radeon/llvm: Handle TGSI KIL opcode for SI.

Fixes piglit fp-kil and glBitmap() with radeonsi.

Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
This commit is contained in:
Michel Dänzer
2012-08-28 15:59:30 +02:00
committed by Michel Dänzer
parent 16e42a5dd0
commit 70f9dbe298
3 changed files with 44 additions and 0 deletions
@@ -129,6 +129,9 @@ MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
case AMDGPU::SI_INTERP_CONST:
LowerSI_INTERP_CONST(MI, *BB, I);
break;
case AMDGPU::SI_KIL:
LowerSI_KIL(MI, *BB, I, MRI);
break;
case AMDGPU::SI_V_CNDLT:
LowerSI_V_CNDLT(MI, *BB, I, MRI);
break;
@@ -193,6 +196,38 @@ void SITargetLowering::LowerSI_INTERP_CONST(MachineInstr *MI,
MI->eraseFromParent();
}
void SITargetLowering::LowerSI_KIL(MachineInstr *MI, MachineBasicBlock &BB,
MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const
{
// Clear this pixel from the exec mask if the operand is negative
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CMPX_LE_F32_e32),
AMDGPU::VCC)
.addReg(AMDGPU::SREG_LIT_0)
.addOperand(MI->getOperand(0));
// If the exec mask is non-zero, skip the next two instructions
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_CBRANCH_EXECNZ))
.addImm(3)
.addReg(AMDGPU::EXEC);
// Exec mask is zero: Export to NULL target...
BuildMI(BB, I, BB.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(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_ENDPGM));
MI->eraseFromParent();
}
void SITargetLowering::LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const
{
@@ -33,6 +33,8 @@ class SITargetLowering : public AMDGPUTargetLowering
MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
void LowerSI_INTERP_CONST(MachineInstr *MI, MachineBasicBlock &BB,
MachineBasicBlock::iterator I) const;
void LowerSI_KIL(MachineInstr *MI, MachineBasicBlock &BB,
MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
void LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
@@ -959,6 +959,13 @@ def SI_INTERP_CONST : InstSI <
imm:$attr, SReg_32:$params))]
>;
def SI_KIL : InstSI <
(outs),
(ins VReg_32:$src),
"SI_KIL $src",
[(int_AMDGPU_kill VReg_32:$src)]
>;
} // end usesCustomInserter
// SI Psuedo branch instructions. These are used by the CFG structurizer pass