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;
36 bool ExpandInputPerspective(MachineInstr& MI);
37 bool ExpandInputConstant(MachineInstr& MI);
40 R600ExpandSpecialInstrsPass(TargetMachine &tm) : MachineFunctionPass(ID),
43 virtual bool runOnMachineFunction(MachineFunction &MF);
45 const char *getPassName() const {
46 return "R600 Expand special instructions pass";
50 } // End anonymous namespace
52 char R600ExpandSpecialInstrsPass::ID = 0;
54 FunctionPass *llvm::createR600ExpandSpecialInstrsPass(TargetMachine &TM) {
55 return new R600ExpandSpecialInstrsPass(TM);
58 bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
59 TII = static_cast<const R600InstrInfo *>(MF.getTarget().getInstrInfo());
61 const R600RegisterInfo &TRI = TII->getRegisterInfo();
63 for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
65 MachineBasicBlock &MBB = *BB;
66 MachineBasicBlock::iterator I = MBB.begin();
67 while (I != MBB.end()) {
68 MachineInstr &MI = *I;
71 switch (MI.getOpcode()) {
73 // Expand PRED_X to one of the PRED_SET instructions.
74 case AMDGPU::PRED_X: {
75 uint64_t Flags = MI.getOperand(3).getImm();
76 // The native opcode used by PRED_X is stored as an immediate in the
78 MachineInstr *PredSet = TII->buildDefaultInstruction(MBB, I,
79 MI.getOperand(2).getImm(), // opcode
80 MI.getOperand(0).getReg(), // dst
81 MI.getOperand(1).getReg(), // src0
82 AMDGPU::ZERO); // src1
83 TII->addFlag(PredSet, 0, MO_FLAG_MASK);
84 if (Flags & MO_FLAG_PUSH) {
85 TII->setImmOperand(PredSet, AMDGPU::OpName::update_exec_mask, 1);
87 TII->setImmOperand(PredSet, AMDGPU::OpName::update_pred, 1);
93 MachineInstr *PredSet = TII->buildDefaultInstruction(MBB, I,
94 AMDGPU::PRED_SETE_INT,
95 AMDGPU::PREDICATE_BIT,
98 TII->addFlag(PredSet, 0, MO_FLAG_MASK);
99 TII->setImmOperand(PredSet, AMDGPU::OpName::update_exec_mask, 1);
101 BuildMI(MBB, I, MBB.findDebugLoc(I),
102 TII->get(AMDGPU::PREDICATED_BREAK))
103 .addReg(AMDGPU::PREDICATE_BIT);
104 MI.eraseFromParent();
108 case AMDGPU::INTERP_PAIR_XY: {
110 unsigned PReg = AMDGPU::R600_ArrayBaseRegClass.getRegister(
111 MI.getOperand(2).getImm());
113 for (unsigned Chan = 0; Chan < 4; ++Chan) {
117 DstReg = MI.getOperand(Chan).getReg();
119 DstReg = Chan == 2 ? AMDGPU::T0_Z : AMDGPU::T0_W;
121 BMI = TII->buildDefaultInstruction(MBB, I, AMDGPU::INTERP_XY,
122 DstReg, MI.getOperand(3 + (Chan % 2)).getReg(), PReg);
125 BMI->bundleWithPred();
128 TII->addFlag(BMI, 0, MO_FLAG_MASK);
130 TII->addFlag(BMI, 0, MO_FLAG_NOT_LAST);
133 MI.eraseFromParent();
137 case AMDGPU::INTERP_PAIR_ZW: {
139 unsigned PReg = AMDGPU::R600_ArrayBaseRegClass.getRegister(
140 MI.getOperand(2).getImm());
142 for (unsigned Chan = 0; Chan < 4; ++Chan) {
146 DstReg = Chan == 0 ? AMDGPU::T0_X : AMDGPU::T0_Y;
148 DstReg = MI.getOperand(Chan-2).getReg();
150 BMI = TII->buildDefaultInstruction(MBB, I, AMDGPU::INTERP_ZW,
151 DstReg, MI.getOperand(3 + (Chan % 2)).getReg(), PReg);
154 BMI->bundleWithPred();
157 TII->addFlag(BMI, 0, MO_FLAG_MASK);
159 TII->addFlag(BMI, 0, MO_FLAG_NOT_LAST);
162 MI.eraseFromParent();
166 case AMDGPU::INTERP_VEC_LOAD: {
167 const R600RegisterInfo &TRI = TII->getRegisterInfo();
169 unsigned PReg = AMDGPU::R600_ArrayBaseRegClass.getRegister(
170 MI.getOperand(1).getImm());
171 unsigned DstReg = MI.getOperand(0).getReg();
173 for (unsigned Chan = 0; Chan < 4; ++Chan) {
174 BMI = TII->buildDefaultInstruction(MBB, I, AMDGPU::INTERP_LOAD_P0,
175 TRI.getSubReg(DstReg, TRI.getSubRegFromChannel(Chan)), PReg);
177 BMI->bundleWithPred();
180 TII->addFlag(BMI, 0, MO_FLAG_NOT_LAST);
183 MI.eraseFromParent();
186 case AMDGPU::DOT_4: {
188 const R600RegisterInfo &TRI = TII->getRegisterInfo();
190 unsigned DstReg = MI.getOperand(0).getReg();
191 unsigned DstBase = TRI.getEncodingValue(DstReg) & HW_REG_MASK;
193 for (unsigned Chan = 0; Chan < 4; ++Chan) {
194 bool Mask = (Chan != TRI.getHWRegChan(DstReg));
196 AMDGPU::R600_TReg32RegClass.getRegister((DstBase * 4) + Chan);
198 TII->buildSlotOfVectorInstruction(MBB, &MI, Chan, SubDstReg);
200 BMI->bundleWithPred();
203 TII->addFlag(BMI, 0, MO_FLAG_MASK);
206 TII->addFlag(BMI, 0, MO_FLAG_NOT_LAST);
207 unsigned Opcode = BMI->getOpcode();
208 // While not strictly necessary from hw point of view, we force
209 // all src operands of a dot4 inst to belong to the same slot.
210 unsigned Src0 = BMI->getOperand(
211 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0))
213 unsigned Src1 = BMI->getOperand(
214 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1))
218 if ((TRI.getEncodingValue(Src0) & 0xff) < 127 &&
219 (TRI.getEncodingValue(Src1) & 0xff) < 127)
220 assert(TRI.getHWRegChan(Src0) == TRI.getHWRegChan(Src1));
222 MI.eraseFromParent();
227 bool IsReduction = TII->isReductionOp(MI.getOpcode());
228 bool IsVector = TII->isVector(MI);
229 bool IsCube = TII->isCubeOp(MI.getOpcode());
230 if (!IsReduction && !IsVector && !IsCube) {
234 // Expand the instruction
236 // Reduction instructions:
237 // T0_X = DP4 T1_XYZW, T2_XYZW
239 // TO_X = DP4 T1_X, T2_X
240 // TO_Y (write masked) = DP4 T1_Y, T2_Y
241 // TO_Z (write masked) = DP4 T1_Z, T2_Z
242 // TO_W (write masked) = DP4 T1_W, T2_W
244 // Vector instructions:
245 // T0_X = MULLO_INT T1_X, T2_X
247 // T0_X = MULLO_INT T1_X, T2_X
248 // T0_Y (write masked) = MULLO_INT T1_X, T2_X
249 // T0_Z (write masked) = MULLO_INT T1_X, T2_X
250 // T0_W (write masked) = MULLO_INT T1_X, T2_X
252 // Cube instructions:
253 // T0_XYZW = CUBE T1_XYZW
255 // TO_X = CUBE T1_Z, T1_Y
256 // T0_Y = CUBE T1_Z, T1_X
257 // T0_Z = CUBE T1_X, T1_Z
258 // T0_W = CUBE T1_Y, T1_Z
259 for (unsigned Chan = 0; Chan < 4; Chan++) {
260 unsigned DstReg = MI.getOperand(
261 TII->getOperandIdx(MI, AMDGPU::OpName::dst)).getReg();
262 unsigned Src0 = MI.getOperand(
263 TII->getOperandIdx(MI, AMDGPU::OpName::src0)).getReg();
266 // Determine the correct source registers
268 int Src1Idx = TII->getOperandIdx(MI, AMDGPU::OpName::src1);
270 Src1 = MI.getOperand(Src1Idx).getReg();
274 unsigned SubRegIndex = TRI.getSubRegFromChannel(Chan);
275 Src0 = TRI.getSubReg(Src0, SubRegIndex);
276 Src1 = TRI.getSubReg(Src1, SubRegIndex);
278 static const int CubeSrcSwz[] = {2, 2, 0, 1};
279 unsigned SubRegIndex0 = TRI.getSubRegFromChannel(CubeSrcSwz[Chan]);
280 unsigned SubRegIndex1 = TRI.getSubRegFromChannel(CubeSrcSwz[3 - Chan]);
281 Src1 = TRI.getSubReg(Src0, SubRegIndex1);
282 Src0 = TRI.getSubReg(Src0, SubRegIndex0);
285 // Determine the correct destination registers;
289 unsigned SubRegIndex = TRI.getSubRegFromChannel(Chan);
290 DstReg = TRI.getSubReg(DstReg, SubRegIndex);
292 // Mask the write if the original instruction does not write to
293 // the current Channel.
294 Mask = (Chan != TRI.getHWRegChan(DstReg));
295 unsigned DstBase = TRI.getEncodingValue(DstReg) & HW_REG_MASK;
296 DstReg = AMDGPU::R600_TReg32RegClass.getRegister((DstBase * 4) + Chan);
299 // Set the IsLast bit
300 NotLast = (Chan != 3 );
302 // Add the new instruction
303 unsigned Opcode = MI.getOpcode();
305 case AMDGPU::CUBE_r600_pseudo:
306 Opcode = AMDGPU::CUBE_r600_real;
308 case AMDGPU::CUBE_eg_pseudo:
309 Opcode = AMDGPU::CUBE_eg_real;
315 MachineInstr *NewMI =
316 TII->buildDefaultInstruction(MBB, I, Opcode, DstReg, Src0, Src1);
319 NewMI->bundleWithPred();
321 TII->addFlag(NewMI, 0, MO_FLAG_MASK);
324 TII->addFlag(NewMI, 0, MO_FLAG_NOT_LAST);
327 MI.eraseFromParent();