//===----------------------------------------------------------------------===//
//
/// \file
-/// \brief This pass lowers the pseudo control flow instructions (SI_IF_NZ, ELSE, ENDIF)
-/// to predicated instructions.
+/// \brief This pass lowers the pseudo control flow instructions to real
+/// machine instructions.
///
-/// All control flow (except loops) is handled using predicated instructions and
+/// All control flow is handled using predicated instructions and
/// a predicate stack. Each Scalar ALU controls the operations of 64 Vector
/// ALUs. The Scalar ALU can update the predicate for any of the Vector ALUs
/// by writting to the 64-bit EXEC register (each bit corresponds to a
///
/// For example:
/// %VCC = V_CMP_GT_F32 %VGPR1, %VGPR2
-/// SI_IF_NZ %VCC
+/// %SGPR0 = SI_IF %VCC
/// %VGPR0 = V_ADD_F32 %VGPR0, %VGPR0
-/// ELSE
+/// %SGPR0 = SI_ELSE %SGPR0
/// %VGPR0 = V_SUB_F32 %VGPR0, %VGPR0
-/// ENDIF
+/// SI_END_CF %SGPR0
///
/// becomes:
///
/// %SGPR0 = S_AND_SAVEEXEC_B64 %VCC // Save and update the exec mask
/// %SGPR0 = S_XOR_B64 %SGPR0, %EXEC // Clear live bits from saved exec mask
-/// S_CBRANCH_EXECZ label0 // This instruction is an
+/// S_CBRANCH_EXECZ label0 // This instruction is an optional
/// // optimization which allows us to
/// // branch if all the bits of
/// // EXEC are zero.
/// // instruction again.
/// %VGPR0 = V_SUB_F32 %VGPR0, %VGPR // Do the THEN block
/// label1:
-/// %EXEC = S_OR_B64 %EXEC, %SGPR2 // Re-enable saved exec mask bits
+/// %EXEC = S_OR_B64 %EXEC, %SGPR0 // Re-enable saved exec mask bits
//===----------------------------------------------------------------------===//
#include "AMDGPU.h"
+#include "AMDGPUSubtarget.h"
#include "SIInstrInfo.h"
#include "SIMachineFunctionInfo.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;
- std::vector<unsigned> PredicateStack;
- std::vector<unsigned> UnusedRegisters;
+ 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 Break(MachineInstr &MI);
+ void IfBreak(MachineInstr &MI);
+ void ElseBreak(MachineInstr &MI);
+ void Loop(MachineInstr &MI);
+ void EndCf(MachineInstr &MI);
+
+ void Kill(MachineInstr &MI);
+ void Branch(MachineInstr &MI);
- unsigned allocReg();
- void freeReg(unsigned Reg);
+ void InitM0ForLDS(MachineBasicBlock::iterator MI);
+ void LoadM0(MachineInstr &MI, MachineInstr *MovRel);
+ 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);
}
-bool SILowerControlFlowPass::runOnMachineFunction(MachineFunction &MF) {
+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) {
- // Find all the unused registers that can be used for the predicate stack.
- for (TargetRegisterClass::iterator I = AMDGPU::SReg_64RegClass.begin(),
- S = AMDGPU::SReg_64RegClass.end();
- I != S; ++I) {
- unsigned Reg = *I;
- if (!MF.getRegInfo().isPhysRegUsed(Reg)) {
- UnusedRegisters.insert(UnusedRegisters.begin(), Reg);
+ if (I->isBundle() || !I->isBundled())
+ if (++NumInstr >= SkipThreshold)
+ return true;
}
}
- for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
- BB != BB_E; ++BB) {
- MachineBasicBlock &MBB = *BB;
- for (MachineBasicBlock::iterator I = MBB.begin(), Next = llvm::next(I);
- I != MBB.end(); I = Next) {
- Next = llvm::next(I);
+ 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();
+
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_AND_SAVEEXEC_B64), Reg)
+ .addReg(Vcc);
+
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_XOR_B64), Reg)
+ .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, 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();
+
+ unsigned Dst = MI.getOperand(0).getReg();
+ unsigned Src = MI.getOperand(1).getReg();
+
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_OR_B64), Dst)
+ .addReg(AMDGPU::EXEC)
+ .addReg(Src);
+
+ MI.eraseFromParent();
+}
+
+void SILowerControlFlowPass::IfBreak(MachineInstr &MI) {
+ MachineBasicBlock &MBB = *MI.getParent();
+ DebugLoc DL = MI.getDebugLoc();
+
+ unsigned Dst = MI.getOperand(0).getReg();
+ unsigned Vcc = MI.getOperand(1).getReg();
+ unsigned Src = MI.getOperand(2).getReg();
+
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_OR_B64), Dst)
+ .addReg(Vcc)
+ .addReg(Src);
+
+ MI.eraseFromParent();
+}
+
+void SILowerControlFlowPass::ElseBreak(MachineInstr &MI) {
+ MachineBasicBlock &MBB = *MI.getParent();
+ DebugLoc DL = MI.getDebugLoc();
+
+ unsigned Dst = MI.getOperand(0).getReg();
+ unsigned Saved = MI.getOperand(1).getReg();
+ unsigned Src = MI.getOperand(2).getReg();
+
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_OR_B64), Dst)
+ .addReg(Saved)
+ .addReg(Src);
+
+ MI.eraseFromParent();
+}
+
+void SILowerControlFlowPass::Loop(MachineInstr &MI) {
+ MachineBasicBlock &MBB = *MI.getParent();
+ DebugLoc DL = MI.getDebugLoc();
+ unsigned Src = MI.getOperand(0).getReg();
+
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_ANDN2_B64), AMDGPU::EXEC)
+ .addReg(AMDGPU::EXEC)
+ .addReg(Src);
+
+ BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
+ .addOperand(MI.getOperand(1))
+ .addReg(AMDGPU::EXEC);
+
+ MI.eraseFromParent();
+}
+
+void SILowerControlFlowPass::EndCf(MachineInstr &MI) {
+ MachineBasicBlock &MBB = *MI.getParent();
+ DebugLoc DL = MI.getDebugLoc();
+ unsigned Reg = MI.getOperand(0).getReg();
+
+ BuildMI(MBB, MBB.getFirstNonPHI(), DL,
+ TII->get(AMDGPU::S_OR_B64), AMDGPU::EXEC)
+ .addReg(AMDGPU::EXEC)
+ .addReg(Reg);
+
+ 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() || Op.isFPImm())) {
+ // Constant operand: Set exec mask to 0 or do nothing
+ if (Op.isImm() ? (Op.getImm() & 0x80000000) :
+ Op.getFPImm()->isNegative()) {
+ 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();
+}
+
+/// The m0 register stores the maximum allowable address for LDS reads and
+/// writes. Its value must be at least the size in bytes of LDS allocated by
+/// the shader. For simplicity, we set it to the maximum possible value.
+void SILowerControlFlowPass::InitM0ForLDS(MachineBasicBlock::iterator MI) {
+ BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32),
+ AMDGPU::M0).addImm(0xffffffff);
+}
+
+void SILowerControlFlowPass::LoadM0(MachineInstr &MI, MachineInstr *MovRel) {
+
+ 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)) {
+ 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::VReg_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);
+
+ // 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);
+
+ }
+ // FIXME: Are there any values other than the LDS address clamp that need to
+ // be stored in the m0 register and may be live for more than a few
+ // instructions? If so, we should save the m0 register at the beginning
+ // of this function and restore it here.
+ // FIXME: Add support for LDS direct loads.
+ InitM0ForLDS(&MI);
+ MI.eraseFromParent();
+}
+
+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();
+ unsigned Off = MI.getOperand(4).getImm();
+ unsigned SubReg = TRI->getSubReg(Vec, AMDGPU::sub0);
+ if (!SubReg)
+ SubReg = Vec;
+
+ MachineInstr *MovRel =
+ BuildMI(*MBB.getParent(), DL, TII->get(AMDGPU::V_MOVRELS_B32_e32), Dst)
+ .addReg(SubReg + Off)
+ .addReg(AMDGPU::M0, RegState::Implicit)
+ .addReg(Vec, RegState::Implicit);
+
+ LoadM0(MI, MovRel);
+}
+
+void SILowerControlFlowPass::IndirectDst(MachineInstr &MI) {
+
+ MachineBasicBlock &MBB = *MI.getParent();
+ DebugLoc DL = MI.getDebugLoc();
+
+ unsigned Dst = MI.getOperand(0).getReg();
+ unsigned Off = MI.getOperand(4).getImm();
+ unsigned Val = MI.getOperand(5).getReg();
+ unsigned SubReg = TRI->getSubReg(Dst, AMDGPU::sub0);
+ if (!SubReg)
+ SubReg = Dst;
+
+ MachineInstr *MovRel =
+ BuildMI(*MBB.getParent(), DL, TII->get(AMDGPU::V_MOVRELD_B32_e32))
+ .addReg(SubReg + Off, RegState::Define)
+ .addReg(Val)
+ .addReg(AMDGPU::M0, RegState::Implicit)
+ .addReg(Dst, RegState::Implicit);
+
+ LoadM0(MI, MovRel);
+}
+
+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 HaveKill = false;
+ bool NeedM0 = false;
+ bool NeedWQM = false;
+ unsigned Depth = 0;
+
+ for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
+ BI != BE; ++BI) {
+
+ MachineBasicBlock &MBB = *BI;
+ MachineBasicBlock::iterator I, Next;
+ for (I = MBB.begin(); I != MBB.end(); I = Next) {
+ Next = std::next(I);
+
MachineInstr &MI = *I;
- unsigned Reg;
+ if (TII->isDS(MI.getOpcode())) {
+ NeedM0 = true;
+ NeedWQM = true;
+ }
+
switch (MI.getOpcode()) {
default: break;
- case AMDGPU::SI_IF_NZ:
- Reg = allocReg();
- BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_AND_SAVEEXEC_B64),
- Reg)
- .addOperand(MI.getOperand(0)); // VCC
- BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_XOR_B64),
- Reg)
- .addReg(Reg)
- .addReg(AMDGPU::EXEC);
- MI.eraseFromParent();
- PredicateStack.push_back(Reg);
+ case AMDGPU::SI_IF:
+ ++Depth;
+ If(MI);
+ break;
+
+ case AMDGPU::SI_ELSE:
+ Else(MI);
break;
- case AMDGPU::ELSE:
- Reg = PredicateStack.back();
- BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_OR_SAVEEXEC_B64),
- Reg)
- .addReg(Reg);
- BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_XOR_B64),
- AMDGPU::EXEC)
- .addReg(Reg)
- .addReg(AMDGPU::EXEC);
- MI.eraseFromParent();
+ case AMDGPU::SI_BREAK:
+ Break(MI);
break;
- case AMDGPU::ENDIF:
- Reg = PredicateStack.back();
- PredicateStack.pop_back();
- BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_OR_B64),
- AMDGPU::EXEC)
- .addReg(AMDGPU::EXEC)
- .addReg(Reg);
- freeReg(Reg);
-
- if (MF.getInfo<SIMachineFunctionInfo>()->ShaderType == ShaderType::PIXEL &&
- PredicateStack.empty()) {
- // If the exec mask is non-zero, skip the next two instructions
- BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_CBRANCH_EXECNZ))
- .addImm(3)
- .addReg(AMDGPU::EXEC);
-
- // Exec mask is zero: Export to NULL target...
- BuildMI(MBB, I, MBB.findDebugLoc(I), 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, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_ENDPGM));
+ case AMDGPU::SI_IF_BREAK:
+ IfBreak(MI);
+ break;
+
+ case AMDGPU::SI_ELSE_BREAK:
+ ElseBreak(MI);
+ break;
+
+ case AMDGPU::SI_LOOP:
+ ++Depth;
+ Loop(MI);
+ break;
+
+ case AMDGPU::SI_END_CF:
+ if (--Depth == 0 && HaveKill) {
+ SkipIfDead(MI);
+ HaveKill = false;
}
- MI.eraseFromParent();
+ 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;
+
+ case AMDGPU::V_INTERP_P1_F32:
+ case AMDGPU::V_INTERP_P2_F32:
+ case AMDGPU::V_INTERP_MOV_F32:
+ NeedWQM = true;
+ break;
+
}
}
}
- return true;
-}
-
-unsigned SILowerControlFlowPass::allocReg() {
- assert(!UnusedRegisters.empty() && "Ran out of registers for predicate stack");
- unsigned Reg = UnusedRegisters.back();
- UnusedRegisters.pop_back();
- return Reg;
-}
+ if (NeedM0) {
+ MachineBasicBlock &MBB = MF.front();
+ // Initialize M0 to a value that won't cause LDS access to be discarded
+ // due to offset clamping
+ InitM0ForLDS(MBB.getFirstNonPHI());
+ }
-void SILowerControlFlowPass::freeReg(unsigned Reg) {
+ 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);
+ }
- UnusedRegisters.push_back(Reg);
+ return true;
}