radeon/llvm: Remove AMDIL LOADCONST* instructions

This obsoletes the R600LowerInstruction and SIPropagateImmReads passes.
This commit is contained in:
Tom Stellard 2012-06-02 06:16:18 -04:00
parent 17e047242e
commit 8d53ddb375
13 changed files with 33 additions and 322 deletions

View file

@ -22,11 +22,9 @@ class AMDGPUTargetMachine;
// R600 Passes
FunctionPass* createR600KernelParametersPass(const TargetData* TD);
FunctionPass *createR600CodeEmitterPass(formatted_raw_ostream &OS);
FunctionPass *createR600LowerInstructionsPass(TargetMachine &tm);
// SI Passes
FunctionPass *createSIAssignInterpRegsPass(TargetMachine &tm);
FunctionPass *createSIPropagateImmReadsPass(TargetMachine &tm);
FunctionPass *createSICodeEmitterPass(formatted_raw_ostream &OS);
// Passes common to R600 and SI

View file

@ -135,9 +135,7 @@ bool AMDGPUPassConfig::addInstSelector() {
bool AMDGPUPassConfig::addPreRegAlloc() {
const AMDILSubtarget &ST = TM->getSubtarget<AMDILSubtarget>();
if (ST.device()->getGeneration() <= AMDILDeviceInfo::HD6XXX) {
PM->add(createR600LowerInstructionsPass(*TM));
} else {
if (ST.device()->getGeneration() > AMDILDeviceInfo::HD6XXX) {
PM->add(createSIAssignInterpRegsPass(*TM));
}
PM->add(createAMDGPULowerInstructionsPass(*TM));
@ -154,12 +152,8 @@ bool AMDGPUPassConfig::addPreSched2() {
}
bool AMDGPUPassConfig::addPreEmitPass() {
const AMDILSubtarget &ST = TM->getSubtarget<AMDILSubtarget>();
PM->add(createAMDILCFGPreparationPass(*TM));
PM->add(createAMDILCFGStructurizerPass(*TM));
if (ST.device()->getGeneration() == AMDILDeviceInfo::HD7XXX) {
PM->add(createSIPropagateImmReadsPass(*TM));
}
return false;
}

View file

@ -490,43 +490,6 @@ AMDILTargetLowering::LowerMemArgument(
//===----------------------------------------------------------------------===//
// TargetLowering Implementation Help Functions End
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// Instruction generation functions
//===----------------------------------------------------------------------===//
MachineOperand
AMDILTargetLowering::convertToReg(MachineOperand op) const
{
if (op.isReg()) {
return op;
} else if (op.isImm()) {
uint32_t loadReg
= genVReg(op.getParent()->getDesc().OpInfo[0].RegClass);
generateMachineInst(AMDIL::LOADCONST_i32, loadReg)
.addImm(op.getImm());
op.ChangeToRegister(loadReg, false);
} else if (op.isFPImm()) {
uint32_t loadReg
= genVReg(op.getParent()->getDesc().OpInfo[0].RegClass);
generateMachineInst(AMDIL::LOADCONST_f32, loadReg)
.addFPImm(op.getFPImm());
op.ChangeToRegister(loadReg, false);
} else if (op.isMBB()) {
op.ChangeToRegister(0, false);
} else if (op.isFI()) {
op.ChangeToRegister(0, false);
} else if (op.isCPI()) {
op.ChangeToRegister(0, false);
} else if (op.isJTI()) {
op.ChangeToRegister(0, false);
} else if (op.isGlobal()) {
op.ChangeToRegister(0, false);
} else if (op.isSymbol()) {
op.ChangeToRegister(0, false);
}/* else if (op.isMetadata()) {
op.ChangeToRegister(0, false);
}*/
return op;
}
//===----------------------------------------------------------------------===//
// TargetLowering Class Implementation Begins

View file

@ -106,10 +106,3 @@ include "AMDILIntrinsics.td"
// Instructions support
//===--------------------------------------------------------------------===//
include "AMDILInstructions.td"
//===--------------------------------------------------------------------===//
// Instruction Pattern support - This Must be the last include in the file
// as it requires items defined in other files
//===--------------------------------------------------------------------===//
include "AMDILInstrPatterns.td"

View file

@ -1,35 +0,0 @@
//===- AMDILInstrPatterns.td - AMDIL Target ------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//==-----------------------------------------------------------------------===//
//===--------------------------------------------------------------------===//
// This file holds all the custom patterns that are used by the amdil backend
//
//===--------------------------------------------------------------------===//
//===--------------------------------------------------------------------===//
// Custom patterns for conversion operations
//===--------------------------------------------------------------------===////
// float ==> long patterns
// unsigned: f32 -> i64
def FTOUL : Pat<(i64 (fp_to_uint GPRF32:$src)),
(LCREATE (FTOU GPRF32:$src), (LOADCONST_i32 0))>;
// unsigned: i64 -> f32
def ULTOF : Pat<(f32 (uint_to_fp GPRI64:$src)),
(UTOF (LLO GPRI64:$src))>;
// LLVM isn't lowering this correctly, so writing a pattern that
// matches it isntead.
def : Pat<(build_vector (i32 imm:$src)),
(VCREATE_v4i32 (LOADCONST_i32 imm:$src))>;
// Calls:
def : Pat<(IL_call tglobaladdr:$dst),
(CALL tglobaladdr:$dst)>;
def : Pat<(IL_call texternalsym:$dst),
(CALL texternalsym:$dst)>;
def : Pat<(IL_call tconstpool:$dst),
(CALL tconstpool:$dst)>;

View file

@ -10,7 +10,6 @@
// Operations in this file are generic to all data types
// This opcode has custom swizzle pattern encoded in Swizzle Encoder
let isReMaterializable = 1, isAsCheapAsAMove = 1 in {
defm LOADCONST : ILConstant<"mov $dst, $val">;
defm MOVE : UnaryOpMC<IL_OP_MOV, IL_mov>;
defm PHIMOVE : UnaryOpMC<IL_OP_MOV, IL_phimov>;
}

View file

@ -100,47 +100,7 @@ AMDILRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
int SPAdj,
RegScavenger *RS) const
{
assert(SPAdj == 0 && "Unexpected");
MachineInstr &MI = *II;
MachineFunction &MF = *MI.getParent()->getParent();
MachineFrameInfo *MFI = MF.getFrameInfo();
unsigned int y = MI.getNumOperands();
for (unsigned int x = 0; x < y; ++x) {
if (!MI.getOperand(x).isFI()) {
continue;
}
const AMDILInstrInfo * AMDILII =
static_cast<const AMDILInstrInfo *>(TM.getInstrInfo());
bool def = AMDILII->isStoreInst(&MI);
int FrameIndex = MI.getOperand(x).getIndex();
int64_t Offset = MFI->getObjectOffset(FrameIndex);
//int64_t Size = MF.getFrameInfo()->getObjectSize(FrameIndex);
// An optimization is to only use the offsets if the size
// is larger than 4, which means we are storing an array
// instead of just a pointer. If we are size 4 then we can
// just do register copies since we don't need to worry about
// indexing dynamically
MachineInstr *nMI = MF.CreateMachineInstr(
TII.get(AMDIL::LOADCONST_i32), MI.getDebugLoc());
nMI->addOperand(MachineOperand::CreateReg(AMDIL::DFP, true));
nMI->addOperand(
MachineOperand::CreateImm(Offset));
MI.getParent()->insert(II, nMI);
nMI = MF.CreateMachineInstr(
TII.get(AMDIL::ADD_INT), MI.getDebugLoc());
nMI->addOperand(MachineOperand::CreateReg(AMDIL::DFP, true));
nMI->addOperand(MachineOperand::CreateReg(AMDIL::DFP, false));
nMI->addOperand(MachineOperand::CreateReg(AMDIL::FP, false));
MI.getParent()->insert(II, nMI);
if (MI.getOperand(x).isReg() == false) {
MI.getOperand(x).ChangeToRegister(
nMI->getOperand(0).getReg(), def);
} else {
MI.getOperand(x).setReg(
nMI->getOperand(0).getReg());
}
}
assert(!"Implement");
}
void

View file

@ -41,7 +41,6 @@ CPP_SOURCES := \
R600ISelLowering.cpp \
R600InstrInfo.cpp \
R600KernelParameters.cpp \
R600LowerInstructions.cpp \
R600MachineFunctionInfo.cpp \
R600RegisterInfo.cpp \
SIAssignInterpRegs.cpp \
@ -49,7 +48,6 @@ CPP_SOURCES := \
SIInstrInfo.cpp \
SIISelLowering.cpp \
SIMachineFunctionInfo.cpp \
SIPropagateImmReads.cpp \
SIRegisterInfo.cpp \
MCTargetDesc/AMDILMCAsmInfo.cpp \
MCTargetDesc/AMDILMCTargetDesc.cpp \

View file

@ -371,6 +371,25 @@ def FLOOR : R600_1OP <
def MOV : R600_1OP <0x19, "MOV", []>;
class MOV_IMM <ValueType vt, Operand immType> : InstR600 <0x19,
(outs R600_Reg32:$dst),
(ins R600_Reg32:$alu_literal, immType:$imm),
"MOV_IMM $dst, $imm",
[], AnyALU
>;
def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
def : Pat <
(imm:$val),
(MOV_IMM_I32 (i32 ALU_LITERAL_X), imm:$val)
>;
def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
def : Pat <
(fpimm:$val),
(MOV_IMM_F32 (i32 ALU_LITERAL_X), fpimm:$val)
>;
def KILLGT : R600_2OP <
0x2D, "KILLGT",
[]
@ -821,7 +840,7 @@ let Predicates = [isEGorCayman] in {
class TRIG_eg <InstR600 trig, Intrinsic intr> : Pat<
(intr R600_Reg32:$src),
(trig (MUL (MOV (LOADCONST_i32 CONST.TWO_PI_INV)), R600_Reg32:$src))
(trig (MUL (MOV_IMM_I32 (i32 ALU_LITERAL_X), CONST.TWO_PI_INV), R600_Reg32:$src))
>;
def MULADD_eg : MULADD_Common<0x14>;

View file

@ -1,106 +0,0 @@
//===-- R600LowerInstructions.cpp - Lower unsupported AMDIL instructions --===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass lowers AMDIL MachineInstrs that aren't supported by the R600
// target to either supported AMDIL MachineInstrs or R600 MachineInstrs.
//
//===----------------------------------------------------------------------===//
#include "AMDGPU.h"
#include "AMDGPUInstrInfo.h"
#include "AMDGPUUtil.h"
#include "AMDIL.h"
#include "AMDILRegisterInfo.h"
#include "R600InstrInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/Constants.h"
#include "llvm/Target/TargetInstrInfo.h"
#include <stdio.h>
using namespace llvm;
namespace {
class R600LowerInstructionsPass : public MachineFunctionPass {
private:
static char ID;
const R600InstrInfo * TII;
public:
R600LowerInstructionsPass(TargetMachine &tm) :
MachineFunctionPass(ID),
TII(static_cast<const R600InstrInfo*>(tm.getInstrInfo()))
{ }
const char *getPassName() const { return "R600 Lower Instructions"; }
virtual bool runOnMachineFunction(MachineFunction &MF);
};
} /* End anonymous namespace */
char R600LowerInstructionsPass::ID = 0;
FunctionPass *llvm::createR600LowerInstructionsPass(TargetMachine &tm) {
return new R600LowerInstructionsPass(tm);
}
bool R600LowerInstructionsPass::runOnMachineFunction(MachineFunction &MF)
{
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) ) {
MachineInstr &MI = *I;
switch(MI.getOpcode()) {
case AMDIL::LOADCONST_f32:
case AMDIL::LOADCONST_i32:
{
bool canInline = false;
unsigned inlineReg;
MachineOperand & dstOp = MI.getOperand(0);
MachineOperand & immOp = MI.getOperand(1);
if (immOp.isFPImm()) {
const ConstantFP * cfp = immOp.getFPImm();
if (cfp->isZero()) {
canInline = true;
inlineReg = AMDIL::ZERO;
} else if (cfp->isExactlyValue(1.0f)) {
canInline = true;
inlineReg = AMDIL::ONE;
} else if (cfp->isExactlyValue(0.5f)) {
canInline = true;
inlineReg = AMDIL::HALF;
}
}
if (canInline) {
BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDIL::COPY))
.addOperand(dstOp)
.addReg(inlineReg);
} else {
BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDIL::MOV))
.addOperand(dstOp)
.addReg(AMDIL::ALU_LITERAL_X)
.addOperand(immOp);
}
break;
}
default:
continue;
}
MI.eraseFromParent();
}
}
return false;
}

View file

@ -109,7 +109,7 @@ unsigned SIInstrInfo::getISAOpcode(unsigned AMDILopcode) const
MachineInstr * SIInstrInfo::getMovImmInstr(MachineFunction *MF, unsigned DstReg,
int64_t Imm) const
{
MachineInstr * MI = MF->CreateMachineInstr(get(AMDIL::V_MOV_IMM), DebugLoc());
MachineInstr * MI = MF->CreateMachineInstr(get(AMDIL::V_MOV_IMM_I32), DebugLoc());
MachineInstrBuilder(MI).addReg(DstReg, RegState::Define);
MachineInstrBuilder(MI).addImm(Imm);

View file

@ -773,22 +773,25 @@ def S_BFE_I64 : SOP2_64 <0x0000002a, "S_BFE_I64", []>;
//def S_CBRANCH_G_FORK : SOP2_ <0x0000002b, "S_CBRANCH_G_FORK", []>;
def S_ABSDIFF_I32 : SOP2_32 <0x0000002c, "S_ABSDIFF_I32", []>;
def V_MOV_IMM : VOP1 <
class V_MOV_IMM <Operand immType, SDNode immNode> : VOP1 <
0x1,
(outs VReg_32:$dst),
(ins f32imm:$src0),
(ins immType:$src0),
"V_MOV_IMM",
[]
[(set VReg_32:$dst, (immNode:$src0))]
>;
def V_MOV_IMM_I32 : V_MOV_IMM<i32imm, imm>;
def V_MOV_IMM_F32 : V_MOV_IMM<f32imm, fpimm>;
def S_MOV_IMM_I32 : SOP1 <
0x3,
(outs SReg_32:$dst),
(ins i32Literal:$src0),
"S_MOV_IMM",
[] > {
let neverHasSideEffects = 1;
}
"S_MOV_IMM_I32",
[(set SReg_32:$dst, (imm:$src0))]
>;
let isCodeGenOnly = 1, isPseudo = 1 in {

View file

@ -1,75 +0,0 @@
//===-- SIPropagateImmReads.cpp - Lower Immediate Reads Pass --------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// We can't do this in the ConvertToISA pass, because later passes might
// create LOADCONST_* instructions that we would miss. This is why we need
// a separate pass for this.
//
//===----------------------------------------------------------------------===//
#include "AMDGPU.h"
#include "AMDGPUUtil.h"
#include "SIInstrInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
using namespace llvm;
namespace {
class SIPropagateImmReadsPass : public MachineFunctionPass {
private:
static char ID;
TargetMachine &TM;
public:
SIPropagateImmReadsPass(TargetMachine &tm) :
MachineFunctionPass(ID), TM(tm) { }
virtual bool runOnMachineFunction(MachineFunction &MF);
};
} // End anonymous namespace
char SIPropagateImmReadsPass::ID = 0;
FunctionPass *llvm::createSIPropagateImmReadsPass(TargetMachine &tm) {
return new SIPropagateImmReadsPass(tm);
}
bool SIPropagateImmReadsPass::runOnMachineFunction(MachineFunction &MF)
{
const SIInstrInfo * TII = static_cast<const SIInstrInfo*>(TM.getInstrInfo());
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)) {
MachineInstr &MI = *I;
switch (MI.getOpcode()) {
case AMDIL::LOADCONST_f32:
case AMDIL::LOADCONST_i32:
case AMDIL::LOADCONST_i64:
break;
default:
continue;
}
// XXX: Create and use S_MOV_IMM for SREGs
BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDIL::V_MOV_IMM))
.addOperand(MI.getOperand(0))
.addOperand(MI.getOperand(1));
MI.eraseFromParent();
}
}
return false;
}