radeon/llvm: Fix MULLO* instructions on Cayman

On Cayman, the MULLO* instructions must fill all slots in an
instruction group.
This commit is contained in:
Tom Stellard 2012-06-06 11:35:48 -04:00
parent 0c4b19ac63
commit 9c46cb2368
4 changed files with 53 additions and 14 deletions

View file

@ -21,6 +21,7 @@
#include "AMDILCodeEmitter.h"
#include "AMDILInstrInfo.h"
#include "AMDILUtilityFunctions.h"
#include "R600InstrInfo.h"
#include "R600RegisterInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
@ -50,6 +51,7 @@ private:
bool isCube;
bool isReduction;
bool isVector;
unsigned currentElement;
bool isLast;
@ -58,7 +60,7 @@ private:
public:
R600CodeEmitter(formatted_raw_ostream &OS) : MachineFunctionPass(ID),
_OS(OS), TM(NULL), isCube(false), isReduction(false),
_OS(OS), TM(NULL), isCube(false), isReduction(false), isVector(false),
isLast(true) { }
const char *getPassName() const { return "AMDGPU Machine Code Emitter"; }
@ -144,6 +146,7 @@ bool R600CodeEmitter::runOnMachineFunction(MachineFunction &MF) {
TM = &MF.getTarget();
MRI = &MF.getRegInfo();
TRI = static_cast<const R600RegisterInfo *>(TM->getRegisterInfo());
const R600InstrInfo * TII = static_cast<const R600InstrInfo *>(TM->getInstrInfo());
const AMDILSubtarget &STM = TM->getSubtarget<AMDILSubtarget>();
std::string gpu = STM.getDeviceName();
@ -157,6 +160,8 @@ bool R600CodeEmitter::runOnMachineFunction(MachineFunction &MF) {
for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
I != E; ++I) {
MachineInstr &MI = *I;
isReduction = AMDGPU::isReductionOp(MI.getOpcode());
isVector = TII->isVector(MI);
if (MI.getNumOperands() > 1 && MI.getOperand(0).isReg() && MI.getOperand(0).isDead()) {
continue;
}
@ -164,14 +169,14 @@ bool R600CodeEmitter::runOnMachineFunction(MachineFunction &MF) {
emitTexInstr(MI);
} else if (AMDGPU::isFCOp(MI.getOpcode())){
emitFCInstr(MI);
} else if (AMDGPU::isReductionOp(MI.getOpcode())) {
isReduction = true;
} else if (isReduction || isVector) {
isLast = false;
for (currentElement = 0; currentElement < 4; currentElement++) {
isLast = (currentElement == 3);
emitALUInstr(MI);
}
isReduction = false;
isVector = false;
} else if (AMDGPU::isCubeOp(MI.getOpcode())) {
isCube = true;
isLast = false;
@ -389,7 +394,7 @@ void R600CodeEmitter::emitDst(const MachineOperand & MO)
emitByte(getHWReg(MO.getReg()));
// Emit the element of the destination register (1 byte)
if (isReduction || isCube) {
if (isReduction || isCube || isVector) {
emitByte(currentElement);
} else {
emitByte(TRI->getHWRegChan(MO.getReg()));
@ -403,8 +408,9 @@ void R600CodeEmitter::emitDst(const MachineOperand & MO)
}
// Emit writemask (1 byte).
if ((isReduction && currentElement != TRI->getHWRegChan(MO.getReg()))
|| MO.getTargetFlags() & MO_FLAG_MASK) {
if (((isReduction || isVector) &&
currentElement != TRI->getHWRegChan(MO.getReg()))
|| MO.getTargetFlags() & MO_FLAG_MASK) {
emitByte(0);
} else {
emitByte(1);

View file

@ -33,6 +33,11 @@ bool R600InstrInfo::isTrig(const MachineInstr &MI) const
return get(MI.getOpcode()).TSFlags & R600_InstFlag::TRIG;
}
bool R600InstrInfo::isVector(const MachineInstr &MI) const
{
return get(MI.getOpcode()).TSFlags & R600_InstFlag::VECTOR;
}
void
R600InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
MachineBasicBlock::iterator MI, DebugLoc DL,

View file

@ -43,6 +43,10 @@ namespace llvm {
bool isTrig(const MachineInstr &MI) const;
/// isVector - Vector instructions are instructions that must fill all
/// instruction slots within an instruction group.
bool isVector(const MachineInstr &MI) const;
virtual MachineInstr * getMovImmInstr(MachineFunction *MF, unsigned DstReg,
int64_t Imm) const;
@ -59,7 +63,8 @@ namespace R600_InstFlag {
REDUCTION = (1 << 2),
FC = (1 << 3),
TRIG = (1 << 4),
OP3 = (1 << 5)
OP3 = (1 << 5),
VECTOR = (1 << 6)
};
}

View file

@ -18,8 +18,9 @@ class InstR600 <bits<32> inst, dag outs, dag ins, string asm, list<dag> pattern,
: AMDGPUInst <outs, ins, asm, pattern> {
field bits<32> Inst;
bit Trig = 0;
bit Trig = 0;
bit Op3 = 0;
bit isVector = 0;
let Inst = inst;
let Namespace = "AMDIL";
@ -31,6 +32,10 @@ class InstR600 <bits<32> inst, dag outs, dag ins, string asm, list<dag> pattern,
let TSFlags{4} = Trig;
let TSFlags{5} = Op3;
// Vector instructions are instructions that must fill all slots in an
// instruction group
let TSFlags{6} = isVector;
}
class InstR600ISA <dag outs, dag ins, string asm, list<dag> pattern> :
@ -784,6 +789,19 @@ class TRIG_HELPER_r700 <InstR600 trig_inst>: Pat <
>;
*/
//===----------------------------------------------------------------------===//
// Evergreen Only instructions
//===----------------------------------------------------------------------===//
let Predicates = [isEG] in {
def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
def MULHI_INT_eg : MULHI_INT_Common<0x90>;
def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
} // End Predicates = [isEG]
/* ------------------------------- */
/* Evergreen / Cayman Instructions */
/* ------------------------------- */
@ -812,10 +830,6 @@ class TRIG_eg <InstR600 trig, Intrinsic intr> : Pat<
def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
def SIN_eg : SIN_Common<0x8D>;
def COS_eg : COS_Common<0x8E>;
def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
def MULHI_INT_eg : MULHI_INT_Common<0x90>;
def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
def DOT4_eg : DOT4_Common<0xBE>;
def CUBE_eg : CUBE_Common<0xC0>;
@ -893,8 +907,17 @@ def VTX_READ_GLOBAL_eg : VTX_READ_eg <1,
let Predicates = [isCayman] in {
/* XXX: I'm not sure if this opcode is correct. */
def RECIP_UINT_cm : RECIP_UINT_Common<0x77>;
let isVector = 1 in {
def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
def MULHI_INT_cm : MULHI_INT_Common<0x90>;
def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
} // End isVector = 1
/* XXX: I'm not sure if this opcode is correct. */
def RECIP_UINT_cm : RECIP_UINT_Common<0x77>;
} // End isCayman