1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-19 11:02:59 +02:00
llvm-mirror/lib/Target/X86/X86VZeroUpper.cpp
Sanjay Patel b2763427c2 [x86, AVX] don't add a vzeroupper if that's what the code is already doing (PR27823)
This isn't the complete fix, but it handles the trivial examples of duplicate vzero* ops in PR27823:
https://llvm.org/bugs/show_bug.cgi?id=27823
...and amusingly, the bogus cases already exist as regression tests, so let's take this baby step.

We'll need to do more in the general case where there's legitimate AVX usage in the function + there's
already a vzero in the code.

Differential Revision: http://reviews.llvm.org/D20477

llvm-svn: 270378
2016-05-22 20:22:47 +00:00

332 lines
12 KiB
C++

//===-- X86VZeroUpper.cpp - AVX vzeroupper instruction inserter -----------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines the pass which inserts x86 AVX vzeroupper instructions
// before calls to SSE encoded functions. This avoids transition latency
// penalty when transferring control between AVX encoded instructions and old
// SSE encoding mode.
//
//===----------------------------------------------------------------------===//
#include "X86.h"
#include "X86InstrInfo.h"
#include "X86Subtarget.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetInstrInfo.h"
using namespace llvm;
#define DEBUG_TYPE "x86-vzeroupper"
STATISTIC(NumVZU, "Number of vzeroupper instructions inserted");
namespace {
class VZeroUpperInserter : public MachineFunctionPass {
public:
VZeroUpperInserter() : MachineFunctionPass(ID) {}
bool runOnMachineFunction(MachineFunction &MF) override;
MachineFunctionProperties getRequiredProperties() const override {
return MachineFunctionProperties().set(
MachineFunctionProperties::Property::AllVRegsAllocated);
}
const char *getPassName() const override {return "X86 vzeroupper inserter";}
private:
void processBasicBlock(MachineBasicBlock &MBB);
void insertVZeroUpper(MachineBasicBlock::iterator I,
MachineBasicBlock &MBB);
void addDirtySuccessor(MachineBasicBlock &MBB);
typedef enum { PASS_THROUGH, EXITS_CLEAN, EXITS_DIRTY } BlockExitState;
static const char* getBlockExitStateName(BlockExitState ST);
// Core algorithm state:
// BlockState - Each block is either:
// - PASS_THROUGH: There are neither YMM dirtying instructions nor
// vzeroupper instructions in this block.
// - EXITS_CLEAN: There is (or will be) a vzeroupper instruction in this
// block that will ensure that YMM is clean on exit.
// - EXITS_DIRTY: An instruction in the block dirties YMM and no
// subsequent vzeroupper in the block clears it.
//
// AddedToDirtySuccessors - This flag is raised when a block is added to the
// DirtySuccessors list to ensure that it's not
// added multiple times.
//
// FirstUnguardedCall - Records the location of the first unguarded call in
// each basic block that may need to be guarded by a
// vzeroupper. We won't know whether it actually needs
// to be guarded until we discover a predecessor that
// is DIRTY_OUT.
struct BlockState {
BlockState() : ExitState(PASS_THROUGH), AddedToDirtySuccessors(false) {}
BlockExitState ExitState;
bool AddedToDirtySuccessors;
MachineBasicBlock::iterator FirstUnguardedCall;
};
typedef SmallVector<BlockState, 8> BlockStateMap;
typedef SmallVector<MachineBasicBlock*, 8> DirtySuccessorsWorkList;
BlockStateMap BlockStates;
DirtySuccessorsWorkList DirtySuccessors;
bool EverMadeChange;
bool IsX86INTR;
const TargetInstrInfo *TII;
static char ID;
};
char VZeroUpperInserter::ID = 0;
}
FunctionPass *llvm::createX86IssueVZeroUpperPass() {
return new VZeroUpperInserter();
}
const char* VZeroUpperInserter::getBlockExitStateName(BlockExitState ST) {
switch (ST) {
case PASS_THROUGH: return "Pass-through";
case EXITS_DIRTY: return "Exits-dirty";
case EXITS_CLEAN: return "Exits-clean";
}
llvm_unreachable("Invalid block exit state.");
}
static bool isYmmReg(unsigned Reg) {
return (Reg >= X86::YMM0 && Reg <= X86::YMM15);
}
static bool checkFnHasLiveInYmm(MachineRegisterInfo &MRI) {
for (MachineRegisterInfo::livein_iterator I = MRI.livein_begin(),
E = MRI.livein_end(); I != E; ++I)
if (isYmmReg(I->first))
return true;
return false;
}
static bool clobbersAllYmmRegs(const MachineOperand &MO) {
for (unsigned reg = X86::YMM0; reg <= X86::YMM15; ++reg) {
if (!MO.clobbersPhysReg(reg))
return false;
}
return true;
}
static bool hasYmmReg(MachineInstr *MI) {
for (const MachineOperand &MO : MI->operands()) {
if (MI->isCall() && MO.isRegMask() && !clobbersAllYmmRegs(MO))
return true;
if (!MO.isReg())
continue;
if (MO.isDebug())
continue;
if (isYmmReg(MO.getReg()))
return true;
}
return false;
}
/// Check if any YMM register will be clobbered by this instruction.
static bool callClobbersAnyYmmReg(MachineInstr *MI) {
assert(MI->isCall() && "Can only be called on call instructions.");
for (const MachineOperand &MO : MI->operands()) {
if (!MO.isRegMask())
continue;
for (unsigned reg = X86::YMM0; reg <= X86::YMM15; ++reg) {
if (MO.clobbersPhysReg(reg))
return true;
}
}
return false;
}
/// Insert a vzeroupper instruction before I.
void VZeroUpperInserter::insertVZeroUpper(MachineBasicBlock::iterator I,
MachineBasicBlock &MBB) {
DebugLoc dl = I->getDebugLoc();
BuildMI(MBB, I, dl, TII->get(X86::VZEROUPPER));
++NumVZU;
EverMadeChange = true;
}
/// Add MBB to the DirtySuccessors list if it hasn't already been added.
void VZeroUpperInserter::addDirtySuccessor(MachineBasicBlock &MBB) {
if (!BlockStates[MBB.getNumber()].AddedToDirtySuccessors) {
DirtySuccessors.push_back(&MBB);
BlockStates[MBB.getNumber()].AddedToDirtySuccessors = true;
}
}
/// Loop over all of the instructions in the basic block, inserting vzeroupper
/// instructions before function calls.
void VZeroUpperInserter::processBasicBlock(MachineBasicBlock &MBB) {
// Start by assuming that the block is PASS_THROUGH which implies no unguarded
// calls.
BlockExitState CurState = PASS_THROUGH;
BlockStates[MBB.getNumber()].FirstUnguardedCall = MBB.end();
for (MachineBasicBlock::iterator I = MBB.begin(); I != MBB.end(); ++I) {
MachineInstr *MI = I;
// No need for vzeroupper before iret in interrupt handler function,
// epilogue will restore YMM registers if needed.
bool IsReturnFromX86INTR = IsX86INTR && MI->isReturn();
bool IsControlFlow = MI->isCall() || MI->isReturn();
// Shortcut: don't need to check regular instructions in dirty state.
if ((!IsControlFlow || IsReturnFromX86INTR) && CurState == EXITS_DIRTY)
continue;
// Ignore existing VZERO* instructions.
// FIXME: The existence of these instructions should be used to modify the
// current state and/or used when deciding whether we need to create a VZU.
if (MI->getOpcode() == X86::VZEROALL || MI->getOpcode() == X86::VZEROUPPER)
continue;
if (hasYmmReg(MI)) {
// We found a ymm-using instruction; this could be an AVX instruction,
// or it could be control flow.
CurState = EXITS_DIRTY;
continue;
}
// Check for control-flow out of the current function (which might
// indirectly execute SSE instructions).
if (!IsControlFlow || IsReturnFromX86INTR)
continue;
// If the call won't clobber any YMM register, skip it as well. It usually
// happens on helper function calls (such as '_chkstk', '_ftol2') where
// standard calling convention is not used (RegMask is not used to mark
// register clobbered and register usage (def/imp-def/use) is well-defined
// and explicitly specified.
if (MI->isCall() && !callClobbersAnyYmmReg(MI))
continue;
// The VZEROUPPER instruction resets the upper 128 bits of all AVX
// registers. In addition, the processor changes back to Clean state, after
// which execution of SSE instructions or AVX instructions has no transition
// penalty. Add the VZEROUPPER instruction before any function call/return
// that might execute SSE code.
// FIXME: In some cases, we may want to move the VZEROUPPER into a
// predecessor block.
if (CurState == EXITS_DIRTY) {
// After the inserted VZEROUPPER the state becomes clean again, but
// other YMM may appear before other subsequent calls or even before
// the end of the BB.
insertVZeroUpper(I, MBB);
CurState = EXITS_CLEAN;
} else if (CurState == PASS_THROUGH) {
// If this block is currently in pass-through state and we encounter a
// call then whether we need a vzeroupper or not depends on whether this
// block has successors that exit dirty. Record the location of the call,
// and set the state to EXITS_CLEAN, but do not insert the vzeroupper yet.
// It will be inserted later if necessary.
BlockStates[MBB.getNumber()].FirstUnguardedCall = I;
CurState = EXITS_CLEAN;
}
}
DEBUG(dbgs() << "MBB #" << MBB.getNumber() << " exit state: "
<< getBlockExitStateName(CurState) << '\n');
if (CurState == EXITS_DIRTY)
for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
SE = MBB.succ_end();
SI != SE; ++SI)
addDirtySuccessor(**SI);
BlockStates[MBB.getNumber()].ExitState = CurState;
}
/// Loop over all of the basic blocks, inserting vzeroupper instructions before
/// function calls.
bool VZeroUpperInserter::runOnMachineFunction(MachineFunction &MF) {
const X86Subtarget &ST = MF.getSubtarget<X86Subtarget>();
if (!ST.hasAVX() || ST.hasAVX512() || ST.hasFastPartialYMMWrite())
return false;
TII = ST.getInstrInfo();
MachineRegisterInfo &MRI = MF.getRegInfo();
EverMadeChange = false;
IsX86INTR = MF.getFunction()->getCallingConv() == CallingConv::X86_INTR;
bool FnHasLiveInYmm = checkFnHasLiveInYmm(MRI);
// Fast check: if the function doesn't use any ymm registers, we don't need
// to insert any VZEROUPPER instructions. This is constant-time, so it is
// cheap in the common case of no ymm use.
bool YMMUsed = FnHasLiveInYmm;
if (!YMMUsed) {
const TargetRegisterClass *RC = &X86::VR256RegClass;
for (TargetRegisterClass::iterator i = RC->begin(), e = RC->end(); i != e;
i++) {
if (!MRI.reg_nodbg_empty(*i)) {
YMMUsed = true;
break;
}
}
}
if (!YMMUsed) {
return false;
}
assert(BlockStates.empty() && DirtySuccessors.empty() &&
"X86VZeroUpper state should be clear");
BlockStates.resize(MF.getNumBlockIDs());
// Process all blocks. This will compute block exit states, record the first
// unguarded call in each block, and add successors of dirty blocks to the
// DirtySuccessors list.
for (MachineBasicBlock &MBB : MF)
processBasicBlock(MBB);
// If any YMM regs are live-in to this function, add the entry block to the
// DirtySuccessors list
if (FnHasLiveInYmm)
addDirtySuccessor(MF.front());
// Re-visit all blocks that are successors of EXITS_DIRTY bsocks. Add
// vzeroupper instructions to unguarded calls, and propagate EXITS_DIRTY
// through PASS_THROUGH blocks.
while (!DirtySuccessors.empty()) {
MachineBasicBlock &MBB = *DirtySuccessors.back();
DirtySuccessors.pop_back();
BlockState &BBState = BlockStates[MBB.getNumber()];
// MBB is a successor of a dirty block, so its first call needs to be
// guarded.
if (BBState.FirstUnguardedCall != MBB.end())
insertVZeroUpper(BBState.FirstUnguardedCall, MBB);
// If this successor was a pass-through block, then it is now dirty. Its
// successors need to be added to the worklist (if they haven't been
// already).
if (BBState.ExitState == PASS_THROUGH) {
DEBUG(dbgs() << "MBB #" << MBB.getNumber()
<< " was Pass-through, is now Dirty-out.\n");
for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
SE = MBB.succ_end();
SI != SE; ++SI)
addDirtySuccessor(**SI);
}
}
BlockStates.clear();
return EverMadeChange;
}