AMDGPU: Mark test as XFAIL in expensive_checks builds
[llvm-project.git] / llvm / lib / Target / VE / VERegisterInfo.cpp
blobf381b7d321598d6cfc958f5bd2430f7b0d7374c4
1 //===-- VERegisterInfo.cpp - VE Register Information ----------------------===//
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 VE implementation of the TargetRegisterInfo class.
11 //===----------------------------------------------------------------------===//
13 #include "VERegisterInfo.h"
14 #include "VE.h"
15 #include "VESubtarget.h"
16 #include "llvm/ADT/BitVector.h"
17 #include "llvm/CodeGen/MachineFrameInfo.h"
18 #include "llvm/CodeGen/MachineFunction.h"
19 #include "llvm/CodeGen/MachineInstrBuilder.h"
20 #include "llvm/CodeGen/TargetInstrInfo.h"
21 #include "llvm/IR/Type.h"
22 #include "llvm/Support/Debug.h"
24 using namespace llvm;
26 #define DEBUG_TYPE "ve-register-info"
28 #define GET_REGINFO_TARGET_DESC
29 #include "VEGenRegisterInfo.inc"
31 // VE uses %s10 == %lp to keep return address
32 VERegisterInfo::VERegisterInfo() : VEGenRegisterInfo(VE::SX10) {}
34 const MCPhysReg *
35 VERegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
36 switch (MF->getFunction().getCallingConv()) {
37 case CallingConv::Fast:
38 // Being explicit (same as standard CC).
39 default:
40 return CSR_SaveList;
41 case CallingConv::PreserveAll:
42 return CSR_preserve_all_SaveList;
46 const uint32_t *VERegisterInfo::getCallPreservedMask(const MachineFunction &MF,
47 CallingConv::ID CC) const {
48 switch (CC) {
49 case CallingConv::Fast:
50 // Being explicit (same as standard CC).
51 default:
52 return CSR_RegMask;
53 case CallingConv::PreserveAll:
54 return CSR_preserve_all_RegMask;
58 const uint32_t *VERegisterInfo::getNoPreservedMask() const {
59 return CSR_NoRegs_RegMask;
62 BitVector VERegisterInfo::getReservedRegs(const MachineFunction &MF) const {
63 BitVector Reserved(getNumRegs());
65 const Register ReservedRegs[] = {
66 VE::SX8, // Stack limit
67 VE::SX9, // Frame pointer
68 VE::SX10, // Link register (return address)
69 VE::SX11, // Stack pointer
71 // FIXME: maybe not need to be reserved
72 VE::SX12, // Outer register
73 VE::SX13, // Id register for dynamic linker
75 VE::SX14, // Thread pointer
76 VE::SX15, // Global offset table register
77 VE::SX16, // Procedure linkage table register
78 VE::SX17, // Linkage-area register
79 // sx18-sx33 are callee-saved registers
80 // sx34-sx63 are temporary registers
83 for (auto R : ReservedRegs)
84 for (MCRegAliasIterator ItAlias(R, this, true); ItAlias.isValid();
85 ++ItAlias)
86 Reserved.set(*ItAlias);
88 // Reserve constant registers.
89 Reserved.set(VE::VM0);
90 Reserved.set(VE::VMP0);
92 return Reserved;
95 const TargetRegisterClass *
96 VERegisterInfo::getPointerRegClass(const MachineFunction &MF,
97 unsigned Kind) const {
98 return &VE::I64RegClass;
101 static unsigned offsetToDisp(MachineInstr &MI) {
102 // Default offset in instruction's operands (reg+reg+imm).
103 unsigned OffDisp = 2;
105 #define RRCAS_multi_cases(NAME) NAME##rir : case NAME##rii
108 using namespace llvm::VE;
109 switch (MI.getOpcode()) {
110 case INLINEASM:
111 case RRCAS_multi_cases(TS1AML):
112 case RRCAS_multi_cases(TS1AMW):
113 case RRCAS_multi_cases(CASL):
114 case RRCAS_multi_cases(CASW):
115 // These instructions use AS format (reg+imm).
116 OffDisp = 1;
117 break;
120 #undef RRCAS_multi_cases
122 return OffDisp;
125 namespace {
126 class EliminateFrameIndex {
127 const TargetInstrInfo &TII;
128 const TargetRegisterInfo &TRI;
129 const DebugLoc &DL;
130 MachineBasicBlock &MBB;
131 MachineBasicBlock::iterator II;
132 Register clobber;
134 // Some helper functions for the ease of instruction building.
135 MachineFunction &getFunc() const { return *MBB.getParent(); }
136 inline MCRegister getSubReg(MCRegister Reg, unsigned Idx) const {
137 return TRI.getSubReg(Reg, Idx);
139 inline const MCInstrDesc &get(unsigned Opcode) const {
140 return TII.get(Opcode);
142 inline MachineInstrBuilder build(const MCInstrDesc &MCID, Register DestReg) {
143 return BuildMI(MBB, II, DL, MCID, DestReg);
145 inline MachineInstrBuilder build(unsigned InstOpc, Register DestReg) {
146 return build(get(InstOpc), DestReg);
148 inline MachineInstrBuilder build(const MCInstrDesc &MCID) {
149 return BuildMI(MBB, II, DL, MCID);
151 inline MachineInstrBuilder build(unsigned InstOpc) {
152 return build(get(InstOpc));
155 // Calculate an address of frame index from a frame register and a given
156 // offset if the offset doesn't fit in the immediate field. Use a clobber
157 // register to hold calculated address.
158 void prepareReplaceFI(MachineInstr &MI, Register &FrameReg, int64_t &Offset,
159 int64_t Bytes = 0);
160 // Replace the frame index in \p MI with a frame register and a given offset
161 // if it fits in the immediate field. Otherwise, use pre-calculated address
162 // in a clobber regsiter.
163 void replaceFI(MachineInstr &MI, Register FrameReg, int64_t Offset,
164 int FIOperandNum);
166 // Expand and eliminate Frame Index of pseudo STQrii and LDQrii.
167 void processSTQ(MachineInstr &MI, Register FrameReg, int64_t Offset,
168 int FIOperandNum);
169 void processLDQ(MachineInstr &MI, Register FrameReg, int64_t Offset,
170 int FIOperandNum);
171 // Expand and eliminate Frame Index of pseudo STVMrii and LDVMrii.
172 void processSTVM(MachineInstr &MI, Register FrameReg, int64_t Offset,
173 int FIOperandNum);
174 void processLDVM(MachineInstr &MI, Register FrameReg, int64_t Offset,
175 int FIOperandNum);
176 // Expand and eliminate Frame Index of pseudo STVM512rii and LDVM512rii.
177 void processSTVM512(MachineInstr &MI, Register FrameReg, int64_t Offset,
178 int FIOperandNum);
179 void processLDVM512(MachineInstr &MI, Register FrameReg, int64_t Offset,
180 int FIOperandNum);
182 public:
183 EliminateFrameIndex(const TargetInstrInfo &TII, const TargetRegisterInfo &TRI,
184 const DebugLoc &DL, MachineBasicBlock &MBB,
185 MachineBasicBlock::iterator II)
186 : TII(TII), TRI(TRI), DL(DL), MBB(MBB), II(II), clobber(VE::SX13) {}
188 // Expand and eliminate Frame Index from MI
189 void processMI(MachineInstr &MI, Register FrameReg, int64_t Offset,
190 int FIOperandNum);
192 } // namespace
194 // Prepare the frame index if it doesn't fit in the immediate field. Use
195 // clobber register to hold calculated address.
196 void EliminateFrameIndex::prepareReplaceFI(MachineInstr &MI, Register &FrameReg,
197 int64_t &Offset, int64_t Bytes) {
198 if (isInt<32>(Offset) && isInt<32>(Offset + Bytes)) {
199 // If the offset is small enough to fit in the immediate field, directly
200 // encode it. So, nothing to prepare here.
201 return;
204 // If the offset doesn't fit, emit following codes. This clobbers SX13
205 // which we always know is available here.
206 // lea %clobber, Offset@lo
207 // and %clobber, %clobber, (32)0
208 // lea.sl %clobber, Offset@hi(FrameReg, %clobber)
209 build(VE::LEAzii, clobber).addImm(0).addImm(0).addImm(Lo_32(Offset));
210 build(VE::ANDrm, clobber).addReg(clobber).addImm(M0(32));
211 build(VE::LEASLrri, clobber)
212 .addReg(clobber)
213 .addReg(FrameReg)
214 .addImm(Hi_32(Offset));
216 // Use clobber register as a frame register and 0 offset
217 FrameReg = clobber;
218 Offset = 0;
221 // Replace the frame index in \p MI with a proper byte and framereg offset.
222 void EliminateFrameIndex::replaceFI(MachineInstr &MI, Register FrameReg,
223 int64_t Offset, int FIOperandNum) {
224 assert(isInt<32>(Offset));
226 // The offset must be small enough to fit in the immediate field after
227 // call of prepareReplaceFI. Therefore, we directly encode it.
228 MI.getOperand(FIOperandNum).ChangeToRegister(FrameReg, false);
229 MI.getOperand(FIOperandNum + offsetToDisp(MI)).ChangeToImmediate(Offset);
232 void EliminateFrameIndex::processSTQ(MachineInstr &MI, Register FrameReg,
233 int64_t Offset, int FIOperandNum) {
234 assert(MI.getOpcode() == VE::STQrii);
235 LLVM_DEBUG(dbgs() << "processSTQ: "; MI.dump());
237 prepareReplaceFI(MI, FrameReg, Offset, 8);
239 Register SrcReg = MI.getOperand(3).getReg();
240 Register SrcHiReg = getSubReg(SrcReg, VE::sub_even);
241 Register SrcLoReg = getSubReg(SrcReg, VE::sub_odd);
242 // VE stores HiReg to 8(addr) and LoReg to 0(addr)
243 MachineInstr *StMI =
244 build(VE::STrii).addReg(FrameReg).addImm(0).addImm(0).addReg(SrcLoReg);
245 replaceFI(*StMI, FrameReg, Offset, 0);
246 // Mutate to 'hi' store.
247 MI.setDesc(get(VE::STrii));
248 MI.getOperand(3).setReg(SrcHiReg);
249 Offset += 8;
250 replaceFI(MI, FrameReg, Offset, FIOperandNum);
253 void EliminateFrameIndex::processLDQ(MachineInstr &MI, Register FrameReg,
254 int64_t Offset, int FIOperandNum) {
255 assert(MI.getOpcode() == VE::LDQrii);
256 LLVM_DEBUG(dbgs() << "processLDQ: "; MI.dump());
258 prepareReplaceFI(MI, FrameReg, Offset, 8);
260 Register DestReg = MI.getOperand(0).getReg();
261 Register DestHiReg = getSubReg(DestReg, VE::sub_even);
262 Register DestLoReg = getSubReg(DestReg, VE::sub_odd);
263 // VE loads HiReg from 8(addr) and LoReg from 0(addr)
264 MachineInstr *StMI =
265 build(VE::LDrii, DestLoReg).addReg(FrameReg).addImm(0).addImm(0);
266 replaceFI(*StMI, FrameReg, Offset, 1);
267 MI.setDesc(get(VE::LDrii));
268 MI.getOperand(0).setReg(DestHiReg);
269 Offset += 8;
270 replaceFI(MI, FrameReg, Offset, FIOperandNum);
273 void EliminateFrameIndex::processSTVM(MachineInstr &MI, Register FrameReg,
274 int64_t Offset, int FIOperandNum) {
275 assert(MI.getOpcode() == VE::STVMrii);
276 LLVM_DEBUG(dbgs() << "processSTVM: "; MI.dump());
278 // Original MI is:
279 // STVMrii frame-index, 0, offset, reg (, memory operand)
280 // Convert it to:
281 // SVMi tmp-reg, reg, 0
282 // STrii frame-reg, 0, offset, tmp-reg
283 // SVMi tmp-reg, reg, 1
284 // STrii frame-reg, 0, offset+8, tmp-reg
285 // SVMi tmp-reg, reg, 2
286 // STrii frame-reg, 0, offset+16, tmp-reg
287 // SVMi tmp-reg, reg, 3
288 // STrii frame-reg, 0, offset+24, tmp-reg
290 prepareReplaceFI(MI, FrameReg, Offset, 24);
292 Register SrcReg = MI.getOperand(3).getReg();
293 bool isKill = MI.getOperand(3).isKill();
294 // FIXME: it would be better to scavenge a register here instead of
295 // reserving SX16 all of the time.
296 Register TmpReg = VE::SX16;
297 for (int i = 0; i < 3; ++i) {
298 build(VE::SVMmr, TmpReg).addReg(SrcReg).addImm(i);
299 MachineInstr *StMI =
300 build(VE::STrii).addReg(FrameReg).addImm(0).addImm(0).addReg(
301 TmpReg, getKillRegState(true));
302 replaceFI(*StMI, FrameReg, Offset, 0);
303 Offset += 8;
305 build(VE::SVMmr, TmpReg).addReg(SrcReg, getKillRegState(isKill)).addImm(3);
306 MI.setDesc(get(VE::STrii));
307 MI.getOperand(3).ChangeToRegister(TmpReg, false, false, true);
308 replaceFI(MI, FrameReg, Offset, FIOperandNum);
311 void EliminateFrameIndex::processLDVM(MachineInstr &MI, Register FrameReg,
312 int64_t Offset, int FIOperandNum) {
313 assert(MI.getOpcode() == VE::LDVMrii);
314 LLVM_DEBUG(dbgs() << "processLDVM: "; MI.dump());
316 // Original MI is:
317 // LDVMri reg, frame-index, 0, offset (, memory operand)
318 // Convert it to:
319 // LDrii tmp-reg, frame-reg, 0, offset
320 // LVMir vm, 0, tmp-reg
321 // LDrii tmp-reg, frame-reg, 0, offset+8
322 // LVMir_m vm, 1, tmp-reg, vm
323 // LDrii tmp-reg, frame-reg, 0, offset+16
324 // LVMir_m vm, 2, tmp-reg, vm
325 // LDrii tmp-reg, frame-reg, 0, offset+24
326 // LVMir_m vm, 3, tmp-reg, vm
328 prepareReplaceFI(MI, FrameReg, Offset, 24);
330 Register DestReg = MI.getOperand(0).getReg();
331 // FIXME: it would be better to scavenge a register here instead of
332 // reserving SX16 all of the time.
333 unsigned TmpReg = VE::SX16;
334 for (int i = 0; i < 4; ++i) {
335 if (i != 3) {
336 MachineInstr *StMI =
337 build(VE::LDrii, TmpReg).addReg(FrameReg).addImm(0).addImm(0);
338 replaceFI(*StMI, FrameReg, Offset, 1);
339 Offset += 8;
340 } else {
341 // Last LDrii replace the target instruction.
342 MI.setDesc(get(VE::LDrii));
343 MI.getOperand(0).ChangeToRegister(TmpReg, true);
345 // First LVM is LVMir. Others are LVMir_m. Last LVM places at the
346 // next of the target instruction.
347 if (i == 0)
348 build(VE::LVMir, DestReg).addImm(i).addReg(TmpReg, getKillRegState(true));
349 else if (i != 3)
350 build(VE::LVMir_m, DestReg)
351 .addImm(i)
352 .addReg(TmpReg, getKillRegState(true))
353 .addReg(DestReg);
354 else
355 BuildMI(*MI.getParent(), std::next(II), DL, get(VE::LVMir_m), DestReg)
356 .addImm(3)
357 .addReg(TmpReg, getKillRegState(true))
358 .addReg(DestReg);
360 replaceFI(MI, FrameReg, Offset, FIOperandNum);
363 void EliminateFrameIndex::processSTVM512(MachineInstr &MI, Register FrameReg,
364 int64_t Offset, int FIOperandNum) {
365 assert(MI.getOpcode() == VE::STVM512rii);
366 LLVM_DEBUG(dbgs() << "processSTVM512: "; MI.dump());
368 prepareReplaceFI(MI, FrameReg, Offset, 56);
370 Register SrcReg = MI.getOperand(3).getReg();
371 Register SrcLoReg = getSubReg(SrcReg, VE::sub_vm_odd);
372 Register SrcHiReg = getSubReg(SrcReg, VE::sub_vm_even);
373 bool isKill = MI.getOperand(3).isKill();
374 // FIXME: it would be better to scavenge a register here instead of
375 // reserving SX16 all of the time.
376 Register TmpReg = VE::SX16;
377 // store low part of VMP
378 MachineInstr *LastMI = nullptr;
379 for (int i = 0; i < 4; ++i) {
380 LastMI = build(VE::SVMmr, TmpReg).addReg(SrcLoReg).addImm(i);
381 MachineInstr *StMI =
382 build(VE::STrii).addReg(FrameReg).addImm(0).addImm(0).addReg(
383 TmpReg, getKillRegState(true));
384 replaceFI(*StMI, FrameReg, Offset, 0);
385 Offset += 8;
387 if (isKill)
388 LastMI->addRegisterKilled(SrcLoReg, &TRI, true);
389 // store high part of VMP
390 for (int i = 0; i < 3; ++i) {
391 build(VE::SVMmr, TmpReg).addReg(SrcHiReg).addImm(i);
392 MachineInstr *StMI =
393 build(VE::STrii).addReg(FrameReg).addImm(0).addImm(0).addReg(
394 TmpReg, getKillRegState(true));
395 replaceFI(*StMI, FrameReg, Offset, 0);
396 Offset += 8;
398 LastMI = build(VE::SVMmr, TmpReg).addReg(SrcHiReg).addImm(3);
399 if (isKill) {
400 LastMI->addRegisterKilled(SrcHiReg, &TRI, true);
401 // Add implicit super-register kills to the particular MI.
402 LastMI->addRegisterKilled(SrcReg, &TRI, true);
404 MI.setDesc(get(VE::STrii));
405 MI.getOperand(3).ChangeToRegister(TmpReg, false, false, true);
406 replaceFI(MI, FrameReg, Offset, FIOperandNum);
409 void EliminateFrameIndex::processLDVM512(MachineInstr &MI, Register FrameReg,
410 int64_t Offset, int FIOperandNum) {
411 assert(MI.getOpcode() == VE::LDVM512rii);
412 LLVM_DEBUG(dbgs() << "processLDVM512: "; MI.dump());
414 prepareReplaceFI(MI, FrameReg, Offset, 56);
416 Register DestReg = MI.getOperand(0).getReg();
417 Register DestLoReg = getSubReg(DestReg, VE::sub_vm_odd);
418 Register DestHiReg = getSubReg(DestReg, VE::sub_vm_even);
419 // FIXME: it would be better to scavenge a register here instead of
420 // reserving SX16 all of the time.
421 Register TmpReg = VE::SX16;
422 build(VE::IMPLICIT_DEF, DestReg);
423 for (int i = 0; i < 4; ++i) {
424 MachineInstr *LdMI =
425 build(VE::LDrii, TmpReg).addReg(FrameReg).addImm(0).addImm(0);
426 replaceFI(*LdMI, FrameReg, Offset, 1);
427 build(VE::LVMir_m, DestLoReg)
428 .addImm(i)
429 .addReg(TmpReg, getKillRegState(true))
430 .addReg(DestLoReg);
431 Offset += 8;
433 for (int i = 0; i < 3; ++i) {
434 MachineInstr *LdMI =
435 build(VE::LDrii, TmpReg).addReg(FrameReg).addImm(0).addImm(0);
436 replaceFI(*LdMI, FrameReg, Offset, 1);
437 build(VE::LVMir_m, DestHiReg)
438 .addImm(i)
439 .addReg(TmpReg, getKillRegState(true))
440 .addReg(DestHiReg);
441 Offset += 8;
443 MI.setDesc(get(VE::LDrii));
444 MI.getOperand(0).ChangeToRegister(TmpReg, true);
445 BuildMI(*MI.getParent(), std::next(II), DL, get(VE::LVMir_m), DestHiReg)
446 .addImm(3)
447 .addReg(TmpReg, getKillRegState(true))
448 .addReg(DestHiReg);
449 replaceFI(MI, FrameReg, Offset, FIOperandNum);
452 void EliminateFrameIndex::processMI(MachineInstr &MI, Register FrameReg,
453 int64_t Offset, int FIOperandNum) {
454 switch (MI.getOpcode()) {
455 case VE::STQrii:
456 processSTQ(MI, FrameReg, Offset, FIOperandNum);
457 return;
458 case VE::LDQrii:
459 processLDQ(MI, FrameReg, Offset, FIOperandNum);
460 return;
461 case VE::STVMrii:
462 processSTVM(MI, FrameReg, Offset, FIOperandNum);
463 return;
464 case VE::LDVMrii:
465 processLDVM(MI, FrameReg, Offset, FIOperandNum);
466 return;
467 case VE::STVM512rii:
468 processSTVM512(MI, FrameReg, Offset, FIOperandNum);
469 return;
470 case VE::LDVM512rii:
471 processLDVM512(MI, FrameReg, Offset, FIOperandNum);
472 return;
474 prepareReplaceFI(MI, FrameReg, Offset);
475 replaceFI(MI, FrameReg, Offset, FIOperandNum);
478 bool VERegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
479 int SPAdj, unsigned FIOperandNum,
480 RegScavenger *RS) const {
481 assert(SPAdj == 0 && "Unexpected");
483 MachineInstr &MI = *II;
484 int FrameIndex = MI.getOperand(FIOperandNum).getIndex();
486 MachineFunction &MF = *MI.getParent()->getParent();
487 const VESubtarget &Subtarget = MF.getSubtarget<VESubtarget>();
488 const VEFrameLowering &TFI = *getFrameLowering(MF);
489 const TargetInstrInfo &TII = *Subtarget.getInstrInfo();
490 const VERegisterInfo &TRI = *Subtarget.getRegisterInfo();
491 DebugLoc DL = MI.getDebugLoc();
492 EliminateFrameIndex EFI(TII, TRI, DL, *MI.getParent(), II);
494 // Retrieve FrameReg and byte offset for stack slot.
495 Register FrameReg;
496 int64_t Offset =
497 TFI.getFrameIndexReference(MF, FrameIndex, FrameReg).getFixed();
498 Offset += MI.getOperand(FIOperandNum + offsetToDisp(MI)).getImm();
500 EFI.processMI(MI, FrameReg, Offset, FIOperandNum);
501 return false;
504 Register VERegisterInfo::getFrameRegister(const MachineFunction &MF) const {
505 return VE::SX9;