1 //===-- R600ExpandSpecialInstrs.cpp - Expand special instructions ---------===//
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 //===----------------------------------------------------------------------===//
11 /// Vector, Reduction, and Cube instructions need to fill the entire instruction
12 /// group to work correctly. This pass expands these individual instructions
13 /// into several instructions that will completely fill the instruction group.
15 //===----------------------------------------------------------------------===//
18 #include "R600Defines.h"
19 #include "R600InstrInfo.h"
20 #include "R600MachineFunctionInfo.h"
21 #include "R600RegisterInfo.h"
22 #include "llvm/CodeGen/MachineFunctionPass.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 class R600ExpandSpecialInstrsPass : public MachineFunctionPass {
34 const R600InstrInfo *TII;
37 R600ExpandSpecialInstrsPass(TargetMachine &tm) : MachineFunctionPass(ID),
40 virtual bool runOnMachineFunction(MachineFunction &MF);
42 const char *getPassName() const {
43 return "R600 Expand special instructions pass";
47 } // End anonymous namespace
49 char R600ExpandSpecialInstrsPass::ID = 0;
51 FunctionPass *llvm::createR600ExpandSpecialInstrsPass(TargetMachine &TM) {
52 return new R600ExpandSpecialInstrsPass(TM);
55 bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
56 TII = static_cast<const R600InstrInfo *>(MF.getTarget().getInstrInfo());
58 const R600RegisterInfo &TRI = TII->getRegisterInfo();
60 for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
62 MachineBasicBlock &MBB = *BB;
63 MachineBasicBlock::iterator I = MBB.begin();
64 while (I != MBB.end()) {
65 MachineInstr &MI = *I;
68 // Expand LDS_*_RET instructions
69 if (TII->isLDSRetInstr(MI.getOpcode())) {
70 int DstIdx = TII->getOperandIdx(MI.getOpcode(), AMDGPU::OpName::dst);
72 MachineOperand &DstOp = MI.getOperand(DstIdx);
73 MachineInstr *Mov = TII->buildMovInstr(&MBB, I,
74 DstOp.getReg(), AMDGPU::OQAP);
75 DstOp.setReg(AMDGPU::OQAP);
76 int LDSPredSelIdx = TII->getOperandIdx(MI.getOpcode(),
77 AMDGPU::OpName::pred_sel);
78 int MovPredSelIdx = TII->getOperandIdx(Mov->getOpcode(),
79 AMDGPU::OpName::pred_sel);
80 // Copy the pred_sel bit
81 Mov->getOperand(MovPredSelIdx).setReg(
82 MI.getOperand(LDSPredSelIdx).getReg());
85 switch (MI.getOpcode()) {
87 // Expand PRED_X to one of the PRED_SET instructions.
88 case AMDGPU::PRED_X: {
89 uint64_t Flags = MI.getOperand(3).getImm();
90 // The native opcode used by PRED_X is stored as an immediate in the
92 MachineInstr *PredSet = TII->buildDefaultInstruction(MBB, I,
93 MI.getOperand(2).getImm(), // opcode
94 MI.getOperand(0).getReg(), // dst
95 MI.getOperand(1).getReg(), // src0
96 AMDGPU::ZERO); // src1
97 TII->addFlag(PredSet, 0, MO_FLAG_MASK);
98 if (Flags & MO_FLAG_PUSH) {
99 TII->setImmOperand(PredSet, AMDGPU::OpName::update_exec_mask, 1);
101 TII->setImmOperand(PredSet, AMDGPU::OpName::update_pred, 1);
103 MI.eraseFromParent();
107 case AMDGPU::INTERP_PAIR_XY: {
109 unsigned PReg = AMDGPU::R600_ArrayBaseRegClass.getRegister(
110 MI.getOperand(2).getImm());
112 for (unsigned Chan = 0; Chan < 4; ++Chan) {
116 DstReg = MI.getOperand(Chan).getReg();
118 DstReg = Chan == 2 ? AMDGPU::T0_Z : AMDGPU::T0_W;
120 BMI = TII->buildDefaultInstruction(MBB, I, AMDGPU::INTERP_XY,
121 DstReg, MI.getOperand(3 + (Chan % 2)).getReg(), PReg);
124 BMI->bundleWithPred();
127 TII->addFlag(BMI, 0, MO_FLAG_MASK);
129 TII->addFlag(BMI, 0, MO_FLAG_NOT_LAST);
132 MI.eraseFromParent();
136 case AMDGPU::INTERP_PAIR_ZW: {
138 unsigned PReg = AMDGPU::R600_ArrayBaseRegClass.getRegister(
139 MI.getOperand(2).getImm());
141 for (unsigned Chan = 0; Chan < 4; ++Chan) {
145 DstReg = Chan == 0 ? AMDGPU::T0_X : AMDGPU::T0_Y;
147 DstReg = MI.getOperand(Chan-2).getReg();
149 BMI = TII->buildDefaultInstruction(MBB, I, AMDGPU::INTERP_ZW,
150 DstReg, MI.getOperand(3 + (Chan % 2)).getReg(), PReg);
153 BMI->bundleWithPred();
156 TII->addFlag(BMI, 0, MO_FLAG_MASK);
158 TII->addFlag(BMI, 0, MO_FLAG_NOT_LAST);
161 MI.eraseFromParent();
165 case AMDGPU::INTERP_VEC_LOAD: {
166 const R600RegisterInfo &TRI = TII->getRegisterInfo();
168 unsigned PReg = AMDGPU::R600_ArrayBaseRegClass.getRegister(
169 MI.getOperand(1).getImm());
170 unsigned DstReg = MI.getOperand(0).getReg();
172 for (unsigned Chan = 0; Chan < 4; ++Chan) {
173 BMI = TII->buildDefaultInstruction(MBB, I, AMDGPU::INTERP_LOAD_P0,
174 TRI.getSubReg(DstReg, TRI.getSubRegFromChannel(Chan)), PReg);
176 BMI->bundleWithPred();
179 TII->addFlag(BMI, 0, MO_FLAG_NOT_LAST);
182 MI.eraseFromParent();
185 case AMDGPU::DOT_4: {
187 const R600RegisterInfo &TRI = TII->getRegisterInfo();
189 unsigned DstReg = MI.getOperand(0).getReg();
190 unsigned DstBase = TRI.getEncodingValue(DstReg) & HW_REG_MASK;
192 for (unsigned Chan = 0; Chan < 4; ++Chan) {
193 bool Mask = (Chan != TRI.getHWRegChan(DstReg));
195 AMDGPU::R600_TReg32RegClass.getRegister((DstBase * 4) + Chan);
197 TII->buildSlotOfVectorInstruction(MBB, &MI, Chan, SubDstReg);
199 BMI->bundleWithPred();
202 TII->addFlag(BMI, 0, MO_FLAG_MASK);
205 TII->addFlag(BMI, 0, MO_FLAG_NOT_LAST);
206 unsigned Opcode = BMI->getOpcode();
207 // While not strictly necessary from hw point of view, we force
208 // all src operands of a dot4 inst to belong to the same slot.
209 unsigned Src0 = BMI->getOperand(
210 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0))
212 unsigned Src1 = BMI->getOperand(
213 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1))
217 if ((TRI.getEncodingValue(Src0) & 0xff) < 127 &&
218 (TRI.getEncodingValue(Src1) & 0xff) < 127)
219 assert(TRI.getHWRegChan(Src0) == TRI.getHWRegChan(Src1));
221 MI.eraseFromParent();
226 bool IsReduction = TII->isReductionOp(MI.getOpcode());
227 bool IsVector = TII->isVector(MI);
228 bool IsCube = TII->isCubeOp(MI.getOpcode());
229 if (!IsReduction && !IsVector && !IsCube) {
233 // Expand the instruction
235 // Reduction instructions:
236 // T0_X = DP4 T1_XYZW, T2_XYZW
238 // TO_X = DP4 T1_X, T2_X
239 // TO_Y (write masked) = DP4 T1_Y, T2_Y
240 // TO_Z (write masked) = DP4 T1_Z, T2_Z
241 // TO_W (write masked) = DP4 T1_W, T2_W
243 // Vector instructions:
244 // T0_X = MULLO_INT T1_X, T2_X
246 // T0_X = MULLO_INT T1_X, T2_X
247 // T0_Y (write masked) = MULLO_INT T1_X, T2_X
248 // T0_Z (write masked) = MULLO_INT T1_X, T2_X
249 // T0_W (write masked) = MULLO_INT T1_X, T2_X
251 // Cube instructions:
252 // T0_XYZW = CUBE T1_XYZW
254 // TO_X = CUBE T1_Z, T1_Y
255 // T0_Y = CUBE T1_Z, T1_X
256 // T0_Z = CUBE T1_X, T1_Z
257 // T0_W = CUBE T1_Y, T1_Z
258 for (unsigned Chan = 0; Chan < 4; Chan++) {
259 unsigned DstReg = MI.getOperand(
260 TII->getOperandIdx(MI, AMDGPU::OpName::dst)).getReg();
261 unsigned Src0 = MI.getOperand(
262 TII->getOperandIdx(MI, AMDGPU::OpName::src0)).getReg();
265 // Determine the correct source registers
267 int Src1Idx = TII->getOperandIdx(MI, AMDGPU::OpName::src1);
269 Src1 = MI.getOperand(Src1Idx).getReg();
273 unsigned SubRegIndex = TRI.getSubRegFromChannel(Chan);
274 Src0 = TRI.getSubReg(Src0, SubRegIndex);
275 Src1 = TRI.getSubReg(Src1, SubRegIndex);
277 static const int CubeSrcSwz[] = {2, 2, 0, 1};
278 unsigned SubRegIndex0 = TRI.getSubRegFromChannel(CubeSrcSwz[Chan]);
279 unsigned SubRegIndex1 = TRI.getSubRegFromChannel(CubeSrcSwz[3 - Chan]);
280 Src1 = TRI.getSubReg(Src0, SubRegIndex1);
281 Src0 = TRI.getSubReg(Src0, SubRegIndex0);
284 // Determine the correct destination registers;
288 unsigned SubRegIndex = TRI.getSubRegFromChannel(Chan);
289 DstReg = TRI.getSubReg(DstReg, SubRegIndex);
291 // Mask the write if the original instruction does not write to
292 // the current Channel.
293 Mask = (Chan != TRI.getHWRegChan(DstReg));
294 unsigned DstBase = TRI.getEncodingValue(DstReg) & HW_REG_MASK;
295 DstReg = AMDGPU::R600_TReg32RegClass.getRegister((DstBase * 4) + Chan);
298 // Set the IsLast bit
299 NotLast = (Chan != 3 );
301 // Add the new instruction
302 unsigned Opcode = MI.getOpcode();
304 case AMDGPU::CUBE_r600_pseudo:
305 Opcode = AMDGPU::CUBE_r600_real;
307 case AMDGPU::CUBE_eg_pseudo:
308 Opcode = AMDGPU::CUBE_eg_real;
314 MachineInstr *NewMI =
315 TII->buildDefaultInstruction(MBB, I, Opcode, DstReg, Src0, Src1);
318 NewMI->bundleWithPred();
320 TII->addFlag(NewMI, 0, MO_FLAG_MASK);
323 TII->addFlag(NewMI, 0, MO_FLAG_NOT_LAST);
326 MI.eraseFromParent();