//===----------------------------------------------------------------------===//
#include "AMDGPU.h"
+#include "AMDGPUSubtarget.h"
#include "SIInstrInfo.h"
#include "SIMachineFunctionInfo.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/IR/Constants.h"
using namespace llvm;
class SILowerControlFlowPass : public MachineFunctionPass {
private:
+ static const unsigned SkipThreshold = 12;
+
static char ID;
- const TargetInstrInfo *TII;
+ const SIRegisterInfo *TRI;
+ const SIInstrInfo *TII;
+
+ bool shouldSkip(MachineBasicBlock *From, MachineBasicBlock *To);
+
+ void Skip(MachineInstr &From, MachineOperand &To);
+ void SkipIfDead(MachineInstr &MI);
void If(MachineInstr &MI);
void Else(MachineInstr &MI);
void Loop(MachineInstr &MI);
void EndCf(MachineInstr &MI);
+ void Kill(MachineInstr &MI);
+ void Branch(MachineInstr &MI);
+
+ void LoadM0(MachineInstr &MI, MachineInstr *MovRel, int Offset = 0);
+ void computeIndirectRegAndOffset(unsigned VecReg, unsigned &Reg, int &Offset);
+ void IndirectSrc(MachineInstr &MI);
+ void IndirectDst(MachineInstr &MI);
+
public:
SILowerControlFlowPass(TargetMachine &tm) :
- MachineFunctionPass(ID), TII(tm.getInstrInfo()) { }
+ MachineFunctionPass(ID), TRI(nullptr), TII(nullptr) { }
- virtual bool runOnMachineFunction(MachineFunction &MF);
+ bool runOnMachineFunction(MachineFunction &MF) override;
- const char *getPassName() const {
+ const char *getPassName() const override {
return "SI Lower control flow instructions";
}
return new SILowerControlFlowPass(tm);
}
-void SILowerControlFlowPass::If(MachineInstr &MI) {
+bool SILowerControlFlowPass::shouldSkip(MachineBasicBlock *From,
+ MachineBasicBlock *To) {
+
+ unsigned NumInstr = 0;
+
+ for (MachineBasicBlock *MBB = From; MBB != To && !MBB->succ_empty();
+ MBB = *MBB->succ_begin()) {
+
+ for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end();
+ NumInstr < SkipThreshold && I != E; ++I) {
+
+ if (I->isBundle() || !I->isBundled())
+ if (++NumInstr >= SkipThreshold)
+ return true;
+ }
+ }
+
+ return false;
+}
+
+void SILowerControlFlowPass::Skip(MachineInstr &From, MachineOperand &To) {
+
+ if (!shouldSkip(*From.getParent()->succ_begin(), To.getMBB()))
+ return;
+
+ DebugLoc DL = From.getDebugLoc();
+ BuildMI(*From.getParent(), &From, DL, TII->get(AMDGPU::S_CBRANCH_EXECZ))
+ .addOperand(To)
+ .addReg(AMDGPU::EXEC);
+}
+
+void SILowerControlFlowPass::SkipIfDead(MachineInstr &MI) {
MachineBasicBlock &MBB = *MI.getParent();
DebugLoc DL = MI.getDebugLoc();
+
+ if (MBB.getParent()->getInfo<SIMachineFunctionInfo>()->getShaderType() !=
+ ShaderType::PIXEL ||
+ !shouldSkip(&MBB, &MBB.getParent()->back()))
+ return;
+
+ MachineBasicBlock::iterator Insert = &MI;
+ ++Insert;
+
+ // If the exec mask is non-zero, skip the next two instructions
+ BuildMI(MBB, Insert, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
+ .addImm(3)
+ .addReg(AMDGPU::EXEC);
+
+ // Exec mask is zero: Export to NULL target...
+ BuildMI(MBB, Insert, DL, TII->get(AMDGPU::EXP))
+ .addImm(0)
+ .addImm(0x09) // V_008DFC_SQ_EXP_NULL
+ .addImm(0)
+ .addImm(1)
+ .addImm(1)
+ .addReg(AMDGPU::VGPR0)
+ .addReg(AMDGPU::VGPR0)
+ .addReg(AMDGPU::VGPR0)
+ .addReg(AMDGPU::VGPR0);
+
+ // ... and terminate wavefront
+ BuildMI(MBB, Insert, DL, TII->get(AMDGPU::S_ENDPGM));
+}
+
+void SILowerControlFlowPass::If(MachineInstr &MI) {
+ MachineBasicBlock &MBB = *MI.getParent();
+ DebugLoc DL = MI.getDebugLoc();
unsigned Reg = MI.getOperand(0).getReg();
unsigned Vcc = MI.getOperand(1).getReg();
.addReg(AMDGPU::EXEC)
.addReg(Reg);
+ Skip(MI, MI.getOperand(2));
+
MI.eraseFromParent();
}
void SILowerControlFlowPass::Else(MachineInstr &MI) {
-
MachineBasicBlock &MBB = *MI.getParent();
DebugLoc DL = MI.getDebugLoc();
unsigned Dst = MI.getOperand(0).getReg();
unsigned Src = MI.getOperand(1).getReg();
- BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_OR_SAVEEXEC_B64), Dst)
+ BuildMI(MBB, MBB.getFirstNonPHI(), DL,
+ TII->get(AMDGPU::S_OR_SAVEEXEC_B64), Dst)
.addReg(Src); // Saved EXEC
BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_XOR_B64), AMDGPU::EXEC)
.addReg(AMDGPU::EXEC)
.addReg(Dst);
+ Skip(MI, MI.getOperand(2));
+
MI.eraseFromParent();
}
void SILowerControlFlowPass::Break(MachineInstr &MI) {
-
MachineBasicBlock &MBB = *MI.getParent();
DebugLoc DL = MI.getDebugLoc();
}
void SILowerControlFlowPass::IfBreak(MachineInstr &MI) {
-
MachineBasicBlock &MBB = *MI.getParent();
DebugLoc DL = MI.getDebugLoc();
}
void SILowerControlFlowPass::ElseBreak(MachineInstr &MI) {
-
MachineBasicBlock &MBB = *MI.getParent();
DebugLoc DL = MI.getDebugLoc();
}
void SILowerControlFlowPass::Loop(MachineInstr &MI) {
-
MachineBasicBlock &MBB = *MI.getParent();
DebugLoc DL = MI.getDebugLoc();
unsigned Src = MI.getOperand(0).getReg();
}
void SILowerControlFlowPass::EndCf(MachineInstr &MI) {
-
MachineBasicBlock &MBB = *MI.getParent();
DebugLoc DL = MI.getDebugLoc();
unsigned Reg = MI.getOperand(0).getReg();
MI.eraseFromParent();
}
+void SILowerControlFlowPass::Branch(MachineInstr &MI) {
+ if (MI.getOperand(0).getMBB() == MI.getParent()->getNextNode())
+ MI.eraseFromParent();
+
+ // If these aren't equal, this is probably an infinite loop.
+}
+
+void SILowerControlFlowPass::Kill(MachineInstr &MI) {
+ MachineBasicBlock &MBB = *MI.getParent();
+ DebugLoc DL = MI.getDebugLoc();
+ const MachineOperand &Op = MI.getOperand(0);
+
+#ifndef NDEBUG
+ const SIMachineFunctionInfo *MFI
+ = MBB.getParent()->getInfo<SIMachineFunctionInfo>();
+ // Kill is only allowed in pixel / geometry shaders.
+ assert(MFI->getShaderType() == ShaderType::PIXEL ||
+ MFI->getShaderType() == ShaderType::GEOMETRY);
+#endif
+
+ // Clear this thread from the exec mask if the operand is negative
+ if ((Op.isImm())) {
+ // Constant operand: Set exec mask to 0 or do nothing
+ if (Op.getImm() & 0x80000000) {
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B64), AMDGPU::EXEC)
+ .addImm(0);
+ }
+ } else {
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMPX_LE_F32_e32), AMDGPU::VCC)
+ .addImm(0)
+ .addOperand(Op);
+ }
+
+ MI.eraseFromParent();
+}
+
+void SILowerControlFlowPass::LoadM0(MachineInstr &MI, MachineInstr *MovRel, int Offset) {
+
+ MachineBasicBlock &MBB = *MI.getParent();
+ DebugLoc DL = MI.getDebugLoc();
+ MachineBasicBlock::iterator I = MI;
+
+ unsigned Save = MI.getOperand(1).getReg();
+ unsigned Idx = MI.getOperand(3).getReg();
+
+ if (AMDGPU::SReg_32RegClass.contains(Idx)) {
+ if (Offset) {
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_ADD_I32), AMDGPU::M0)
+ .addReg(Idx)
+ .addImm(Offset);
+ } else {
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
+ .addReg(Idx);
+ }
+ MBB.insert(I, MovRel);
+ } else {
+
+ assert(AMDGPU::SReg_64RegClass.contains(Save));
+ assert(AMDGPU::VGPR_32RegClass.contains(Idx));
+
+ // Save the EXEC mask
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B64), Save)
+ .addReg(AMDGPU::EXEC);
+
+ // Read the next variant into VCC (lower 32 bits) <- also loop target
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_READFIRSTLANE_B32),
+ AMDGPU::VCC_LO)
+ .addReg(Idx);
+
+ // Move index from VCC into M0
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
+ .addReg(AMDGPU::VCC_LO);
+
+ // Compare the just read M0 value to all possible Idx values
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMP_EQ_U32_e32), AMDGPU::VCC)
+ .addReg(AMDGPU::M0)
+ .addReg(Idx);
+
+ // Update EXEC, save the original EXEC value to VCC
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_AND_SAVEEXEC_B64), AMDGPU::VCC)
+ .addReg(AMDGPU::VCC);
+
+ if (Offset) {
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_ADD_I32), AMDGPU::M0)
+ .addReg(AMDGPU::M0)
+ .addImm(Offset);
+ }
+ // Do the actual move
+ MBB.insert(I, MovRel);
+
+ // Update EXEC, switch all done bits to 0 and all todo bits to 1
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_XOR_B64), AMDGPU::EXEC)
+ .addReg(AMDGPU::EXEC)
+ .addReg(AMDGPU::VCC);
+
+ // Loop back to V_READFIRSTLANE_B32 if there are still variants to cover
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
+ .addImm(-7)
+ .addReg(AMDGPU::EXEC);
+
+ // Restore EXEC
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B64), AMDGPU::EXEC)
+ .addReg(Save);
+
+ }
+ MI.eraseFromParent();
+}
+
+/// \param @VecReg The register which holds element zero of the vector
+/// being addressed into.
+/// \param[out] @Reg The base register to use in the indirect addressing instruction.
+/// \param[in,out] @Offset As an input, this is the constant offset part of the
+// indirect Index. e.g. v0 = v[VecReg + Offset]
+// As an output, this is a constant value that needs
+// to be added to the value stored in M0.
+void SILowerControlFlowPass::computeIndirectRegAndOffset(unsigned VecReg,
+ unsigned &Reg,
+ int &Offset) {
+ unsigned SubReg = TRI->getSubReg(VecReg, AMDGPU::sub0);
+ if (!SubReg)
+ SubReg = VecReg;
+
+ const TargetRegisterClass *RC = TRI->getPhysRegClass(SubReg);
+ int RegIdx = TRI->getHWRegIndex(SubReg) + Offset;
+
+ if (RegIdx < 0) {
+ Offset = RegIdx;
+ RegIdx = 0;
+ } else {
+ Offset = 0;
+ }
+
+ Reg = RC->getRegister(RegIdx);
+}
+
+void SILowerControlFlowPass::IndirectSrc(MachineInstr &MI) {
+
+ MachineBasicBlock &MBB = *MI.getParent();
+ DebugLoc DL = MI.getDebugLoc();
+
+ unsigned Dst = MI.getOperand(0).getReg();
+ unsigned Vec = MI.getOperand(2).getReg();
+ int Off = MI.getOperand(4).getImm();
+ unsigned Reg;
+
+ computeIndirectRegAndOffset(Vec, Reg, Off);
+
+ MachineInstr *MovRel =
+ BuildMI(*MBB.getParent(), DL, TII->get(AMDGPU::V_MOVRELS_B32_e32), Dst)
+ .addReg(Reg)
+ .addReg(AMDGPU::M0, RegState::Implicit)
+ .addReg(Vec, RegState::Implicit);
+
+ LoadM0(MI, MovRel, Off);
+}
+
+void SILowerControlFlowPass::IndirectDst(MachineInstr &MI) {
+
+ MachineBasicBlock &MBB = *MI.getParent();
+ DebugLoc DL = MI.getDebugLoc();
+
+ unsigned Dst = MI.getOperand(0).getReg();
+ int Off = MI.getOperand(4).getImm();
+ unsigned Val = MI.getOperand(5).getReg();
+ unsigned Reg;
+
+ computeIndirectRegAndOffset(Dst, Reg, Off);
+
+ MachineInstr *MovRel =
+ BuildMI(*MBB.getParent(), DL, TII->get(AMDGPU::V_MOVRELD_B32_e32))
+ .addReg(Reg, RegState::Define)
+ .addReg(Val)
+ .addReg(AMDGPU::M0, RegState::Implicit)
+ .addReg(Dst, RegState::Implicit);
+
+ LoadM0(MI, MovRel, Off);
+}
+
bool SILowerControlFlowPass::runOnMachineFunction(MachineFunction &MF) {
+ TII = static_cast<const SIInstrInfo *>(MF.getSubtarget().getInstrInfo());
+ TRI =
+ static_cast<const SIRegisterInfo *>(MF.getSubtarget().getRegisterInfo());
+ SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
- bool HaveCf = false;
+ bool HaveKill = false;
+ bool NeedWQM = false;
+ bool NeedFlat = false;
+ unsigned Depth = 0;
for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
BI != BE; ++BI) {
MachineBasicBlock &MBB = *BI;
- for (MachineBasicBlock::iterator I = MBB.begin(), Next = llvm::next(I);
- I != MBB.end(); I = Next) {
+ MachineBasicBlock::iterator I, Next;
+ for (I = MBB.begin(); I != MBB.end(); I = Next) {
+ Next = std::next(I);
- Next = llvm::next(I);
MachineInstr &MI = *I;
+ if (TII->isWQM(MI.getOpcode()) || TII->isDS(MI.getOpcode()))
+ NeedWQM = true;
+
+ // Flat uses m0 in case it needs to access LDS.
+ if (TII->isFLAT(MI.getOpcode()))
+ NeedFlat = true;
+
switch (MI.getOpcode()) {
default: break;
case AMDGPU::SI_IF:
+ ++Depth;
If(MI);
break;
break;
case AMDGPU::SI_LOOP:
+ ++Depth;
Loop(MI);
break;
case AMDGPU::SI_END_CF:
- HaveCf = true;
+ if (--Depth == 0 && HaveKill) {
+ SkipIfDead(MI);
+ HaveKill = false;
+ }
EndCf(MI);
break;
+
+ case AMDGPU::SI_KILL:
+ if (Depth == 0)
+ SkipIfDead(MI);
+ else
+ HaveKill = true;
+ Kill(MI);
+ break;
+
+ case AMDGPU::S_BRANCH:
+ Branch(MI);
+ break;
+
+ case AMDGPU::SI_INDIRECT_SRC:
+ IndirectSrc(MI);
+ break;
+
+ case AMDGPU::SI_INDIRECT_DST_V1:
+ case AMDGPU::SI_INDIRECT_DST_V2:
+ case AMDGPU::SI_INDIRECT_DST_V4:
+ case AMDGPU::SI_INDIRECT_DST_V8:
+ case AMDGPU::SI_INDIRECT_DST_V16:
+ IndirectDst(MI);
+ break;
}
}
}
- // TODO: What is this good for?
- unsigned ShaderType = MF.getInfo<SIMachineFunctionInfo>()->ShaderType;
- if (HaveCf && ShaderType == ShaderType::PIXEL) {
- for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
- BI != BE; ++BI) {
-
- MachineBasicBlock &MBB = *BI;
- if (MBB.succ_empty()) {
-
- MachineInstr &MI = *MBB.getFirstNonPHI();
- DebugLoc DL = MI.getDebugLoc();
-
- // If the exec mask is non-zero, skip the next two instructions
- BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
- .addImm(3)
- .addReg(AMDGPU::EXEC);
-
- // Exec mask is zero: Export to NULL target...
- BuildMI(MBB, &MI, DL, TII->get(AMDGPU::EXP))
- .addImm(0)
- .addImm(0x09) // V_008DFC_SQ_EXP_NULL
- .addImm(0)
- .addImm(1)
- .addImm(1)
- .addReg(AMDGPU::SREG_LIT_0)
- .addReg(AMDGPU::SREG_LIT_0)
- .addReg(AMDGPU::SREG_LIT_0)
- .addReg(AMDGPU::SREG_LIT_0);
-
- // ... and terminate wavefront
- BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_ENDPGM));
- }
- }
+ if (NeedWQM && MFI->getShaderType() == ShaderType::PIXEL) {
+ MachineBasicBlock &MBB = MF.front();
+ BuildMI(MBB, MBB.getFirstNonPHI(), DebugLoc(), TII->get(AMDGPU::S_WQM_B64),
+ AMDGPU::EXEC).addReg(AMDGPU::EXEC);
+ }
+
+ // FIXME: This seems inappropriate to do here.
+ if (NeedFlat && MFI->IsKernel) {
+ // Insert the prologue initializing the SGPRs pointing to the scratch space
+ // for flat accesses.
+ const MachineFrameInfo *FrameInfo = MF.getFrameInfo();
+
+ // TODO: What to use with function calls?
+
+ // FIXME: This is reporting stack size that is used in a scratch buffer
+ // rather than registers as well.
+ uint64_t StackSizeBytes = FrameInfo->getStackSize();
+
+ int IndirectBegin
+ = static_cast<const AMDGPUInstrInfo*>(TII)->getIndirectIndexBegin(MF);
+ // Convert register index to 256-byte unit.
+ uint64_t StackOffset = IndirectBegin < 0 ? 0 : (4 * IndirectBegin / 256);
+
+ assert((StackSizeBytes < 0xffff) && StackOffset < 0xffff &&
+ "Stack limits should be smaller than 16-bits");
+
+ // Initialize the flat scratch register pair.
+ // TODO: Can we use one s_mov_b64 here?
+
+ // Offset is in units of 256-bytes.
+ MachineBasicBlock &MBB = MF.front();
+ DebugLoc NoDL;
+ MachineBasicBlock::iterator Start = MBB.getFirstNonPHI();
+ const MCInstrDesc &SMovK = TII->get(AMDGPU::S_MOVK_I32);
+
+ assert(isInt<16>(StackOffset) && isInt<16>(StackSizeBytes));
+
+ BuildMI(MBB, Start, NoDL, SMovK, AMDGPU::FLAT_SCR_LO)
+ .addImm(StackOffset);
+
+ // Documentation says size is "per-thread scratch size in bytes"
+ BuildMI(MBB, Start, NoDL, SMovK, AMDGPU::FLAT_SCR_HI)
+ .addImm(StackSizeBytes);
}
return true;