[llvm] r208432 - R600/SI: Teach SIInstrInfo::moveToVALU() how to move S_LOAD_*_IMM instructions
Tom Stellard
thomas.stellard at amd.com
Fri May 9 09:42:23 PDT 2014
Author: tstellar
Date: Fri May 9 11:42:22 2014
New Revision: 208432
URL: http://llvm.org/viewvc/llvm-project?rev=208432&view=rev
Log:
R600/SI: Teach SIInstrInfo::moveToVALU() how to move S_LOAD_*_IMM instructions
Modified:
llvm/trunk/lib/Target/R600/SIInstrInfo.cpp
llvm/trunk/test/CodeGen/R600/salu-to-valu.ll
Modified: llvm/trunk/lib/Target/R600/SIInstrInfo.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/SIInstrInfo.cpp?rev=208432&r1=208431&r2=208432&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/SIInstrInfo.cpp (original)
+++ llvm/trunk/lib/Target/R600/SIInstrInfo.cpp Fri May 9 11:42:22 2014
@@ -661,8 +661,11 @@ unsigned SIInstrInfo::getVALUOp(const Ma
case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
+ case AMDGPU::S_LOAD_DWORD_IMM:
case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
+ case AMDGPU::S_LOAD_DWORDX2_IMM:
case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
+ case AMDGPU::S_LOAD_DWORDX4_IMM:
case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
}
}
@@ -1029,15 +1032,39 @@ void SIInstrInfo::legalizeOperands(Machi
void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
MachineBasicBlock *MBB = MI->getParent();
switch (MI->getOpcode()) {
+ case AMDGPU::S_LOAD_DWORD_IMM:
case AMDGPU::S_LOAD_DWORD_SGPR:
+ case AMDGPU::S_LOAD_DWORDX2_IMM:
case AMDGPU::S_LOAD_DWORDX2_SGPR:
+ case AMDGPU::S_LOAD_DWORDX4_IMM:
case AMDGPU::S_LOAD_DWORDX4_SGPR:
unsigned NewOpcode = getVALUOp(*MI);
- unsigned Offset = MI->getOperand(2).getReg();
+ unsigned RegOffset;
+ unsigned ImmOffset;
+ if (MI->getOperand(2).isReg()) {
+ RegOffset = MI->getOperand(2).getReg();
+ ImmOffset = 0;
+ } else {
+ assert(MI->getOperand(2).isImm());
+ // SMRD instructions take a dword offsets and MUBUF instructions
+ // take a byte offset.
+ ImmOffset = MI->getOperand(2).getImm() << 2;
+ RegOffset = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
+ if (isUInt<12>(ImmOffset)) {
+ BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
+ RegOffset)
+ .addImm(0);
+ } else {
+ BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32),
+ RegOffset)
+ .addImm(ImmOffset);
+ ImmOffset = 0;
+ }
+ }
unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
- unsigned DWord0 = Offset;
+ unsigned DWord0 = RegOffset;
unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
@@ -1058,9 +1085,13 @@ void SIInstrInfo::moveSMRDToVALU(Machine
.addReg(DWord3)
.addImm(AMDGPU::sub3);
MI->setDesc(get(NewOpcode));
- MI->getOperand(2).setReg(MI->getOperand(1).getReg());
+ if (MI->getOperand(2).isReg()) {
+ MI->getOperand(2).setReg(MI->getOperand(1).getReg());
+ } else {
+ MI->getOperand(2).ChangeToRegister(MI->getOperand(1).getReg(), false);
+ }
MI->getOperand(1).setReg(SRsrc);
- MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(0));
+ MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(ImmOffset));
}
}
Modified: llvm/trunk/test/CodeGen/R600/salu-to-valu.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/R600/salu-to-valu.ll?rev=208432&r1=208431&r2=208432&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/R600/salu-to-valu.ll (original)
+++ llvm/trunk/test/CodeGen/R600/salu-to-valu.ll Fri May 9 11:42:22 2014
@@ -74,3 +74,17 @@ endif:
store i32 %6, i32 addrspace(1)* %out
ret void
}
+
+; Test moving ann SMRD with an immediate offset to the VALU
+
+; CHECK-LABEL: @smrd_valu2
+; CHECK: BUFFER_LOAD_DWORD
+define void @smrd_valu2(i32 addrspace(1)* %out, [8 x i32] addrspace(2)* %in) {
+entry:
+ %0 = call i32 @llvm.r600.read.tidig.x() nounwind readnone
+ %1 = add i32 %0, 4
+ %2 = getelementptr [8 x i32] addrspace(2)* %in, i32 %0, i32 4
+ %3 = load i32 addrspace(2)* %2
+ store i32 %3, i32 addrspace(1)* %out
+ ret void
+}
More information about the llvm-commits
mailing list