[X86] Optimize stackmap shadows on X86.

This patch minimizes the number of nops that must be emitted on X86 to satisfy
stackmap shadow constraints.

To minimize the number of nops inserted, the X86AsmPrinter now records the
size of the most recent stackmap's shadow in the StackMapShadowTracker class,
and tracks the number of instruction bytes emitted since the that stackmap
instruction was encountered. Padding is emitted (if it is required at all)
immediately before the next stackmap/patchpoint instruction, or at the end of
the basic block.

This optimization should reduce code-size and improve performance for people
using the llvm stackmap intrinsic on X86.

<rdar://problem/14959522>



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213892 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Lang Hames 2014-07-24 20:40:55 +00:00
parent 5b93c8af72
commit b96e833817
8 changed files with 171 additions and 45 deletions

View File

@ -264,6 +264,9 @@ public:
/// function. /// function.
virtual void EmitFunctionBodyEnd() {} virtual void EmitFunctionBodyEnd() {}
/// Targets can override this to emit stuff at the end of a basic block.
virtual void EmitBasicBlockEnd(const MachineBasicBlock &MBB) {}
/// Targets should implement this to emit instructions. /// Targets should implement this to emit instructions.
virtual void EmitInstruction(const MachineInstr *) { virtual void EmitInstruction(const MachineInstr *) {
llvm_unreachable("EmitInstruction not implemented"); llvm_unreachable("EmitInstruction not implemented");

View File

@ -804,6 +804,8 @@ void AsmPrinter::EmitFunctionBody() {
} }
} }
} }
EmitBasicBlockEnd(MBB);
} }
// If the last instruction was a prolog label, then we have a situation where // If the last instruction was a prolog label, then we have a situation where

View File

@ -47,6 +47,8 @@ using namespace llvm;
/// runOnMachineFunction - Emit the function body. /// runOnMachineFunction - Emit the function body.
/// ///
bool X86AsmPrinter::runOnMachineFunction(MachineFunction &MF) { bool X86AsmPrinter::runOnMachineFunction(MachineFunction &MF) {
SMShadowTracker.startFunction(MF);
SetupMachineFunction(MF); SetupMachineFunction(MF);
if (Subtarget->isTargetCOFF()) { if (Subtarget->isTargetCOFF()) {

View File

@ -15,6 +15,11 @@
#include "llvm/CodeGen/StackMaps.h" #include "llvm/CodeGen/StackMaps.h"
#include "llvm/Target/TargetMachine.h" #include "llvm/Target/TargetMachine.h"
// Implemented in X86MCInstLower.cpp
namespace {
class X86MCInstLower;
}
namespace llvm { namespace llvm {
class MCStreamer; class MCStreamer;
class MCSymbol; class MCSymbol;
@ -25,9 +30,52 @@ class LLVM_LIBRARY_VISIBILITY X86AsmPrinter : public AsmPrinter {
void GenerateExportDirective(const MCSymbol *Sym, bool IsData); void GenerateExportDirective(const MCSymbol *Sym, bool IsData);
// This utility class tracks the length of a stackmap instruction's 'shadow'.
// It is used by the X86AsmPrinter to ensure that the stackmap shadow
// invariants (i.e. no other stackmaps, patchpoints, or control flow within
// the shadow) are met, while outputting a minimal number of NOPs for padding.
//
// To minimise the number of NOPs used, the shadow tracker counts the number
// of instruction bytes output since the last stackmap. Only if there are too
// few instruction bytes to cover the shadow are NOPs used for padding.
class StackMapShadowTracker {
public:
StackMapShadowTracker(TargetMachine &TM);
~StackMapShadowTracker();
void startFunction(MachineFunction &MF);
void count(MCInst &Inst, const MCSubtargetInfo &STI);
void reset(unsigned RequiredSize) {
RequiredShadowSize = RequiredSize;
CurrentShadowSize = 0;
Count = true;
}
void emitShadowPadding(MCStreamer &OutStreamer, const MCSubtargetInfo &STI);
private:
TargetMachine &TM;
std::unique_ptr<MCCodeEmitter> CodeEmitter;
bool Count;
unsigned RequiredShadowSize, CurrentShadowSize;
};
StackMapShadowTracker SMShadowTracker;
// All instructions emitted by the X86AsmPrinter should use this helper
// method.
//
// This helper function invokes the SMShadowTracker on each instruction before
// outputting it to the OutStream. This allows the shadow tracker to minimise
// the number of NOPs used for stackmap padding.
void EmitAndCountInstruction(MCInst &Inst);
void InsertStackMapShadows(MachineFunction &MF);
void LowerSTACKMAP(const MachineInstr &MI);
void LowerPATCHPOINT(const MachineInstr &MI);
void LowerTlsAddr(X86MCInstLower &MCInstLowering, const MachineInstr &MI);
public: public:
explicit X86AsmPrinter(TargetMachine &TM, MCStreamer &Streamer) explicit X86AsmPrinter(TargetMachine &TM, MCStreamer &Streamer)
: AsmPrinter(TM, Streamer), SM(*this) { : AsmPrinter(TM, Streamer), SM(*this), SMShadowTracker(TM) {
Subtarget = &TM.getSubtarget<X86Subtarget>(); Subtarget = &TM.getSubtarget<X86Subtarget>();
} }
@ -43,6 +91,10 @@ class LLVM_LIBRARY_VISIBILITY X86AsmPrinter : public AsmPrinter {
void EmitInstruction(const MachineInstr *MI) override; void EmitInstruction(const MachineInstr *MI) override;
void EmitBasicBlockEnd(const MachineBasicBlock &MBB) override {
SMShadowTracker.emitShadowPadding(OutStreamer, getSubtargetInfo());
}
bool PrintAsmOperand(const MachineInstr *MI, unsigned OpNo, bool PrintAsmOperand(const MachineInstr *MI, unsigned OpNo,
unsigned AsmVariant, const char *ExtraCode, unsigned AsmVariant, const char *ExtraCode,
raw_ostream &OS) override; raw_ostream &OS) override;

View File

@ -24,12 +24,14 @@
#include "llvm/IR/GlobalValue.h" #include "llvm/IR/GlobalValue.h"
#include "llvm/IR/Mangler.h" #include "llvm/IR/Mangler.h"
#include "llvm/MC/MCAsmInfo.h" #include "llvm/MC/MCAsmInfo.h"
#include "llvm/MC/MCCodeEmitter.h"
#include "llvm/MC/MCContext.h" #include "llvm/MC/MCContext.h"
#include "llvm/MC/MCExpr.h" #include "llvm/MC/MCExpr.h"
#include "llvm/MC/MCInst.h" #include "llvm/MC/MCInst.h"
#include "llvm/MC/MCInstBuilder.h" #include "llvm/MC/MCInstBuilder.h"
#include "llvm/MC/MCStreamer.h" #include "llvm/MC/MCStreamer.h"
#include "llvm/MC/MCSymbol.h" #include "llvm/MC/MCSymbol.h"
#include "llvm/Support/TargetRegistry.h"
using namespace llvm; using namespace llvm;
namespace { namespace {
@ -58,6 +60,52 @@ private:
} // end anonymous namespace } // end anonymous namespace
// Emit a minimal sequence of nops spanning NumBytes bytes.
static void EmitNops(MCStreamer &OS, unsigned NumBytes, bool Is64Bit,
const MCSubtargetInfo &STI);
namespace llvm {
X86AsmPrinter::StackMapShadowTracker::StackMapShadowTracker(TargetMachine &TM)
: TM(TM), Count(false), RequiredShadowSize(0), CurrentShadowSize(0) {}
X86AsmPrinter::StackMapShadowTracker::~StackMapShadowTracker() {}
void
X86AsmPrinter::StackMapShadowTracker::startFunction(MachineFunction &MF) {
CodeEmitter.reset(TM.getTarget().createMCCodeEmitter(*TM.getInstrInfo(),
*TM.getRegisterInfo(),
*TM.getSubtargetImpl(),
MF.getContext()));
}
void X86AsmPrinter::StackMapShadowTracker::count(MCInst &Inst,
const MCSubtargetInfo &STI) {
if (Count) {
SmallString<256> Code;
SmallVector<MCFixup, 4> Fixups;
raw_svector_ostream VecOS(Code);
CodeEmitter->EncodeInstruction(Inst, VecOS, Fixups, STI);
VecOS.flush();
CurrentShadowSize += Code.size();
if (CurrentShadowSize >= RequiredShadowSize)
Count = false; // The shadow is big enough. Stop counting.
}
}
void X86AsmPrinter::StackMapShadowTracker::emitShadowPadding(
MCStreamer &OutStreamer, const MCSubtargetInfo &STI) {
if (Count && CurrentShadowSize < RequiredShadowSize)
EmitNops(OutStreamer, RequiredShadowSize - CurrentShadowSize,
TM.getSubtarget<X86Subtarget>().is64Bit(), STI);
Count = false;
}
void X86AsmPrinter::EmitAndCountInstruction(MCInst &Inst) {
OutStreamer.EmitInstruction(Inst, getSubtargetInfo());
SMShadowTracker.count(Inst, getSubtargetInfo());
}
} // end llvm namespace
X86MCInstLower::X86MCInstLower(const MachineFunction &mf, X86MCInstLower::X86MCInstLower(const MachineFunction &mf,
X86AsmPrinter &asmprinter) X86AsmPrinter &asmprinter)
: Ctx(mf.getContext()), MF(mf), TM(mf.getTarget()), : Ctx(mf.getContext()), MF(mf), TM(mf.getTarget()),
@ -602,10 +650,8 @@ ReSimplify:
} }
} }
static void LowerTlsAddr(MCStreamer &OutStreamer, void X86AsmPrinter::LowerTlsAddr(X86MCInstLower &MCInstLowering,
X86MCInstLower &MCInstLowering, const MachineInstr &MI) {
const MachineInstr &MI,
const MCSubtargetInfo& STI) {
bool is64Bits = MI.getOpcode() == X86::TLS_addr64 || bool is64Bits = MI.getOpcode() == X86::TLS_addr64 ||
MI.getOpcode() == X86::TLS_base_addr64; MI.getOpcode() == X86::TLS_base_addr64;
@ -615,7 +661,7 @@ static void LowerTlsAddr(MCStreamer &OutStreamer,
MCContext &context = OutStreamer.getContext(); MCContext &context = OutStreamer.getContext();
if (needsPadding) if (needsPadding)
OutStreamer.EmitInstruction(MCInstBuilder(X86::DATA16_PREFIX), STI); EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
MCSymbolRefExpr::VariantKind SRVK; MCSymbolRefExpr::VariantKind SRVK;
switch (MI.getOpcode()) { switch (MI.getOpcode()) {
@ -662,12 +708,12 @@ static void LowerTlsAddr(MCStreamer &OutStreamer,
LEA.addOperand(MCOperand::CreateExpr(symRef)); // disp LEA.addOperand(MCOperand::CreateExpr(symRef)); // disp
LEA.addOperand(MCOperand::CreateReg(0)); // seg LEA.addOperand(MCOperand::CreateReg(0)); // seg
} }
OutStreamer.EmitInstruction(LEA, STI); EmitAndCountInstruction(LEA);
if (needsPadding) { if (needsPadding) {
OutStreamer.EmitInstruction(MCInstBuilder(X86::DATA16_PREFIX), STI); EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
OutStreamer.EmitInstruction(MCInstBuilder(X86::DATA16_PREFIX), STI); EmitAndCountInstruction(MCInstBuilder(X86::DATA16_PREFIX));
OutStreamer.EmitInstruction(MCInstBuilder(X86::REX64_PREFIX), STI); EmitAndCountInstruction(MCInstBuilder(X86::REX64_PREFIX));
} }
StringRef name = is64Bits ? "__tls_get_addr" : "___tls_get_addr"; StringRef name = is64Bits ? "__tls_get_addr" : "___tls_get_addr";
@ -677,9 +723,9 @@ static void LowerTlsAddr(MCStreamer &OutStreamer,
MCSymbolRefExpr::VK_PLT, MCSymbolRefExpr::VK_PLT,
context); context);
OutStreamer.EmitInstruction(MCInstBuilder(is64Bits ? X86::CALL64pcrel32 EmitAndCountInstruction(MCInstBuilder(is64Bits ? X86::CALL64pcrel32
: X86::CALLpcrel32) : X86::CALLpcrel32)
.addExpr(tlsRef), STI); .addExpr(tlsRef));
} }
/// \brief Emit the optimal amount of multi-byte nops on X86. /// \brief Emit the optimal amount of multi-byte nops on X86.
@ -725,10 +771,9 @@ static void EmitNops(MCStreamer &OS, unsigned NumBytes, bool Is64Bit, const MCSu
break; break;
case X86::NOOPL: case X86::NOOPL:
case X86::NOOPW: case X86::NOOPW:
OS.EmitInstruction(MCInstBuilder(Opc).addReg(BaseReg).addImm(ScaleVal) OS.EmitInstruction(MCInstBuilder(Opc).addReg(BaseReg)
.addReg(IndexReg) .addImm(ScaleVal).addReg(IndexReg)
.addImm(Displacement) .addImm(Displacement).addReg(SegmentReg), STI);
.addReg(SegmentReg), STI);
break; break;
} }
} // while (NumBytes) } // while (NumBytes)
@ -736,22 +781,20 @@ static void EmitNops(MCStreamer &OS, unsigned NumBytes, bool Is64Bit, const MCSu
// Lower a stackmap of the form: // Lower a stackmap of the form:
// <id>, <shadowBytes>, ... // <id>, <shadowBytes>, ...
static void LowerSTACKMAP(MCStreamer &OS, StackMaps &SM, void X86AsmPrinter::LowerSTACKMAP(const MachineInstr &MI) {
const MachineInstr &MI, bool Is64Bit, const MCSubtargetInfo& STI) { SMShadowTracker.emitShadowPadding(OutStreamer, getSubtargetInfo());
unsigned NumBytes = MI.getOperand(1).getImm();
SM.recordStackMap(MI); SM.recordStackMap(MI);
// Emit padding. unsigned NumShadowBytes = MI.getOperand(1).getImm();
// FIXME: These nops ensure that the stackmap's shadow is covered by SMShadowTracker.reset(NumShadowBytes);
// instructions from the same basic block, but the nops should not be
// necessary if instructions from the same block follow the stackmap.
EmitNops(OS, NumBytes, Is64Bit, STI);
} }
// Lower a patchpoint of the form: // Lower a patchpoint of the form:
// [<def>], <id>, <numBytes>, <target>, <numArgs>, <cc>, ... // [<def>], <id>, <numBytes>, <target>, <numArgs>, <cc>, ...
static void LowerPATCHPOINT(MCStreamer &OS, StackMaps &SM, void X86AsmPrinter::LowerPATCHPOINT(const MachineInstr &MI) {
const MachineInstr &MI, bool Is64Bit, const MCSubtargetInfo& STI) { assert(Subtarget->is64Bit() && "Patchpoint currently only supports X86-64");
assert(Is64Bit && "Patchpoint currently only supports X86-64");
SMShadowTracker.emitShadowPadding(OutStreamer, getSubtargetInfo());
SM.recordPatchPoint(MI); SM.recordPatchPoint(MI);
PatchPointOpers opers(&MI); PatchPointOpers opers(&MI);
@ -766,16 +809,17 @@ static void LowerPATCHPOINT(MCStreamer &OS, StackMaps &SM,
EncodedBytes = 13; EncodedBytes = 13;
else else
EncodedBytes = 12; EncodedBytes = 12;
OS.EmitInstruction(MCInstBuilder(X86::MOV64ri).addReg(ScratchReg) EmitAndCountInstruction(MCInstBuilder(X86::MOV64ri).addReg(ScratchReg)
.addImm(CallTarget), STI); .addImm(CallTarget));
OS.EmitInstruction(MCInstBuilder(X86::CALL64r).addReg(ScratchReg), STI); EmitAndCountInstruction(MCInstBuilder(X86::CALL64r).addReg(ScratchReg));
} }
// Emit padding. // Emit padding.
unsigned NumBytes = opers.getMetaOper(PatchPointOpers::NBytesPos).getImm(); unsigned NumBytes = opers.getMetaOper(PatchPointOpers::NBytesPos).getImm();
assert(NumBytes >= EncodedBytes && assert(NumBytes >= EncodedBytes &&
"Patchpoint can't request size less than the length of a call."); "Patchpoint can't request size less than the length of a call.");
EmitNops(OS, NumBytes - EncodedBytes, Is64Bit, STI); EmitNops(OutStreamer, NumBytes - EncodedBytes, Subtarget->is64Bit(),
getSubtargetInfo());
} }
void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) { void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) {
@ -812,7 +856,7 @@ void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) {
case X86::TLS_addr64: case X86::TLS_addr64:
case X86::TLS_base_addr32: case X86::TLS_base_addr32:
case X86::TLS_base_addr64: case X86::TLS_base_addr64:
return LowerTlsAddr(OutStreamer, MCInstLowering, *MI, getSubtargetInfo()); return LowerTlsAddr(MCInstLowering, *MI);
case X86::MOVPC32r: { case X86::MOVPC32r: {
// This is a pseudo op for a two instruction sequence with a label, which // This is a pseudo op for a two instruction sequence with a label, which
@ -825,15 +869,15 @@ void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) {
MCSymbol *PICBase = MF->getPICBaseSymbol(); MCSymbol *PICBase = MF->getPICBaseSymbol();
// FIXME: We would like an efficient form for this, so we don't have to do a // FIXME: We would like an efficient form for this, so we don't have to do a
// lot of extra uniquing. // lot of extra uniquing.
EmitToStreamer(OutStreamer, MCInstBuilder(X86::CALLpcrel32) EmitAndCountInstruction(MCInstBuilder(X86::CALLpcrel32)
.addExpr(MCSymbolRefExpr::Create(PICBase, OutContext))); .addExpr(MCSymbolRefExpr::Create(PICBase, OutContext)));
// Emit the label. // Emit the label.
OutStreamer.EmitLabel(PICBase); OutStreamer.EmitLabel(PICBase);
// popl $reg // popl $reg
EmitToStreamer(OutStreamer, MCInstBuilder(X86::POP32r) EmitAndCountInstruction(MCInstBuilder(X86::POP32r)
.addReg(MI->getOperand(0).getReg())); .addReg(MI->getOperand(0).getReg()));
return; return;
} }
@ -863,7 +907,7 @@ void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) {
DotExpr = MCBinaryExpr::CreateAdd(MCSymbolRefExpr::Create(OpSym,OutContext), DotExpr = MCBinaryExpr::CreateAdd(MCSymbolRefExpr::Create(OpSym,OutContext),
DotExpr, OutContext); DotExpr, OutContext);
EmitToStreamer(OutStreamer, MCInstBuilder(X86::ADD32ri) EmitAndCountInstruction(MCInstBuilder(X86::ADD32ri)
.addReg(MI->getOperand(0).getReg()) .addReg(MI->getOperand(0).getReg())
.addReg(MI->getOperand(1).getReg()) .addReg(MI->getOperand(1).getReg())
.addExpr(DotExpr)); .addExpr(DotExpr));
@ -871,21 +915,21 @@ void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) {
} }
case TargetOpcode::STACKMAP: case TargetOpcode::STACKMAP:
return LowerSTACKMAP(OutStreamer, SM, *MI, Subtarget->is64Bit(), getSubtargetInfo()); return LowerSTACKMAP(*MI);
case TargetOpcode::PATCHPOINT: case TargetOpcode::PATCHPOINT:
return LowerPATCHPOINT(OutStreamer, SM, *MI, Subtarget->is64Bit(), getSubtargetInfo()); return LowerPATCHPOINT(*MI);
case X86::MORESTACK_RET: case X86::MORESTACK_RET:
EmitToStreamer(OutStreamer, MCInstBuilder(getRetOpcode(*Subtarget))); EmitAndCountInstruction(MCInstBuilder(getRetOpcode(*Subtarget)));
return; return;
case X86::MORESTACK_RET_RESTORE_R10: case X86::MORESTACK_RET_RESTORE_R10:
// Return, then restore R10. // Return, then restore R10.
EmitToStreamer(OutStreamer, MCInstBuilder(getRetOpcode(*Subtarget))); EmitAndCountInstruction(MCInstBuilder(getRetOpcode(*Subtarget)));
EmitToStreamer(OutStreamer, MCInstBuilder(X86::MOV64rr) EmitAndCountInstruction(MCInstBuilder(X86::MOV64rr)
.addReg(X86::R10) .addReg(X86::R10)
.addReg(X86::RAX)); .addReg(X86::RAX));
return; return;
case X86::SEH_PushReg: case X86::SEH_PushReg:
@ -922,5 +966,5 @@ void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) {
MCInst TmpInst; MCInst TmpInst;
MCInstLowering.Lower(MI, TmpInst); MCInstLowering.Lower(MI, TmpInst);
EmitToStreamer(OutStreamer, TmpInst); EmitAndCountInstruction(TmpInst);
} }

View File

@ -224,6 +224,7 @@ entry:
tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 28, i32 28) tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 28, i32 28)
tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 29, i32 29) tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 29, i32 29)
tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 30, i32 30) tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 30, i32 30)
tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 31, i32 0)
ret void ret void
} }

View File

@ -0,0 +1,21 @@
; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=corei7 -disable-fp-elim | FileCheck %s
; Check that the X86 stackmap shadow optimization is only outputting a 1-byte
; nop here. 8-bytes are requested, but 7 are covered by the code for the call to
; bar, the frame teardown and the return.
define void @shadow_optimization_test() {
entry:
; CHECK-LABEL: shadow_optimization_test:
; CHECK: callq _bar
; CHECK-NOT: nop
; CHECK: callq _bar
; CHECK: retq
; CHECK: nop
call void @bar()
tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 0, i32 8)
call void @bar()
ret void
}
declare void @llvm.experimental.stackmap(i64, i32, ...)
declare void @bar()

View File

@ -41,6 +41,7 @@ entry:
tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 13, i32 13) tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 13, i32 13)
tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 14, i32 14) tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 14, i32 14)
tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 15, i32 15) tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 15, i32 15)
tail call void (i64, i32, ...)* @llvm.experimental.stackmap(i64 16, i32 0)
ret void ret void
} }