llvm-project/llvm/lib/Target/PowerPC/PPCCTRLoops.cpp
Chen Zheng df9d60af1f [PowerPC] handle more than two predecessors loop header in ctrloop pass
After ISEL, the "valid" loop header which has two predecessors
(one is preheader and the other one is latch) may be transformed
to have more than two predecessors by some optimizations, like tail
duplicator, if the old header's successor(will be changed to new
header) is a sub loop.

The predecessors of the new loop header are preheader, loop latch
and the loop latch(es) of the sub loop(old header's successor).

Before the patch, ctrloop pass assumes two predecessors for candidate
loop header. This patch fixes this case.

Reviewed By: lkail

Differential Revision: https://reviews.llvm.org/D135846
2022-10-19 01:11:58 +00:00

379 lines
13 KiB
C++

//===-- PPCCTRLoops.cpp - Generate CTR loops ------------------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
// This pass generates machine instructions for the CTR loops related pseudos:
// 1: MTCTRloop/DecreaseCTRloop
// 2: MTCTR8loop/DecreaseCTR8loop
//
// If a CTR loop can be generated:
// 1: MTCTRloop/MTCTR8loop will be converted to "mtctr"
// 2: DecreaseCTRloop/DecreaseCTR8loop will be converted to "bdnz/bdz" and
// its user branch instruction can be deleted.
//
// If a CTR loop can not be generated due to clobber of CTR:
// 1: MTCTRloop/MTCTR8loop can be deleted.
// 2: DecreaseCTRloop/DecreaseCTR8loop will be converted to "addi -1" and
// a "cmplwi/cmpldi".
//
// This pass runs just before register allocation, because we don't want
// register allocator to allocate register for DecreaseCTRloop if a CTR can be
// generated or if a CTR loop can not be generated, we don't have any condition
// register for the new added "cmplwi/cmpldi".
//
//===----------------------------------------------------------------------===//
#include "PPC.h"
#include "PPCInstrInfo.h"
#include "PPCSubtarget.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/Register.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/PassRegistry.h"
#include "llvm/Support/CodeGen.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include <cassert>
using namespace llvm;
#define DEBUG_TYPE "ppc-ctrloops"
STATISTIC(NumCTRLoops, "Number of CTR loops generated");
STATISTIC(NumNormalLoops, "Number of normal compare + branch loops generated");
namespace {
class PPCCTRLoops : public MachineFunctionPass {
public:
static char ID;
PPCCTRLoops() : MachineFunctionPass(ID) {
initializePPCCTRLoopsPass(*PassRegistry::getPassRegistry());
}
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.addRequired<MachineLoopInfo>();
MachineFunctionPass::getAnalysisUsage(AU);
}
bool runOnMachineFunction(MachineFunction &MF) override;
private:
const PPCInstrInfo *TII = nullptr;
MachineRegisterInfo *MRI = nullptr;
bool processLoop(MachineLoop *ML);
bool isCTRClobber(MachineInstr *MI, bool CheckReads) const;
void expandNormalLoops(MachineLoop *ML, MachineInstr *Start,
MachineInstr *Dec);
void expandCTRLoops(MachineLoop *ML, MachineInstr *Start, MachineInstr *Dec);
};
} // namespace
char PPCCTRLoops::ID = 0;
INITIALIZE_PASS_BEGIN(PPCCTRLoops, DEBUG_TYPE, "PowerPC CTR loops generation",
false, false)
INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
INITIALIZE_PASS_END(PPCCTRLoops, DEBUG_TYPE, "PowerPC CTR loops generation",
false, false)
FunctionPass *llvm::createPPCCTRLoopsPass() { return new PPCCTRLoops(); }
bool PPCCTRLoops::runOnMachineFunction(MachineFunction &MF) {
bool Changed = false;
auto &MLI = getAnalysis<MachineLoopInfo>();
TII = static_cast<const PPCInstrInfo *>(MF.getSubtarget().getInstrInfo());
MRI = &MF.getRegInfo();
for (auto *ML : MLI) {
if (ML->isOutermost())
Changed |= processLoop(ML);
}
return Changed;
}
bool PPCCTRLoops::isCTRClobber(MachineInstr *MI, bool CheckReads) const {
if (!CheckReads) {
// If we are only checking for defs, that is we are going to find
// definitions before MTCTRloop, for this case:
// CTR defination inside the callee of a call instruction will not impact
// the defination of MTCTRloop, so we can use definesRegister() for the
// check, no need to check the regmask.
return (MI->definesRegister(PPC::CTR) &&
!MI->registerDefIsDead(PPC::CTR)) ||
(MI->definesRegister(PPC::CTR8) &&
!MI->registerDefIsDead(PPC::CTR8));
}
if ((MI->modifiesRegister(PPC::CTR) && !MI->registerDefIsDead(PPC::CTR)) ||
(MI->modifiesRegister(PPC::CTR8) && !MI->registerDefIsDead(PPC::CTR8)))
return true;
if (MI->getDesc().isCall())
return true;
// We define the CTR in the loop preheader, so if there is any CTR reader in
// the loop, we also can not use CTR loop form.
if (MI->readsRegister(PPC::CTR) || MI->readsRegister(PPC::CTR8))
return true;
return false;
}
bool PPCCTRLoops::processLoop(MachineLoop *ML) {
bool Changed = false;
// Align with HardwareLoop pass, process inner loops first.
for (MachineLoop *I : *ML)
Changed |= processLoop(I);
// If any inner loop is changed, outter loop must be without hardware loop
// intrinsics.
if (Changed)
return true;
auto IsLoopStart = [](MachineInstr &MI) {
return MI.getOpcode() == PPC::MTCTRloop ||
MI.getOpcode() == PPC::MTCTR8loop;
};
auto SearchForStart =
[&IsLoopStart](MachineBasicBlock *MBB) -> MachineInstr * {
for (auto &MI : *MBB) {
if (IsLoopStart(MI))
return &MI;
}
return nullptr;
};
MachineInstr *Start = nullptr;
MachineInstr *Dec = nullptr;
bool InvalidCTRLoop = false;
MachineBasicBlock *Preheader = ML->getLoopPreheader();
// If there is no preheader for this loop, there must be no MTCTRloop
// either.
if (!Preheader)
return false;
Start = SearchForStart(Preheader);
// This is not a CTR loop candidate.
if (!Start)
return false;
// If CTR is live to the preheader, we can not redefine the CTR register.
if (Preheader->isLiveIn(PPC::CTR) || Preheader->isLiveIn(PPC::CTR8))
InvalidCTRLoop = true;
// Make sure there is also no CTR clobber in the block preheader between the
// begin and MTCTR.
for (MachineBasicBlock::reverse_instr_iterator I =
std::next(Start->getReverseIterator());
I != Preheader->instr_rend(); ++I)
// Only check the definitions of CTR. If there is non-dead definition for
// the CTR, we conservatively don't generate a CTR loop.
if (isCTRClobber(&*I, /* CheckReads */ false)) {
InvalidCTRLoop = true;
break;
}
// Make sure there is also no CTR clobber/user in the block preheader between
// MTCTR and the end.
for (MachineBasicBlock::instr_iterator I = std::next(Start->getIterator());
I != Preheader->instr_end(); ++I)
if (isCTRClobber(&*I, /* CheckReads */ true)) {
InvalidCTRLoop = true;
break;
}
// Find the CTR loop components and decide whether or not to fall back to a
// normal loop.
for (auto *MBB : reverse(ML->getBlocks())) {
for (auto &MI : *MBB) {
if (MI.getOpcode() == PPC::DecreaseCTRloop ||
MI.getOpcode() == PPC::DecreaseCTR8loop)
Dec = &MI;
else if (!InvalidCTRLoop)
// If any instruction clobber CTR, then we can not generate a CTR loop.
InvalidCTRLoop |= isCTRClobber(&MI, /* CheckReads */ true);
}
if (Dec && InvalidCTRLoop)
break;
}
assert(Dec && "CTR loop is not complete!");
if (InvalidCTRLoop) {
expandNormalLoops(ML, Start, Dec);
++NumNormalLoops;
}
else {
expandCTRLoops(ML, Start, Dec);
++NumCTRLoops;
}
return true;
}
void PPCCTRLoops::expandNormalLoops(MachineLoop *ML, MachineInstr *Start,
MachineInstr *Dec) {
bool Is64Bit =
Start->getParent()->getParent()->getSubtarget<PPCSubtarget>().isPPC64();
MachineBasicBlock *Preheader = Start->getParent();
MachineBasicBlock *Exiting = Dec->getParent();
assert((Preheader && Exiting) &&
"Preheader and exiting should exist for CTR loop!");
assert(Dec->getOperand(1).getImm() == 1 &&
"Loop decrement stride must be 1");
unsigned ADDIOpcode = Is64Bit ? PPC::ADDI8 : PPC::ADDI;
unsigned CMPOpcode = Is64Bit ? PPC::CMPLDI : PPC::CMPLWI;
Register PHIDef =
MRI->createVirtualRegister(Is64Bit ? &PPC::G8RC_and_G8RC_NOX0RegClass
: &PPC::GPRC_and_GPRC_NOR0RegClass);
Start->getParent()->getParent()->getProperties().reset(
MachineFunctionProperties::Property::NoPHIs);
// Generate "PHI" in the header block.
auto PHIMIB = BuildMI(*ML->getHeader(), ML->getHeader()->getFirstNonPHI(),
DebugLoc(), TII->get(TargetOpcode::PHI), PHIDef);
PHIMIB.addReg(Start->getOperand(0).getReg()).addMBB(Preheader);
Register ADDIDef =
MRI->createVirtualRegister(Is64Bit ? &PPC::G8RC_and_G8RC_NOX0RegClass
: &PPC::GPRC_and_GPRC_NOR0RegClass);
// Generate "addi -1" in the exiting block.
BuildMI(*Exiting, Dec, Dec->getDebugLoc(), TII->get(ADDIOpcode), ADDIDef)
.addReg(PHIDef)
.addImm(-1);
// Add other inputs for the PHI node.
if (ML->isLoopLatch(Exiting)) {
// Normally there must be only two predecessors for the loop header, one is
// the Preheader and the other one is loop latch Exiting. In hardware loop
// insertion pass, the block containing DecreaseCTRloop must dominate all
// loop latches. So there must be only one latch.
// But there are some optimizations after ISEL, like tail duplicator, may
// merge the two-predecessor loop header with its successor. If the
// successor happens to be a header of nest loop, then we will have a header
// which has more than 2 predecessors.
assert(std::find(ML->getHeader()->predecessors().begin(),
ML->getHeader()->predecessors().end(),
Exiting) != ML->getHeader()->predecessors().end() &&
"Loop latch is not loop header predecessor!");
assert(std::find(ML->getHeader()->predecessors().begin(),
ML->getHeader()->predecessors().end(),
Preheader) != ML->getHeader()->predecessors().end() &&
"Loop preheader is not loop header predecessor!");
PHIMIB.addReg(ADDIDef).addMBB(Exiting);
if (ML->getHeader()->pred_size() > 2) {
Register HeaderIncoming = MRI->createVirtualRegister(
Is64Bit ? &PPC::G8RC_and_G8RC_NOX0RegClass
: &PPC::GPRC_and_GPRC_NOR0RegClass);
BuildMI(*ML->getHeader(), ML->getHeader()->getFirstNonPHI(), DebugLoc(),
TII->get(TargetOpcode::COPY), HeaderIncoming)
.addReg(PHIDef);
for (MachineBasicBlock *P : ML->getHeader()->predecessors()) {
if (P != Preheader && P != Exiting)
PHIMIB.addReg(HeaderIncoming).addMBB(P);
}
}
} else {
// If the block containing DecreaseCTRloop is not a loop latch, we can use
// ADDIDef as the value for all other blocks for the PHI. In hardware loop
// insertion pass, the block containing DecreaseCTRloop must dominate all
// loop latches.
for (MachineBasicBlock *P : ML->getHeader()->predecessors()) {
if (ML->contains(P)) {
assert(ML->isLoopLatch(P) &&
"Loop's header in-loop predecessor is not loop latch!");
PHIMIB.addReg(ADDIDef).addMBB(P);
} else
assert(P == Preheader &&
"CTR loop should not be generated for irreducible loop!");
}
}
// Generate the compare in the exiting block.
Register CMPDef = MRI->createVirtualRegister(&PPC::CRRCRegClass);
auto CMPMIB =
BuildMI(*Exiting, Dec, Dec->getDebugLoc(), TII->get(CMPOpcode), CMPDef)
.addReg(ADDIDef)
.addImm(0);
BuildMI(*Exiting, Dec, Dec->getDebugLoc(), TII->get(TargetOpcode::COPY),
Dec->getOperand(0).getReg())
.addReg(CMPMIB->getOperand(0).getReg(), 0, PPC::sub_gt);
// Remove the pseudo instructions.
Start->eraseFromParent();
Dec->eraseFromParent();
}
void PPCCTRLoops::expandCTRLoops(MachineLoop *ML, MachineInstr *Start,
MachineInstr *Dec) {
bool Is64Bit =
Start->getParent()->getParent()->getSubtarget<PPCSubtarget>().isPPC64();
MachineBasicBlock *Preheader = Start->getParent();
MachineBasicBlock *Exiting = Dec->getParent();
(void)Preheader;
assert((Preheader && Exiting) &&
"Preheader and exiting should exist for CTR loop!");
assert(Dec->getOperand(1).getImm() == 1 && "Loop decrement must be 1!");
unsigned BDNZOpcode = Is64Bit ? PPC::BDNZ8 : PPC::BDNZ;
unsigned BDZOpcode = Is64Bit ? PPC::BDZ8 : PPC::BDZ;
auto BrInstr = MRI->use_instr_begin(Dec->getOperand(0).getReg());
assert(MRI->hasOneUse(Dec->getOperand(0).getReg()) &&
"There should be only one user for loop decrement pseudo!");
unsigned Opcode = 0;
switch (BrInstr->getOpcode()) {
case PPC::BC:
Opcode = BDNZOpcode;
(void) ML;
assert(ML->contains(BrInstr->getOperand(1).getMBB()) &&
"Invalid ctr loop!");
break;
case PPC::BCn:
Opcode = BDZOpcode;
assert(!ML->contains(BrInstr->getOperand(1).getMBB()) &&
"Invalid ctr loop!");
break;
default:
llvm_unreachable("Unhandled branch user for DecreaseCTRloop.");
}
// Generate "bdnz/bdz" in the exiting block just before the terminator.
BuildMI(*Exiting, &*BrInstr, BrInstr->getDebugLoc(), TII->get(Opcode))
.addMBB(BrInstr->getOperand(1).getMBB());
// Remove the pseudo instructions.
BrInstr->eraseFromParent();
Dec->eraseFromParent();
}