#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/IR/Constants.h"
using namespace llvm;
static char ID;
const TargetRegisterInfo *TRI;
- const TargetInstrInfo *TII;
+ const SIInstrInfo *TII;
bool shouldSkip(MachineBasicBlock *From, MachineBasicBlock *To);
MachineBasicBlock &MBB = *MI.getParent();
DebugLoc DL = MI.getDebugLoc();
- if (!shouldSkip(&MBB, &MBB.getParent()->back()))
+ if (MBB.getParent()->getInfo<SIMachineFunctionInfo>()->ShaderType !=
+ ShaderType::PIXEL ||
+ !shouldSkip(&MBB, &MBB.getParent()->back()))
return;
MachineBasicBlock::iterator Insert = &MI;
}
void SILowerControlFlowPass::Branch(MachineInstr &MI) {
- assert(MI.getOperand(0).getMBB() == MI.getParent()->getNextNode());
- MI.eraseFromParent();
+ 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);
- // Kill is only allowed in pixel shaders
+ // Kill is only allowed in pixel / geometry shaders
assert(MBB.getParent()->getInfo<SIMachineFunctionInfo>()->ShaderType ==
- ShaderType::PIXEL);
-
- // Clear this pixel from the exec mask if the operand is negative
- BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMPX_LE_F32_e32), AMDGPU::VCC)
- .addImm(0)
- .addOperand(MI.getOperand(0));
+ ShaderType::PIXEL ||
+ MBB.getParent()->getInfo<SIMachineFunctionInfo>()->ShaderType ==
+ ShaderType::GEOMETRY);
+
+ // 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();
}
.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_e32), AMDGPU::VCC)
+ 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);
+ .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)
}
bool SILowerControlFlowPass::runOnMachineFunction(MachineFunction &MF) {
- TII = MF.getTarget().getInstrInfo();
+ TII = static_cast<const SIInstrInfo*>(MF.getTarget().getInstrInfo());
TRI = MF.getTarget().getRegisterInfo();
SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
BI != BE; ++BI) {
MachineBasicBlock &MBB = *BI;
- for (MachineBasicBlock::iterator I = MBB.begin(), Next = llvm::next(I);
+ for (MachineBasicBlock::iterator I = MBB.begin(), Next = std::next(I);
I != MBB.end(); I = Next) {
- Next = llvm::next(I);
+ Next = std::next(I);
MachineInstr &MI = *I;
+ if (TII->isDS(MI.getOpcode())) {
+ NeedM0 = true;
+ NeedWQM = true;
+ }
+
switch (MI.getOpcode()) {
default: break;
case AMDGPU::SI_IF:
IndirectDst(MI);
break;
- case AMDGPU::DS_READ_B32:
- NeedWQM = true;
- // Fall through
- case AMDGPU::DS_WRITE_B32:
- case AMDGPU::DS_ADD_U32_RTN:
- NeedM0 = true;
- break;
-
case AMDGPU::V_INTERP_P1_F32:
case AMDGPU::V_INTERP_P2_F32:
case AMDGPU::V_INTERP_MOV_F32:
AMDGPU::M0).addImm(0xffffffff);
}
- if (NeedWQM && MFI->ShaderType != ShaderType::COMPUTE) {
+ if (NeedWQM && MFI->ShaderType == ShaderType::PIXEL) {
MachineBasicBlock &MBB = MF.front();
BuildMI(MBB, MBB.getFirstNonPHI(), DebugLoc(), TII->get(AMDGPU::S_WQM_B64),
AMDGPU::EXEC).addReg(AMDGPU::EXEC);