = SRegList;
262
263 // Mark all regs as unavailable (AAPCS rule C.2.vfp for VFP, C.6 for core)
264 for (auto Reg : RegList)
265 State.AllocateReg(Reg);
266
267 // Clamp the alignment between 4 and 8.
268 if (State.getMachineFunction().getSubtarget<ARMSubtarget>().isTargetAEABI())
269 Alignment = ArgFlags.getNonZeroMemAlign() <= 4 ? Align(4) : Align(8);
270
271 // After the first item has been allocated, the rest are packed as tightly as
272 // possible. (E.g. an incoming i64 would have starting Align of 8, but we'll
273 // be allocating a bunch of i32 slots).
274 for (auto &It : PendingMembers) {
275 It.convertToMem(Offset: State.AllocateStack(Size, Alignment));
276 State.addLoc(V: It);
277 Alignment = Align(1);
278 }
279
280 // All pending members have now been allocated
281 PendingMembers.clear();
282
283 // This will be allocated by the last member of the aggregate
284 return true;
285}
286
287static bool CustomAssignInRegList(unsigned ValNo, MVT ValVT, MVT LocVT,
288 CCValAssign::LocInfo LocInfo, CCState &State,
289 ArrayRef<MCPhysReg> RegList) {
290 unsigned Reg = State.AllocateReg(Regs: RegList);
291 if (Reg) {
292 State.addLoc(V: CCValAssign::getCustomReg(ValNo, ValVT, RegNo: Reg, LocVT, HTP: LocInfo));
293 return true;
294 }
295 return false;
296}
297
298static bool CC_ARM_AAPCS_Custom_f16(unsigned ValNo, MVT ValVT, MVT LocVT,
299 CCValAssign::LocInfo LocInfo,
300 ISD::ArgFlagsTy ArgFlags, CCState &State) {
301 // f16 arguments are extended to i32 and assigned to a register in [r0, r3]
302 return CustomAssignInRegList(ValNo, ValVT, LocVT: MVT::i32, LocInfo, State,
303 RegList: RRegList);
304}
305
306static bool CC_ARM_AAPCS_VFP_Custom_f16(unsigned ValNo, MVT ValVT, MVT LocVT,
307 CCValAssign::LocInfo LocInfo,
308 ISD::ArgFlagsTy ArgFlags,
309 CCState &State) {
310 // f16 arguments are extended to f32 and assigned to a register in [s0, s15]
311 return CustomAssignInRegList(ValNo, ValVT, LocVT: MVT::f32, LocInfo, State,
312 RegList: SRegList);
313}
314
315// Include the table generated calling convention implementations.
316#include "ARMGenCallingConv.inc"
317