llvm-project/llvm/lib/Target/AMDGPU/SIAddIMGInit.cpp

181 lines
6.4 KiB
C++
Raw Normal View History

[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try TFE and LWE support requires extra result registers that are written in the event of a failure in order to detect that failure case. The specific use-case that initiated these changes is sparse texture support. This means that if image intrinsics are used with either option turned on, the programmer must ensure that the return type can contain all of the expected results. This can result in redundant registers since the vector size must be a power-of-2. This change takes roughly 6 parts: 1. Modify the instruction defs in tablegen to add new instruction variants that can accomodate the extra return values. 2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE (where the bulk of the work for these instruction types is now done) 3. Extra verification code to catch cases where intrinsics have been used but insufficient return registers are used. 4. Modification to the adjustWritemask optimisation to account for TFE/LWE being enabled (requires extra registers to be maintained for error return value). 5. An extra pass to zero initialize the error value return - this is because if the error does not occur, the register is not written and thus must be zeroed before use. Also added a new (on by default) option to ensure ALL return values are zero-initialized that is required for sparse texture support. 6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO for this to re-enable and handle correctly). There's an additional fix now to avoid a dmask=0 For an image intrinsic with tfe where all result channels except tfe were unused, I was getting an image instruction with dmask=0 and only a single vgpr result for tfe. That is incorrect because the hardware assumes there is at least one vgpr result, plus the one for tfe. Fixed by forcing dmask to 1, which gives the desired two vgpr result with tfe in the second one. The TFE or LWE result is returned from the intrinsics using an aggregate type. Look in the test code provided to see how this works, but in essence IR code to invoke the intrinsic looks as follows: %v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15, i32 %s, <8 x i32> %rsrc, i32 1, i32 0) %v.vec = extractvalue {<4 x float>, i32} %v, 0 %v.err = extractvalue {<4 x float>, i32} %v, 1 This re-submit of the change also includes a slight modification in SIISelLowering.cpp to work-around a compiler bug for the powerpc_le platform that caused a buildbot failure on a previous submission. Differential revision: https://reviews.llvm.org/D48826 Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda Work around for ppcle compiler bug Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b llvm-svn: 351054
2019-01-14 19:55:24 +08:00
//===-- SIAddIMGInit.cpp - Add any required IMG inits ---------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try TFE and LWE support requires extra result registers that are written in the event of a failure in order to detect that failure case. The specific use-case that initiated these changes is sparse texture support. This means that if image intrinsics are used with either option turned on, the programmer must ensure that the return type can contain all of the expected results. This can result in redundant registers since the vector size must be a power-of-2. This change takes roughly 6 parts: 1. Modify the instruction defs in tablegen to add new instruction variants that can accomodate the extra return values. 2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE (where the bulk of the work for these instruction types is now done) 3. Extra verification code to catch cases where intrinsics have been used but insufficient return registers are used. 4. Modification to the adjustWritemask optimisation to account for TFE/LWE being enabled (requires extra registers to be maintained for error return value). 5. An extra pass to zero initialize the error value return - this is because if the error does not occur, the register is not written and thus must be zeroed before use. Also added a new (on by default) option to ensure ALL return values are zero-initialized that is required for sparse texture support. 6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO for this to re-enable and handle correctly). There's an additional fix now to avoid a dmask=0 For an image intrinsic with tfe where all result channels except tfe were unused, I was getting an image instruction with dmask=0 and only a single vgpr result for tfe. That is incorrect because the hardware assumes there is at least one vgpr result, plus the one for tfe. Fixed by forcing dmask to 1, which gives the desired two vgpr result with tfe in the second one. The TFE or LWE result is returned from the intrinsics using an aggregate type. Look in the test code provided to see how this works, but in essence IR code to invoke the intrinsic looks as follows: %v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15, i32 %s, <8 x i32> %rsrc, i32 1, i32 0) %v.vec = extractvalue {<4 x float>, i32} %v, 0 %v.err = extractvalue {<4 x float>, i32} %v, 1 This re-submit of the change also includes a slight modification in SIISelLowering.cpp to work-around a compiler bug for the powerpc_le platform that caused a buildbot failure on a previous submission. Differential revision: https://reviews.llvm.org/D48826 Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda Work around for ppcle compiler bug Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b llvm-svn: 351054
2019-01-14 19:55:24 +08:00
//
//===----------------------------------------------------------------------===//
//
/// \file
/// Any MIMG instructions that use tfe or lwe require an initialization of the
/// result register that will be written in the case of a memory access failure
/// The required code is also added to tie this init code to the result of the
/// img instruction
///
//===----------------------------------------------------------------------===//
//
#include "AMDGPU.h"
#include "AMDGPUSubtarget.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "SIInstrInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/Function.h"
#include "llvm/Support/Debug.h"
#include "llvm/Target/TargetMachine.h"
#define DEBUG_TYPE "si-img-init"
using namespace llvm;
namespace {
class SIAddIMGInit : public MachineFunctionPass {
public:
static char ID;
public:
SIAddIMGInit() : MachineFunctionPass(ID) {
initializeSIAddIMGInitPass(*PassRegistry::getPassRegistry());
}
bool runOnMachineFunction(MachineFunction &MF) override;
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
MachineFunctionPass::getAnalysisUsage(AU);
}
};
} // End anonymous namespace.
INITIALIZE_PASS(SIAddIMGInit, DEBUG_TYPE, "SI Add IMG Init", false, false)
char SIAddIMGInit::ID = 0;
char &llvm::SIAddIMGInitID = SIAddIMGInit::ID;
FunctionPass *llvm::createSIAddIMGInitPass() { return new SIAddIMGInit(); }
bool SIAddIMGInit::runOnMachineFunction(MachineFunction &MF) {
MachineRegisterInfo &MRI = MF.getRegInfo();
const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
const SIInstrInfo *TII = ST.getInstrInfo();
const SIRegisterInfo *RI = ST.getRegisterInfo();
bool Changed = false;
for (MachineFunction::iterator BI = MF.begin(), BE = MF.end(); BI != BE;
++BI) {
MachineBasicBlock &MBB = *BI;
MachineBasicBlock::iterator I, Next;
for (I = MBB.begin(); I != MBB.end(); I = Next) {
Next = std::next(I);
MachineInstr &MI = *I;
auto Opcode = MI.getOpcode();
if (TII->isMIMG(Opcode) && !MI.mayStore()) {
MachineOperand *TFE = TII->getNamedOperand(MI, AMDGPU::OpName::tfe);
MachineOperand *LWE = TII->getNamedOperand(MI, AMDGPU::OpName::lwe);
MachineOperand *D16 = TII->getNamedOperand(MI, AMDGPU::OpName::d16);
// Check for instructions that don't have tfe or lwe fields
// There shouldn't be any at this point.
assert( (TFE && LWE) && "Expected tfe and lwe operands in instruction");
unsigned TFEVal = TFE->getImm();
unsigned LWEVal = LWE->getImm();
unsigned D16Val = D16 ? D16->getImm() : 0;
if (TFEVal || LWEVal) {
// At least one of TFE or LWE are non-zero
// We have to insert a suitable initialization of the result value and
// tie this to the dest of the image instruction.
const DebugLoc &DL = MI.getDebugLoc();
int DstIdx =
AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vdata);
// Calculate which dword we have to initialize to 0.
MachineOperand *MO_Dmask =
TII->getNamedOperand(MI, AMDGPU::OpName::dmask);
// check that dmask operand is found.
assert(MO_Dmask && "Expected dmask operand in instruction");
unsigned dmask = MO_Dmask->getImm();
// Determine the number of active lanes taking into account the
// Gather4 special case
unsigned ActiveLanes =
TII->isGather4(Opcode) ? 4 : countPopulation(dmask);
// Subreg indices are counted from 1
// When D16 then we want next whole VGPR after write data.
static_assert(AMDGPU::sub0 == 1 && AMDGPU::sub4 == 5, "Subreg indices different from expected");
bool Packed = !ST.hasUnpackedD16VMem();
unsigned InitIdx =
D16Val && Packed ? ((ActiveLanes + 1) >> 1) + 1 : ActiveLanes + 1;
// Abandon attempt if the dst size isn't large enough
// - this is in fact an error but this is picked up elsewhere and
// reported correctly.
uint32_t DstSize =
RI->getRegSizeInBits(*TII->getOpRegClass(MI, DstIdx)) / 32;
if (DstSize < InitIdx)
continue;
// Create a register for the intialization value.
unsigned PrevDst =
MRI.createVirtualRegister(TII->getOpRegClass(MI, DstIdx));
unsigned NewDst = 0; // Final initialized value will be in here
// If PRTStrictNull feature is enabled (the default) then initialize
// all the result registers to 0, otherwise just the error indication
// register (VGPRn+1)
unsigned SizeLeft = ST.usePRTStrictNull() ? InitIdx : 1;
unsigned CurrIdx = ST.usePRTStrictNull() ? 1 : InitIdx;
if (DstSize == 1) {
// In this case we can just initialize the result directly
BuildMI(MBB, MI, DL, TII->get(AMDGPU::V_MOV_B32_e32), PrevDst)
.addImm(0);
NewDst = PrevDst;
} else {
BuildMI(MBB, MI, DL, TII->get(AMDGPU::IMPLICIT_DEF), PrevDst);
for (; SizeLeft; SizeLeft--, CurrIdx++) {
NewDst =
MRI.createVirtualRegister(TII->getOpRegClass(MI, DstIdx));
// Initialize dword
unsigned SubReg =
MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
BuildMI(MBB, MI, DL, TII->get(AMDGPU::V_MOV_B32_e32), SubReg)
.addImm(0);
// Insert into the super-reg
BuildMI(MBB, I, DL, TII->get(TargetOpcode::INSERT_SUBREG), NewDst)
.addReg(PrevDst)
.addReg(SubReg)
.addImm(CurrIdx);
PrevDst = NewDst;
}
}
// Add as an implicit operand
MachineInstrBuilder(MF, MI).addReg(NewDst, RegState::Implicit);
// Tie the just added implicit operand to the dst
MI.tieOperands(DstIdx, MI.getNumOperands() - 1);
Changed = true;
}
}
}
}
return Changed;
}