llvm-project/llvm/lib/Target/ARM/ARMSubtarget.cpp

206 lines
6.4 KiB
C++
Raw Normal View History

//===-- ARMSubtarget.cpp - ARM Subtarget Information ----------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file implements the ARM specific subclass of TargetSubtargetInfo.
//
//===----------------------------------------------------------------------===//
#include "ARMSubtarget.h"
#include "ARMBaseInstrInfo.h"
#include "ARMBaseRegisterInfo.h"
#include "llvm/IR/GlobalValue.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Target/TargetInstrInfo.h"
#define GET_SUBTARGETINFO_TARGET_DESC
#define GET_SUBTARGETINFO_CTOR
#include "ARMGenSubtargetInfo.inc"
using namespace llvm;
static cl::opt<bool>
ReserveR9("arm-reserve-r9", cl::Hidden,
cl::desc("Reserve R9, making it unavailable as GPR"));
static cl::opt<bool>
DarwinUseMOVT("arm-darwin-use-movt", cl::init(true), cl::Hidden);
static cl::opt<bool>
UseFusedMulOps("arm-use-mulops",
cl::init(true), cl::Hidden);
static cl::opt<bool>
StrictAlign("arm-strict-align", cl::Hidden,
cl::desc("Disallow all unaligned memory accesses"));
ARMSubtarget::ARMSubtarget(const std::string &TT, const std::string &CPU,
const std::string &FS)
: ARMGenSubtargetInfo(TT, CPU, FS)
, ARMProcFamily(Others)
, HasV4TOps(false)
, HasV5TOps(false)
, HasV5TEOps(false)
, HasV6Ops(false)
, HasV6T2Ops(false)
, HasV7Ops(false)
, HasVFPv2(false)
, HasVFPv3(false)
, HasVFPv4(false)
, HasNEON(false)
, UseNEONForSinglePrecisionFP(false)
, UseMulOps(UseFusedMulOps)
Making use of VFP / NEON floating point multiply-accumulate / subtraction is difficult on current ARM implementations for a few reasons. 1. Even though a single vmla has latency that is one cycle shorter than a pair of vmul + vadd, a RAW hazard during the first (4? on Cortex-a8) can cause additional pipeline stall. So it's frequently better to single codegen vmul + vadd. 2. A vmla folowed by a vmul, vmadd, or vsub causes the second fp instruction to stall for 4 cycles. We need to schedule them apart. 3. A vmla followed vmla is a special case. Obvious issuing back to back RAW vmla + vmla is very bad. But this isn't ideal either: vmul vadd vmla Instead, we want to expand the second vmla: vmla vmul vadd Even with the 4 cycle vmul stall, the second sequence is still 2 cycles faster. Up to now, isel simply avoid codegen'ing fp vmla / vmls. This works well enough but it isn't the optimial solution. This patch attempts to make it possible to use vmla / vmls in cases where it is profitable. A. Add missing isel predicates which cause vmla to be codegen'ed. B. Make sure the fmul in (fadd (fmul)) has a single use. We don't want to compute a fmul and a fmla. C. Add additional isel checks for vmla, avoid cases where vmla is feeding into fp instructions (except for the #3 exceptional case). D. Add ARM hazard recognizer to model the vmla / vmls hazards. E. Add a special pre-regalloc case to expand vmla / vmls when it's likely the vmla / vmls will trigger one of the special hazards. Work in progress, only A+B are enabled. llvm-svn: 120960
2010-12-06 06:04:16 +08:00
, SlowFPVMLx(false)
, HasVMLxForwarding(false)
, SlowFPBrcc(false)
, InThumbMode(false)
, HasThumb2(false)
, IsMClass(false)
, NoARM(false)
, PostRAScheduler(false)
, IsR9Reserved(ReserveR9)
, UseMovt(false)
, SupportsTailCall(false)
, HasFP16(false)
, HasD16(false)
, HasHardwareDivide(false)
, HasHardwareDivideInARM(false)
, HasT2ExtractPack(false)
, HasDataBarrier(false)
, Pref32BitThumb(false)
, AvoidCPSRPartialUpdate(false)
, AvoidMOVsShifterOperand(false)
, HasRAS(false)
, HasMPExtension(false)
, FPOnlySP(false)
, AllowsUnalignedMem(false)
, Thumb2DSP(false)
, UseNaClTrap(false)
, stackAlignment(4)
, CPUString(CPU)
, TargetTriple(TT)
, TargetABI(ARM_ABI_APCS) {
// Determine default and user specified characteristics
if (CPUString.empty())
CPUString = "generic";
// Insert the architecture feature derived from the target triple into the
// feature string. This is important for setting features that are implied
// based on the architecture version.
std::string ArchFS = ARM_MC::ParseARMTriple(TT, CPUString);
if (!FS.empty()) {
if (!ArchFS.empty())
ArchFS = ArchFS + "," + FS;
else
ArchFS = FS;
}
ParseSubtargetFeatures(CPUString, ArchFS);
// Thumb2 implies at least V6T2. FIXME: Fix tests to explicitly specify a
// ARM version or CPU and then remove this.
if (!HasV6T2Ops && hasThumb2())
HasV4TOps = HasV5TOps = HasV5TEOps = HasV6Ops = HasV6T2Ops = true;
// Keep a pointer to static instruction cost data for the specified CPU.
SchedModel = getSchedModelForCPU(CPUString);
// Initialize scheduling itinerary for the specified CPU.
InstrItins = getInstrItineraryForCPU(CPUString);
if ((TT.find("eabi") != std::string::npos) || (isTargetIOS() && isMClass()))
// FIXME: We might want to separate AAPCS and EABI. Some systems, e.g.
// Darwin-EABI conforms to AACPS but not the rest of EABI.
TargetABI = ARM_ABI_AAPCS;
if (isAAPCS_ABI())
stackAlignment = 8;
if (!isTargetIOS())
UseMovt = hasV6T2Ops();
else {
IsR9Reserved = ReserveR9 | !HasV6Ops;
UseMovt = DarwinUseMOVT && hasV6T2Ops();
SupportsTailCall = !getTargetTriple().isOSVersionLT(5, 0);
}
if (!isThumb() || hasThumb2())
PostRAScheduler = true;
// v6+ may or may not support unaligned mem access depending on the system
// configuration.
if (!StrictAlign && hasV6Ops() && isTargetDarwin())
AllowsUnalignedMem = true;
}
/// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol.
bool
ARMSubtarget::GVIsIndirectSymbol(const GlobalValue *GV,
Reloc::Model RelocM) const {
if (RelocM == Reloc::Static)
return false;
// Materializable GVs (in JIT lazy compilation mode) do not require an extra
// load from stub.
bool isDecl = GV->hasAvailableExternallyLinkage();
if (GV->isDeclaration() && !GV->isMaterializable())
isDecl = true;
if (!isTargetDarwin()) {
// Extra load is needed for all externally visible.
if (GV->hasLocalLinkage() || GV->hasHiddenVisibility())
return false;
return true;
} else {
if (RelocM == Reloc::PIC_) {
// If this is a strong reference to a definition, it is definitely not
// through a stub.
if (!isDecl && !GV->isWeakForLinker())
return false;
// Unless we have a symbol with hidden visibility, we have to go through a
// normal $non_lazy_ptr stub because this symbol might be resolved late.
if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
return true;
// If symbol visibility is hidden, we have a stub for common symbol
// references and external declarations.
if (isDecl || GV->hasCommonLinkage())
// Hidden $non_lazy_ptr reference.
return true;
return false;
} else {
// If this is a strong reference to a definition, it is definitely not
// through a stub.
if (!isDecl && !GV->isWeakForLinker())
return false;
2010-12-24 12:28:06 +08:00
// Unless we have a symbol with hidden visibility, we have to go through a
// normal $non_lazy_ptr stub because this symbol might be resolved late.
if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
return true;
}
}
return false;
}
unsigned ARMSubtarget::getMispredictionPenalty() const {
return SchedModel->MispredictPenalty;
}
bool ARMSubtarget::enablePostRAScheduler(
CodeGenOpt::Level OptLevel,
TargetSubtargetInfo::AntiDepBreakMode& Mode,
RegClassVector& CriticalPathRCs) const {
Mode = TargetSubtargetInfo::ANTIDEP_CRITICAL;
CriticalPathRCs.clear();
CriticalPathRCs.push_back(&ARM::GPRRegClass);
return PostRAScheduler && OptLevel >= CodeGenOpt::Default;
}