1 //===-- SILowerControlFlow.cpp - Use predicates for control flow ----------===//
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 /// \brief This pass lowers the pseudo control flow instructions to real
12 /// machine instructions.
14 /// All control flow is handled using predicated instructions and
15 /// a predicate stack. Each Scalar ALU controls the operations of 64 Vector
16 /// ALUs. The Scalar ALU can update the predicate for any of the Vector ALUs
17 /// by writting to the 64-bit EXEC register (each bit corresponds to a
18 /// single vector ALU). Typically, for predicates, a vector ALU will write
19 /// to its bit of the VCC register (like EXEC VCC is 64-bits, one for each
20 /// Vector ALU) and then the ScalarALU will AND the VCC register with the
21 /// EXEC to update the predicates.
24 /// %VCC = V_CMP_GT_F32 %VGPR1, %VGPR2
25 /// %SGPR0 = SI_IF %VCC
26 /// %VGPR0 = V_ADD_F32 %VGPR0, %VGPR0
27 /// %SGPR0 = SI_ELSE %SGPR0
28 /// %VGPR0 = V_SUB_F32 %VGPR0, %VGPR0
33 /// %SGPR0 = S_AND_SAVEEXEC_B64 %VCC // Save and update the exec mask
34 /// %SGPR0 = S_XOR_B64 %SGPR0, %EXEC // Clear live bits from saved exec mask
35 /// S_CBRANCH_EXECZ label0 // This instruction is an optional
36 /// // optimization which allows us to
37 /// // branch if all the bits of
39 /// %VGPR0 = V_ADD_F32 %VGPR0, %VGPR0 // Do the IF block of the branch
42 /// %SGPR0 = S_OR_SAVEEXEC_B64 %EXEC // Restore the exec mask for the Then block
43 /// %EXEC = S_XOR_B64 %SGPR0, %EXEC // Clear live bits from saved exec mask
44 /// S_BRANCH_EXECZ label1 // Use our branch optimization
45 /// // instruction again.
46 /// %VGPR0 = V_SUB_F32 %VGPR0, %VGPR // Do the THEN block
48 /// %EXEC = S_OR_B64 %EXEC, %SGPR0 // Re-enable saved exec mask bits
49 //===----------------------------------------------------------------------===//
52 #include "SIInstrInfo.h"
53 #include "SIMachineFunctionInfo.h"
54 #include "llvm/CodeGen/MachineFunction.h"
55 #include "llvm/CodeGen/MachineFunctionPass.h"
56 #include "llvm/CodeGen/MachineInstrBuilder.h"
57 #include "llvm/CodeGen/MachineRegisterInfo.h"
63 class SILowerControlFlowPass : public MachineFunctionPass {
67 const TargetInstrInfo *TII;
69 void If(MachineInstr &MI);
70 void Else(MachineInstr &MI);
71 void Break(MachineInstr &MI);
72 void IfBreak(MachineInstr &MI);
73 void ElseBreak(MachineInstr &MI);
74 void Loop(MachineInstr &MI);
75 void EndCf(MachineInstr &MI);
78 SILowerControlFlowPass(TargetMachine &tm) :
79 MachineFunctionPass(ID), TII(tm.getInstrInfo()) { }
81 virtual bool runOnMachineFunction(MachineFunction &MF);
83 const char *getPassName() const {
84 return "SI Lower control flow instructions";
89 } // End anonymous namespace
91 char SILowerControlFlowPass::ID = 0;
93 FunctionPass *llvm::createSILowerControlFlowPass(TargetMachine &tm) {
94 return new SILowerControlFlowPass(tm);
97 void SILowerControlFlowPass::If(MachineInstr &MI) {
99 MachineBasicBlock &MBB = *MI.getParent();
100 DebugLoc DL = MI.getDebugLoc();
101 unsigned Reg = MI.getOperand(0).getReg();
102 unsigned Vcc = MI.getOperand(1).getReg();
104 BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_AND_SAVEEXEC_B64), Reg)
107 BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_XOR_B64), Reg)
108 .addReg(AMDGPU::EXEC)
111 MI.eraseFromParent();
114 void SILowerControlFlowPass::Else(MachineInstr &MI) {
116 MachineBasicBlock &MBB = *MI.getParent();
117 DebugLoc DL = MI.getDebugLoc();
118 unsigned Dst = MI.getOperand(0).getReg();
119 unsigned Src = MI.getOperand(1).getReg();
121 BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_OR_SAVEEXEC_B64), Dst)
122 .addReg(Src); // Saved EXEC
124 BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_XOR_B64), AMDGPU::EXEC)
125 .addReg(AMDGPU::EXEC)
128 MI.eraseFromParent();
131 void SILowerControlFlowPass::Break(MachineInstr &MI) {
133 MachineBasicBlock &MBB = *MI.getParent();
134 DebugLoc DL = MI.getDebugLoc();
136 unsigned Dst = MI.getOperand(0).getReg();
137 unsigned Src = MI.getOperand(1).getReg();
139 BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_OR_B64), Dst)
140 .addReg(AMDGPU::EXEC)
143 MI.eraseFromParent();
146 void SILowerControlFlowPass::IfBreak(MachineInstr &MI) {
148 MachineBasicBlock &MBB = *MI.getParent();
149 DebugLoc DL = MI.getDebugLoc();
151 unsigned Dst = MI.getOperand(0).getReg();
152 unsigned Vcc = MI.getOperand(1).getReg();
153 unsigned Src = MI.getOperand(2).getReg();
155 BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_OR_B64), Dst)
159 MI.eraseFromParent();
162 void SILowerControlFlowPass::ElseBreak(MachineInstr &MI) {
164 MachineBasicBlock &MBB = *MI.getParent();
165 DebugLoc DL = MI.getDebugLoc();
167 unsigned Dst = MI.getOperand(0).getReg();
168 unsigned Saved = MI.getOperand(1).getReg();
169 unsigned Src = MI.getOperand(2).getReg();
171 BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_OR_B64), Dst)
175 MI.eraseFromParent();
178 void SILowerControlFlowPass::Loop(MachineInstr &MI) {
180 MachineBasicBlock &MBB = *MI.getParent();
181 DebugLoc DL = MI.getDebugLoc();
182 unsigned Src = MI.getOperand(0).getReg();
184 BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_ANDN2_B64), AMDGPU::EXEC)
185 .addReg(AMDGPU::EXEC)
188 BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
189 .addOperand(MI.getOperand(1))
190 .addReg(AMDGPU::EXEC);
192 MI.eraseFromParent();
195 void SILowerControlFlowPass::EndCf(MachineInstr &MI) {
197 MachineBasicBlock &MBB = *MI.getParent();
198 DebugLoc DL = MI.getDebugLoc();
199 unsigned Reg = MI.getOperand(0).getReg();
201 BuildMI(MBB, MBB.getFirstNonPHI(), DL,
202 TII->get(AMDGPU::S_OR_B64), AMDGPU::EXEC)
203 .addReg(AMDGPU::EXEC)
206 MI.eraseFromParent();
209 bool SILowerControlFlowPass::runOnMachineFunction(MachineFunction &MF) {
213 for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
216 MachineBasicBlock &MBB = *BI;
217 for (MachineBasicBlock::iterator I = MBB.begin(), Next = llvm::next(I);
218 I != MBB.end(); I = Next) {
220 Next = llvm::next(I);
221 MachineInstr &MI = *I;
222 switch (MI.getOpcode()) {
228 case AMDGPU::SI_ELSE:
232 case AMDGPU::SI_BREAK:
236 case AMDGPU::SI_IF_BREAK:
240 case AMDGPU::SI_ELSE_BREAK:
244 case AMDGPU::SI_LOOP:
248 case AMDGPU::SI_END_CF:
256 // TODO: What is this good for?
257 unsigned ShaderType = MF.getInfo<SIMachineFunctionInfo>()->ShaderType;
258 if (HaveCf && ShaderType == ShaderType::PIXEL) {
259 for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
262 MachineBasicBlock &MBB = *BI;
263 if (MBB.succ_empty()) {
265 MachineInstr &MI = *MBB.getFirstNonPHI();
266 DebugLoc DL = MI.getDebugLoc();
268 // If the exec mask is non-zero, skip the next two instructions
269 BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
271 .addReg(AMDGPU::EXEC);
273 // Exec mask is zero: Export to NULL target...
274 BuildMI(MBB, &MI, DL, TII->get(AMDGPU::EXP))
276 .addImm(0x09) // V_008DFC_SQ_EXP_NULL
280 .addReg(AMDGPU::SREG_LIT_0)
281 .addReg(AMDGPU::SREG_LIT_0)
282 .addReg(AMDGPU::SREG_LIT_0)
283 .addReg(AMDGPU::SREG_LIT_0);
285 // ... and terminate wavefront
286 BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_ENDPGM));