2005-01-07 07:44:22 +00:00
|
|
|
//===-- llvm/Target/TargetLowering.h - Target Lowering Info -----*- C++ -*-===//
|
2005-04-21 20:59:05 +00:00
|
|
|
//
|
2005-01-07 07:44:22 +00:00
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
2007-12-29 19:59:42 +00:00
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
2005-04-21 20:59:05 +00:00
|
|
|
//
|
2005-01-07 07:44:22 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// This file describes how to lower LLVM code to machine code. This has two
|
|
|
|
// main components:
|
|
|
|
//
|
|
|
|
// 1. Which ValueTypes are natively supported by the target.
|
|
|
|
// 2. Which operations are supported for supported ValueTypes.
|
2005-07-19 04:52:44 +00:00
|
|
|
// 3. Cost thresholds for alternative implementations of certain operations.
|
2005-01-07 07:44:22 +00:00
|
|
|
//
|
|
|
|
// In addition it has a few other components, like information about FP
|
|
|
|
// immediates.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#ifndef LLVM_TARGET_TARGETLOWERING_H
|
|
|
|
#define LLVM_TARGET_TARGETLOWERING_H
|
|
|
|
|
2012-12-03 17:02:12 +00:00
|
|
|
#include "llvm/ADT/DenseMap.h"
|
2012-10-08 23:06:34 +00:00
|
|
|
#include "llvm/AddressingMode.h"
|
2012-12-03 17:02:12 +00:00
|
|
|
#include "llvm/Attributes.h"
|
2009-08-14 20:10:52 +00:00
|
|
|
#include "llvm/CallingConv.h"
|
2012-12-03 17:02:12 +00:00
|
|
|
#include "llvm/CodeGen/RuntimeLibcalls.h"
|
|
|
|
#include "llvm/CodeGen/SelectionDAGNodes.h"
|
2008-02-26 02:33:44 +00:00
|
|
|
#include "llvm/InlineAsm.h"
|
2012-05-25 16:35:28 +00:00
|
|
|
#include "llvm/Support/CallSite.h"
|
2009-06-19 22:08:58 +00:00
|
|
|
#include "llvm/Support/DebugLoc.h"
|
2010-07-07 15:28:42 +00:00
|
|
|
#include "llvm/Target/TargetCallingConv.h"
|
2009-02-27 13:37:18 +00:00
|
|
|
#include "llvm/Target/TargetMachine.h"
|
2009-01-05 19:47:30 +00:00
|
|
|
#include <climits>
|
2006-04-04 00:25:10 +00:00
|
|
|
#include <map>
|
2007-01-12 23:21:42 +00:00
|
|
|
#include <vector>
|
2005-01-07 07:44:22 +00:00
|
|
|
|
|
|
|
namespace llvm {
|
2009-01-05 17:59:02 +00:00
|
|
|
class CallInst;
|
2011-02-28 17:17:53 +00:00
|
|
|
class CCState;
|
2008-08-19 21:26:29 +00:00
|
|
|
class FastISel;
|
2010-07-07 16:29:44 +00:00
|
|
|
class FunctionLoweringInfo;
|
2010-09-13 18:15:37 +00:00
|
|
|
class ImmutableCallSite;
|
2012-03-13 20:59:56 +00:00
|
|
|
class IntrinsicInst;
|
2008-05-12 20:08:05 +00:00
|
|
|
class MachineBasicBlock;
|
2008-08-19 21:26:29 +00:00
|
|
|
class MachineFunction;
|
2008-05-12 20:08:05 +00:00
|
|
|
class MachineInstr;
|
2010-01-26 04:05:28 +00:00
|
|
|
class MachineJumpTableInfo;
|
|
|
|
class MCContext;
|
|
|
|
class MCExpr;
|
2010-11-06 11:45:59 +00:00
|
|
|
template<typename T> class SmallVectorImpl;
|
2012-10-08 16:38:25 +00:00
|
|
|
class DataLayout;
|
2008-05-12 20:08:05 +00:00
|
|
|
class TargetRegisterClass;
|
2012-08-03 04:06:28 +00:00
|
|
|
class TargetLibraryInfo;
|
2009-07-28 03:13:23 +00:00
|
|
|
class TargetLoweringObjectFile;
|
2008-05-12 20:08:05 +00:00
|
|
|
class Value;
|
2005-01-07 07:44:22 +00:00
|
|
|
|
2012-01-12 02:35:23 +00:00
|
|
|
namespace Sched {
|
|
|
|
enum Preference {
|
|
|
|
None, // No preference
|
2012-01-12 18:27:52 +00:00
|
|
|
Source, // Follow source order.
|
2012-01-12 02:35:23 +00:00
|
|
|
RegPressure, // Scheduling for lowest register pressure.
|
|
|
|
Hybrid, // Scheduling for both latency and register pressure.
|
2012-02-01 22:13:57 +00:00
|
|
|
ILP, // Scheduling for ILP in low register pressure mode.
|
|
|
|
VLIW // Scheduling for VLIW targets.
|
2012-01-12 02:35:23 +00:00
|
|
|
};
|
|
|
|
}
|
|
|
|
|
2009-02-27 13:37:18 +00:00
|
|
|
|
2005-01-07 07:44:22 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
/// TargetLowering - This class defines information used to lower LLVM code to
|
|
|
|
/// legal SelectionDAG operators that the target instruction selector can accept
|
|
|
|
/// natively.
|
|
|
|
///
|
|
|
|
/// This class also defines callbacks that targets must implement to lower
|
|
|
|
/// target-specific constructs to SelectionDAG operators.
|
|
|
|
///
|
|
|
|
class TargetLowering {
|
2012-09-17 06:59:23 +00:00
|
|
|
TargetLowering(const TargetLowering&) LLVM_DELETED_FUNCTION;
|
|
|
|
void operator=(const TargetLowering&) LLVM_DELETED_FUNCTION;
|
2005-01-07 07:44:22 +00:00
|
|
|
public:
|
2005-01-16 07:27:49 +00:00
|
|
|
/// LegalizeAction - This enum indicates whether operations are valid for a
|
|
|
|
/// target, and if not, what action should be used to make them valid.
|
|
|
|
enum LegalizeAction {
|
|
|
|
Legal, // The target natively supports this operation.
|
|
|
|
Promote, // This operation should be executed in a larger type.
|
|
|
|
Expand, // Try to expand this to other ops, otherwise use a libcall.
|
2006-02-22 16:23:43 +00:00
|
|
|
Custom // Use the LowerOperation hook to implement custom lowering.
|
2005-01-16 07:27:49 +00:00
|
|
|
};
|
|
|
|
|
2011-11-13 22:11:24 +00:00
|
|
|
/// LegalizeTypeAction - This enum indicates whether a types are legal for a
|
2011-05-28 17:57:14 +00:00
|
|
|
/// target, and if not, what action should be used to make them valid.
|
|
|
|
enum LegalizeTypeAction {
|
|
|
|
TypeLegal, // The target natively supports this type.
|
|
|
|
TypePromoteInteger, // Replace this integer with a larger one.
|
|
|
|
TypeExpandInteger, // Split this integer into two of half the size.
|
|
|
|
TypeSoftenFloat, // Convert this float to a same size integer type.
|
|
|
|
TypeExpandFloat, // Split this float into two of half the size.
|
|
|
|
TypeScalarizeVector, // Replace this one-element vector with its element.
|
|
|
|
TypeSplitVector, // Split this vector into two of half the size.
|
|
|
|
TypeWidenVector // This vector should be widened into a larger vector.
|
|
|
|
};
|
|
|
|
|
2012-10-24 20:59:17 +00:00
|
|
|
/// LegalizeKind holds the legalization kind that needs to happen to EVT
|
|
|
|
/// in order to type-legalize it.
|
|
|
|
typedef std::pair<LegalizeTypeAction, EVT> LegalizeKind;
|
|
|
|
|
2008-11-23 15:47:28 +00:00
|
|
|
enum BooleanContent { // How the target represents true/false values.
|
|
|
|
UndefinedBooleanContent, // Only bit 0 counts, the rest can hold garbage.
|
|
|
|
ZeroOrOneBooleanContent, // All bits zero except for bit 0.
|
|
|
|
ZeroOrNegativeOneBooleanContent // All bits equal to bit 0.
|
2005-04-07 19:41:18 +00:00
|
|
|
};
|
|
|
|
|
2012-09-02 12:10:19 +00:00
|
|
|
enum SelectSupportKind {
|
|
|
|
ScalarValSelect, // The target supports scalar selects (ex: cmov).
|
|
|
|
ScalarCondVectorVal, // The target supports selects with a scalar condition
|
|
|
|
// and vector values (ex: cmov).
|
|
|
|
VectorMaskSelect // The target supports vector selects with a vector
|
|
|
|
// mask (ex: x86 blends).
|
|
|
|
};
|
|
|
|
|
2011-09-06 19:07:46 +00:00
|
|
|
static ISD::NodeType getExtendForContent(BooleanContent Content) {
|
|
|
|
switch (Content) {
|
|
|
|
case UndefinedBooleanContent:
|
|
|
|
// Extend by adding rubbish bits.
|
|
|
|
return ISD::ANY_EXTEND;
|
|
|
|
case ZeroOrOneBooleanContent:
|
|
|
|
// Extend by adding zero bits.
|
|
|
|
return ISD::ZERO_EXTEND;
|
|
|
|
case ZeroOrNegativeOneBooleanContent:
|
|
|
|
// Extend by copying the sign bit.
|
|
|
|
return ISD::SIGN_EXTEND;
|
|
|
|
}
|
2012-01-10 18:08:01 +00:00
|
|
|
llvm_unreachable("Invalid content kind");
|
2011-09-06 19:07:46 +00:00
|
|
|
}
|
|
|
|
|
2009-07-28 03:13:23 +00:00
|
|
|
/// NOTE: The constructor takes ownership of TLOF.
|
2010-04-21 01:34:56 +00:00
|
|
|
explicit TargetLowering(const TargetMachine &TM,
|
|
|
|
const TargetLoweringObjectFile *TLOF);
|
2005-01-16 07:27:49 +00:00
|
|
|
virtual ~TargetLowering();
|
2005-01-07 07:44:22 +00:00
|
|
|
|
2010-04-21 01:34:56 +00:00
|
|
|
const TargetMachine &getTargetMachine() const { return TM; }
|
2012-10-08 16:38:25 +00:00
|
|
|
const DataLayout *getDataLayout() const { return TD; }
|
2010-04-21 01:34:56 +00:00
|
|
|
const TargetLoweringObjectFile &getObjFileLowering() const { return TLOF; }
|
2005-04-21 20:59:05 +00:00
|
|
|
|
2008-02-11 10:37:04 +00:00
|
|
|
bool isBigEndian() const { return !IsLittleEndian; }
|
2005-01-07 07:44:22 +00:00
|
|
|
bool isLittleEndian() const { return IsLittleEndian; }
|
2012-10-09 16:06:12 +00:00
|
|
|
// Return the pointer type for the given address space, defaults to
|
|
|
|
// the pointer type from the data layout.
|
|
|
|
// FIXME: The default needs to be removed once all the code is updated.
|
2012-10-15 16:24:29 +00:00
|
|
|
virtual MVT getPointerTy(uint32_t AS = 0) const { return PointerTy; }
|
2011-02-25 21:41:48 +00:00
|
|
|
virtual MVT getShiftAmountTy(EVT LHSTy) const;
|
2005-01-16 23:59:30 +00:00
|
|
|
|
2007-01-05 23:31:08 +00:00
|
|
|
/// isSelectExpensive - Return true if the select operation is expensive for
|
2005-05-12 02:02:21 +00:00
|
|
|
/// this target.
|
2007-01-05 23:31:08 +00:00
|
|
|
bool isSelectExpensive() const { return SelectIsExpensive; }
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2012-09-02 12:10:19 +00:00
|
|
|
virtual bool isSelectSupported(SelectSupportKind kind) const { return true; }
|
|
|
|
|
2012-11-29 14:26:24 +00:00
|
|
|
/// shouldSplitVectorElementType - Return true if a vector of the given type
|
|
|
|
/// should be split (TypeSplitVector) instead of promoted
|
|
|
|
/// (TypePromoteInteger) during type legalization.
|
|
|
|
virtual bool shouldSplitVectorElementType(EVT VT) const { return false; }
|
|
|
|
|
2005-10-21 00:02:42 +00:00
|
|
|
/// isIntDivCheap() - Return true if integer divide is usually cheaper than
|
2005-10-20 02:14:14 +00:00
|
|
|
/// a sequence of several shifts, adds, and multiplies for this target.
|
2005-10-21 00:02:42 +00:00
|
|
|
bool isIntDivCheap() const { return IntDivIsCheap; }
|
2005-05-12 02:02:21 +00:00
|
|
|
|
2012-09-04 18:22:17 +00:00
|
|
|
/// isSlowDivBypassed - Returns true if target has indicated at least one
|
|
|
|
/// type should be bypassed.
|
2012-10-04 21:33:40 +00:00
|
|
|
bool isSlowDivBypassed() const { return !BypassSlowDivWidths.empty(); }
|
2012-09-04 18:22:17 +00:00
|
|
|
|
|
|
|
/// getBypassSlowDivTypes - Returns map of slow types for division or
|
|
|
|
/// remainder with corresponding fast types
|
2012-10-04 21:33:40 +00:00
|
|
|
const DenseMap<unsigned int, unsigned int> &getBypassSlowDivWidths() const {
|
|
|
|
return BypassSlowDivWidths;
|
2012-09-04 18:22:17 +00:00
|
|
|
}
|
|
|
|
|
2005-10-21 00:02:42 +00:00
|
|
|
/// isPow2DivCheap() - Return true if pow2 div is cheaper than a chain of
|
|
|
|
/// srl/add/sra.
|
|
|
|
bool isPow2DivCheap() const { return Pow2DivIsCheap; }
|
Loosen up iv reuse to allow reuse of the same stride but a larger type when truncating from the larger type to smaller type is free.
e.g.
Turns this loop:
LBB1_1: # entry.bb_crit_edge
xorl %ecx, %ecx
xorw %dx, %dx
movw %dx, %si
LBB1_2: # bb
movl L_X$non_lazy_ptr, %edi
movw %si, (%edi)
movl L_Y$non_lazy_ptr, %edi
movw %dx, (%edi)
addw $4, %dx
incw %si
incl %ecx
cmpl %eax, %ecx
jne LBB1_2 # bb
into
LBB1_1: # entry.bb_crit_edge
xorl %ecx, %ecx
xorw %dx, %dx
LBB1_2: # bb
movl L_X$non_lazy_ptr, %esi
movw %cx, (%esi)
movl L_Y$non_lazy_ptr, %esi
movw %dx, (%esi)
addw $4, %dx
incl %ecx
cmpl %eax, %ecx
jne LBB1_2 # bb
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43375 91177308-0d34-0410-b5e6-96231b3b80d8
2007-10-26 01:56:11 +00:00
|
|
|
|
2010-11-30 18:12:52 +00:00
|
|
|
/// isJumpExpensive() - Return true if Flow Control is an expensive operation
|
|
|
|
/// that should be avoided.
|
|
|
|
bool isJumpExpensive() const { return JumpIsExpensive; }
|
|
|
|
|
2012-05-05 12:49:14 +00:00
|
|
|
/// isPredictableSelectExpensive - Return true if selects are only cheaper
|
|
|
|
/// than branches if the branch is unlikely to be predicted right.
|
|
|
|
bool isPredictableSelectExpensive() const {
|
|
|
|
return predictableSelectIsExpensive;
|
|
|
|
}
|
|
|
|
|
2009-01-05 21:33:06 +00:00
|
|
|
/// getSetCCResultType - Return the ValueType of the result of SETCC
|
|
|
|
/// operations. Also used to obtain the target's preferred type for
|
|
|
|
/// the condition operand of SELECT and BRCOND nodes. In the case of
|
2009-08-11 20:47:22 +00:00
|
|
|
/// BRCOND the argument passed is MVT::Other since there are no other
|
2009-01-05 21:33:06 +00:00
|
|
|
/// operands to get a type hint from.
|
2011-09-06 19:07:46 +00:00
|
|
|
virtual EVT getSetCCResultType(EVT VT) const;
|
2005-04-07 19:41:18 +00:00
|
|
|
|
2010-06-23 15:49:18 +00:00
|
|
|
/// getCmpLibcallReturnType - Return the ValueType for comparison
|
2009-12-28 02:40:33 +00:00
|
|
|
/// libcalls. Comparions libcalls include floating point comparion calls,
|
|
|
|
/// and Ordered/Unordered check calls on floating point numbers.
|
2010-06-23 15:49:18 +00:00
|
|
|
virtual
|
2009-12-28 02:40:33 +00:00
|
|
|
MVT::SimpleValueType getCmpLibcallReturnType() const;
|
|
|
|
|
2008-11-23 15:47:28 +00:00
|
|
|
/// getBooleanContents - For targets without i1 registers, this gives the
|
|
|
|
/// nature of the high-bits of boolean values held in types wider than i1.
|
|
|
|
/// "Boolean values" are special true/false values produced by nodes like
|
|
|
|
/// SETCC and consumed (as the condition) by nodes like SELECT and BRCOND.
|
|
|
|
/// Not to be confused with general values promoted from i1.
|
2011-09-06 19:07:46 +00:00
|
|
|
/// Some cpus distinguish between vectors of boolean and scalars; the isVec
|
|
|
|
/// parameter selects between the two kinds. For example on X86 a scalar
|
|
|
|
/// boolean should be zero extended from i1, while the elements of a vector
|
|
|
|
/// of booleans should be sign extended from i1.
|
|
|
|
BooleanContent getBooleanContents(bool isVec) const {
|
|
|
|
return isVec ? BooleanVectorContents : BooleanContents;
|
|
|
|
}
|
2005-04-07 19:41:18 +00:00
|
|
|
|
2006-01-25 09:09:02 +00:00
|
|
|
/// getSchedulingPreference - Return target scheduling preference.
|
2010-05-19 20:19:50 +00:00
|
|
|
Sched::Preference getSchedulingPreference() const {
|
2006-01-25 09:09:02 +00:00
|
|
|
return SchedPreferenceInfo;
|
|
|
|
}
|
|
|
|
|
2010-05-20 23:26:43 +00:00
|
|
|
/// getSchedulingPreference - Some scheduler, e.g. hybrid, can switch to
|
|
|
|
/// different scheduling heuristics for different nodes. This function returns
|
|
|
|
/// the preference (or none) for the given node.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual Sched::Preference getSchedulingPreference(SDNode *) const {
|
2010-05-20 23:26:43 +00:00
|
|
|
return Sched::None;
|
|
|
|
}
|
|
|
|
|
2005-01-16 23:59:30 +00:00
|
|
|
/// getRegClassFor - Return the register class that should be used for the
|
2010-05-15 02:18:07 +00:00
|
|
|
/// specified value type.
|
2012-12-13 06:34:11 +00:00
|
|
|
virtual const TargetRegisterClass *getRegClassFor(MVT VT) const {
|
|
|
|
const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy];
|
2005-01-07 07:44:22 +00:00
|
|
|
assert(RC && "This value type is not natively supported!");
|
|
|
|
return RC;
|
|
|
|
}
|
Disable some DAG combiner optimizations that may be
wrong for volatile loads and stores. In fact this
is almost all of them! There are three types of
problems: (1) it is wrong to change the width of
a volatile memory access. These may be used to
do memory mapped i/o, in which case a load can have
an effect even if the result is not used. Consider
loading an i32 but only using the lower 8 bits. It
is wrong to change this into a load of an i8, because
you are no longer tickling the other three bytes. It
is also unwise to make a load/store wider. For
example, changing an i16 load into an i32 load is
wrong no matter how aligned things are, since the
fact of loading an additional 2 bytes can have
i/o side-effects. (2) it is wrong to change the
number of volatile load/stores: they may be counted
by the hardware. (3) it is wrong to change a volatile
load/store that requires one memory access into one
that requires several. For example on x86-32, you
can store a double in one processor operation, but to
store an i64 requires two (two i32 stores). In a
multi-threaded program you may want to bitcast an i64
to a double and store as a double because that will
occur atomically, and be indivisible to other threads.
So it would be wrong to convert the store-of-double
into a store of an i64, because this will become two
i32 stores - no longer atomic. My policy here is
to say that the number of processor operations for
an illegal operation is undefined. So it is alright
to change a store of an i64 (requires at least two
stores; but could be validly lowered to memcpy for
example) into a store of double (one processor op).
In short, if the new store is legal and has the same
size then I say that the transform is ok. It would
also be possible to say that transforms are always
ok if before they were illegal, whether after they
are illegal or not, but that's more awkward to do
and I doubt it buys us anything much.
However this exposed an interesting thing - on x86-32
a store of i64 is considered legal! That is because
operations are marked legal by default, regardless of
whether the type is legal or not. In some ways this
is clever: before type legalization this means that
operations on illegal types are considered legal;
after type legalization there are no illegal types
so now operations are only legal if they really are.
But I consider this to be too cunning for mere mortals.
Better to do things explicitly by testing AfterLegalize.
So I have changed things so that operations with illegal
types are considered illegal - indeed they can never
map to a machine operation. However this means that
the DAG combiner is more conservative because before
it was "accidentally" performing transforms where the
type was illegal because the operation was nonetheless
marked legal. So in a few such places I added a check
on AfterLegalize, which I suppose was actually just
forgotten before. This causes the DAG combiner to do
slightly more than it used to, which resulted in the X86
backend blowing up because it got a slightly surprising
node it wasn't expecting, so I tweaked it.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52254 91177308-0d34-0410-b5e6-96231b3b80d8
2008-06-13 19:07:40 +00:00
|
|
|
|
2010-07-19 18:47:01 +00:00
|
|
|
/// getRepRegClassFor - Return the 'representative' register class for the
|
|
|
|
/// specified value type. The 'representative' register class is the largest
|
|
|
|
/// legal super-reg register class for the register class of the value type.
|
|
|
|
/// For example, on i386 the rep register class for i8, i16, and i32 are GR32;
|
|
|
|
/// while the rep register class is GR64 on x86_64.
|
2012-12-13 18:45:35 +00:00
|
|
|
virtual const TargetRegisterClass *getRepRegClassFor(MVT VT) const {
|
|
|
|
const TargetRegisterClass *RC = RepRegClassForVT[VT.SimpleTy];
|
2010-07-19 18:47:01 +00:00
|
|
|
return RC;
|
|
|
|
}
|
|
|
|
|
2010-07-21 06:09:07 +00:00
|
|
|
/// getRepRegClassCostFor - Return the cost of the 'representative' register
|
|
|
|
/// class for the specified value type.
|
2012-12-11 11:14:33 +00:00
|
|
|
virtual uint8_t getRepRegClassCostFor(EVT VT) const {
|
|
|
|
assert(VT.isSimple() && "getRepRegClassCostFor called on illegal type!");
|
|
|
|
return RepRegClassCostForVT[VT.getSimpleVT().SimpleTy];
|
2010-07-21 06:09:07 +00:00
|
|
|
}
|
|
|
|
|
2005-08-24 16:34:59 +00:00
|
|
|
/// isTypeLegal - Return true if the target has native support for the
|
2005-01-16 01:10:58 +00:00
|
|
|
/// specified value type. This means that it has a register that directly
|
|
|
|
/// holds it without promotions or expansions.
|
2009-08-10 22:56:29 +00:00
|
|
|
bool isTypeLegal(EVT VT) const {
|
2008-06-06 12:08:01 +00:00
|
|
|
assert(!VT.isSimple() ||
|
2009-08-11 20:47:22 +00:00
|
|
|
(unsigned)VT.getSimpleVT().SimpleTy < array_lengthof(RegClassForVT));
|
|
|
|
return VT.isSimple() && RegClassForVT[VT.getSimpleVT().SimpleTy] != 0;
|
2005-01-07 07:44:22 +00:00
|
|
|
}
|
2005-01-16 01:10:58 +00:00
|
|
|
|
2006-01-29 08:40:37 +00:00
|
|
|
class ValueTypeActionImpl {
|
2011-05-28 17:57:14 +00:00
|
|
|
/// ValueTypeActions - For each value type, keep a LegalizeTypeAction enum
|
2010-05-14 07:37:40 +00:00
|
|
|
/// that indicates how instruction selection should deal with the type.
|
|
|
|
uint8_t ValueTypeActions[MVT::LAST_VALUETYPE];
|
2011-02-25 21:41:48 +00:00
|
|
|
|
2006-01-29 08:40:37 +00:00
|
|
|
public:
|
|
|
|
ValueTypeActionImpl() {
|
2010-04-16 20:08:45 +00:00
|
|
|
std::fill(ValueTypeActions, array_endof(ValueTypeActions), 0);
|
2006-01-29 08:40:37 +00:00
|
|
|
}
|
2011-02-25 21:41:48 +00:00
|
|
|
|
2011-05-28 17:57:14 +00:00
|
|
|
LegalizeTypeAction getTypeAction(MVT VT) const {
|
|
|
|
return (LegalizeTypeAction)ValueTypeActions[VT.SimpleTy];
|
2006-01-29 08:40:37 +00:00
|
|
|
}
|
2011-02-25 21:41:48 +00:00
|
|
|
|
2012-12-13 20:42:43 +00:00
|
|
|
void setTypeAction(MVT VT, LegalizeTypeAction Action) {
|
|
|
|
unsigned I = VT.SimpleTy;
|
2010-05-14 07:37:40 +00:00
|
|
|
ValueTypeActions[I] = Action;
|
2006-01-29 08:40:37 +00:00
|
|
|
}
|
|
|
|
};
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-01-29 08:40:37 +00:00
|
|
|
const ValueTypeActionImpl &getValueTypeActions() const {
|
|
|
|
return ValueTypeActions;
|
|
|
|
}
|
Disable some DAG combiner optimizations that may be
wrong for volatile loads and stores. In fact this
is almost all of them! There are three types of
problems: (1) it is wrong to change the width of
a volatile memory access. These may be used to
do memory mapped i/o, in which case a load can have
an effect even if the result is not used. Consider
loading an i32 but only using the lower 8 bits. It
is wrong to change this into a load of an i8, because
you are no longer tickling the other three bytes. It
is also unwise to make a load/store wider. For
example, changing an i16 load into an i32 load is
wrong no matter how aligned things are, since the
fact of loading an additional 2 bytes can have
i/o side-effects. (2) it is wrong to change the
number of volatile load/stores: they may be counted
by the hardware. (3) it is wrong to change a volatile
load/store that requires one memory access into one
that requires several. For example on x86-32, you
can store a double in one processor operation, but to
store an i64 requires two (two i32 stores). In a
multi-threaded program you may want to bitcast an i64
to a double and store as a double because that will
occur atomically, and be indivisible to other threads.
So it would be wrong to convert the store-of-double
into a store of an i64, because this will become two
i32 stores - no longer atomic. My policy here is
to say that the number of processor operations for
an illegal operation is undefined. So it is alright
to change a store of an i64 (requires at least two
stores; but could be validly lowered to memcpy for
example) into a store of double (one processor op).
In short, if the new store is legal and has the same
size then I say that the transform is ok. It would
also be possible to say that transforms are always
ok if before they were illegal, whether after they
are illegal or not, but that's more awkward to do
and I doubt it buys us anything much.
However this exposed an interesting thing - on x86-32
a store of i64 is considered legal! That is because
operations are marked legal by default, regardless of
whether the type is legal or not. In some ways this
is clever: before type legalization this means that
operations on illegal types are considered legal;
after type legalization there are no illegal types
so now operations are only legal if they really are.
But I consider this to be too cunning for mere mortals.
Better to do things explicitly by testing AfterLegalize.
So I have changed things so that operations with illegal
types are considered illegal - indeed they can never
map to a machine operation. However this means that
the DAG combiner is more conservative because before
it was "accidentally" performing transforms where the
type was illegal because the operation was nonetheless
marked legal. So in a few such places I added a check
on AfterLegalize, which I suppose was actually just
forgotten before. This causes the DAG combiner to do
slightly more than it used to, which resulted in the X86
backend blowing up because it got a slightly surprising
node it wasn't expecting, so I tweaked it.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52254 91177308-0d34-0410-b5e6-96231b3b80d8
2008-06-13 19:07:40 +00:00
|
|
|
|
2005-01-16 01:10:58 +00:00
|
|
|
/// getTypeAction - Return how we should legalize values of this type, either
|
2005-01-16 07:27:49 +00:00
|
|
|
/// it is already legal (return 'Legal') or we need to promote it to a larger
|
|
|
|
/// type (return 'Promote'), or we need to expand it into multiple registers
|
|
|
|
/// of smaller integer type (return 'Expand'). 'Custom' is not an option.
|
2011-05-28 17:57:14 +00:00
|
|
|
LegalizeTypeAction getTypeAction(LLVMContext &Context, EVT VT) const {
|
2011-05-27 21:03:13 +00:00
|
|
|
return getTypeConversion(Context, VT).first;
|
2005-01-16 01:10:58 +00:00
|
|
|
}
|
2011-05-28 17:57:14 +00:00
|
|
|
LegalizeTypeAction getTypeAction(MVT VT) const {
|
2010-08-25 23:05:45 +00:00
|
|
|
return ValueTypeActions.getTypeAction(VT);
|
|
|
|
}
|
2011-02-25 21:41:48 +00:00
|
|
|
|
2005-01-16 01:10:58 +00:00
|
|
|
/// getTypeToTransformTo - For types supported by the target, this is an
|
|
|
|
/// identity function. For types that must be promoted to larger types, this
|
2006-12-13 06:12:35 +00:00
|
|
|
/// returns the larger type to promote to. For integer types that are larger
|
|
|
|
/// than the largest integer register, this contains one step in the expansion
|
|
|
|
/// to get to the smaller register. For illegal floating point types, this
|
|
|
|
/// returns the integer type to transform to.
|
2009-08-12 00:36:31 +00:00
|
|
|
EVT getTypeToTransformTo(LLVMContext &Context, EVT VT) const {
|
2011-06-08 23:55:35 +00:00
|
|
|
return getTypeConversion(Context, VT).second;
|
2005-01-16 01:10:58 +00:00
|
|
|
}
|
2007-10-16 09:56:48 +00:00
|
|
|
|
2010-03-23 22:44:42 +00:00
|
|
|
/// getTypeToExpandTo - For types supported by the target, this is an
|
|
|
|
/// identity function. For types that must be expanded (i.e. integer types
|
|
|
|
/// that are larger than the largest integer register or illegal floating
|
|
|
|
/// point types), this returns the largest legal type it will be expanded to.
|
|
|
|
EVT getTypeToExpandTo(LLVMContext &Context, EVT VT) const {
|
|
|
|
assert(!VT.isVector());
|
|
|
|
while (true) {
|
2011-05-27 21:03:13 +00:00
|
|
|
switch (getTypeAction(Context, VT)) {
|
2011-08-04 16:01:54 +00:00
|
|
|
case TypeLegal:
|
2010-03-23 22:44:42 +00:00
|
|
|
return VT;
|
2011-08-04 16:01:54 +00:00
|
|
|
case TypeExpandInteger:
|
2010-03-23 22:44:42 +00:00
|
|
|
VT = getTypeToTransformTo(Context, VT);
|
|
|
|
break;
|
|
|
|
default:
|
2012-02-05 22:14:15 +00:00
|
|
|
llvm_unreachable("Type is not legal nor is it to be expanded!");
|
2010-03-23 22:44:42 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2007-02-15 03:39:18 +00:00
|
|
|
/// getVectorTypeBreakdown - Vector types are broken down into some number of
|
2009-08-10 22:56:29 +00:00
|
|
|
/// legal first class types. For example, EVT::v8f32 maps to 2 EVT::v4f32
|
|
|
|
/// with Altivec or SSE1, or 8 promoted EVT::f64 values with the X86 FP stack.
|
|
|
|
/// Similarly, EVT::v2i64 turns into 4 EVT::i32 values with both PPC and X86.
|
2006-03-31 00:28:23 +00:00
|
|
|
///
|
2006-03-31 00:46:36 +00:00
|
|
|
/// This method returns the number of registers needed, and the VT for each
|
2007-06-28 23:29:44 +00:00
|
|
|
/// register. It also returns the VT and quantity of the intermediate values
|
|
|
|
/// before they are promoted/expanded.
|
2006-03-31 00:28:23 +00:00
|
|
|
///
|
2009-08-12 00:36:31 +00:00
|
|
|
unsigned getVectorTypeBreakdown(LLVMContext &Context, EVT VT,
|
2009-08-10 22:56:29 +00:00
|
|
|
EVT &IntermediateVT,
|
2007-06-28 23:29:44 +00:00
|
|
|
unsigned &NumIntermediates,
|
2012-12-11 11:14:33 +00:00
|
|
|
EVT &RegisterVT) const;
|
2008-11-01 20:24:53 +00:00
|
|
|
|
|
|
|
/// getTgtMemIntrinsic: Given an intrinsic, checks if on the target the
|
|
|
|
/// intrinsic will need to map to a MemIntrinsicNode (touches memory). If
|
|
|
|
/// this is the case, it returns true and store the intrinsic
|
|
|
|
/// information into the IntrinsicInfo that was passed to the function.
|
2010-06-23 15:49:18 +00:00
|
|
|
struct IntrinsicInfo {
|
2008-11-01 20:24:53 +00:00
|
|
|
unsigned opc; // target opcode
|
2009-08-10 22:56:29 +00:00
|
|
|
EVT memVT; // memory VT
|
2008-11-01 20:24:53 +00:00
|
|
|
const Value* ptrVal; // value representing memory location
|
2010-06-23 15:49:18 +00:00
|
|
|
int offset; // offset off of ptrVal
|
2008-11-01 20:24:53 +00:00
|
|
|
unsigned align; // alignment
|
|
|
|
bool vol; // is volatile?
|
|
|
|
bool readMem; // reads memory?
|
|
|
|
bool writeMem; // writes memory?
|
2010-03-17 18:43:25 +00:00
|
|
|
};
|
2008-11-01 20:24:53 +00:00
|
|
|
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool getTgtMemIntrinsic(IntrinsicInfo &, const CallInst &,
|
|
|
|
unsigned /*Intrinsic*/) const {
|
2008-11-01 20:24:53 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2009-10-27 19:56:55 +00:00
|
|
|
/// isFPImmLegal - Returns true if the target can instruction select the
|
|
|
|
/// specified FP immediate natively. If false, the legalizer will materialize
|
|
|
|
/// the FP immediate as a load from a constant pool.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool isFPImmLegal(const APFloat &/*Imm*/, EVT /*VT*/) const {
|
2009-10-27 19:56:55 +00:00
|
|
|
return false;
|
2005-01-07 07:44:22 +00:00
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-04-04 17:21:22 +00:00
|
|
|
/// isShuffleMaskLegal - Targets can use this to indicate that they only
|
|
|
|
/// support *some* VECTOR_SHUFFLE operations, those with specific masks.
|
|
|
|
/// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
|
|
|
|
/// are assumed to be legal.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &/*Mask*/,
|
|
|
|
EVT /*VT*/) const {
|
2006-04-04 17:21:22 +00:00
|
|
|
return true;
|
|
|
|
}
|
2005-01-16 07:27:49 +00:00
|
|
|
|
2010-02-10 23:37:45 +00:00
|
|
|
/// canOpTrap - Returns true if the operation can trap for the value type.
|
|
|
|
/// VT must be a legal type. By default, we optimistically assume most
|
|
|
|
/// operations don't trap except for divide and remainder.
|
|
|
|
virtual bool canOpTrap(unsigned Op, EVT VT) const;
|
|
|
|
|
2006-04-20 08:54:13 +00:00
|
|
|
/// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is
|
|
|
|
/// used by Targets can use this to indicate if there is a suitable
|
|
|
|
/// VECTOR_SHUFFLE that can be used to replace a VAND with a constant
|
|
|
|
/// pool entry.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &/*Mask*/,
|
|
|
|
EVT /*VT*/) const {
|
2006-04-20 08:54:13 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2005-08-24 16:34:59 +00:00
|
|
|
/// getOperationAction - Return how this operation should be treated: either
|
|
|
|
/// it is legal, needs to be promoted to a larger size, needs to be
|
|
|
|
/// expanded to some other code sequence, or the target has a custom expander
|
|
|
|
/// for it.
|
2009-08-10 22:56:29 +00:00
|
|
|
LegalizeAction getOperationAction(unsigned Op, EVT VT) const {
|
2008-06-06 12:08:01 +00:00
|
|
|
if (VT.isExtended()) return Expand;
|
2012-08-08 23:31:14 +00:00
|
|
|
// If a target-specific SDNode requires legalization, require the target
|
|
|
|
// to provide custom legalization for it.
|
|
|
|
if (Op > array_lengthof(OpActions[0])) return Custom;
|
2009-08-11 20:47:22 +00:00
|
|
|
unsigned I = (unsigned) VT.getSimpleVT().SimpleTy;
|
2010-05-14 07:37:40 +00:00
|
|
|
return (LegalizeAction)OpActions[I][Op];
|
2005-01-07 07:44:22 +00:00
|
|
|
}
|
Disable some DAG combiner optimizations that may be
wrong for volatile loads and stores. In fact this
is almost all of them! There are three types of
problems: (1) it is wrong to change the width of
a volatile memory access. These may be used to
do memory mapped i/o, in which case a load can have
an effect even if the result is not used. Consider
loading an i32 but only using the lower 8 bits. It
is wrong to change this into a load of an i8, because
you are no longer tickling the other three bytes. It
is also unwise to make a load/store wider. For
example, changing an i16 load into an i32 load is
wrong no matter how aligned things are, since the
fact of loading an additional 2 bytes can have
i/o side-effects. (2) it is wrong to change the
number of volatile load/stores: they may be counted
by the hardware. (3) it is wrong to change a volatile
load/store that requires one memory access into one
that requires several. For example on x86-32, you
can store a double in one processor operation, but to
store an i64 requires two (two i32 stores). In a
multi-threaded program you may want to bitcast an i64
to a double and store as a double because that will
occur atomically, and be indivisible to other threads.
So it would be wrong to convert the store-of-double
into a store of an i64, because this will become two
i32 stores - no longer atomic. My policy here is
to say that the number of processor operations for
an illegal operation is undefined. So it is alright
to change a store of an i64 (requires at least two
stores; but could be validly lowered to memcpy for
example) into a store of double (one processor op).
In short, if the new store is legal and has the same
size then I say that the transform is ok. It would
also be possible to say that transforms are always
ok if before they were illegal, whether after they
are illegal or not, but that's more awkward to do
and I doubt it buys us anything much.
However this exposed an interesting thing - on x86-32
a store of i64 is considered legal! That is because
operations are marked legal by default, regardless of
whether the type is legal or not. In some ways this
is clever: before type legalization this means that
operations on illegal types are considered legal;
after type legalization there are no illegal types
so now operations are only legal if they really are.
But I consider this to be too cunning for mere mortals.
Better to do things explicitly by testing AfterLegalize.
So I have changed things so that operations with illegal
types are considered illegal - indeed they can never
map to a machine operation. However this means that
the DAG combiner is more conservative because before
it was "accidentally" performing transforms where the
type was illegal because the operation was nonetheless
marked legal. So in a few such places I added a check
on AfterLegalize, which I suppose was actually just
forgotten before. This causes the DAG combiner to do
slightly more than it used to, which resulted in the X86
backend blowing up because it got a slightly surprising
node it wasn't expecting, so I tweaked it.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52254 91177308-0d34-0410-b5e6-96231b3b80d8
2008-06-13 19:07:40 +00:00
|
|
|
|
2009-01-28 17:46:25 +00:00
|
|
|
/// isOperationLegalOrCustom - Return true if the specified operation is
|
|
|
|
/// legal on this target or can be made legal with custom lowering. This
|
|
|
|
/// is used to help guide high-level lowering decisions.
|
2009-08-10 22:56:29 +00:00
|
|
|
bool isOperationLegalOrCustom(unsigned Op, EVT VT) const {
|
2009-08-11 20:47:22 +00:00
|
|
|
return (VT == MVT::Other || isTypeLegal(VT)) &&
|
2009-01-28 17:46:25 +00:00
|
|
|
(getOperationAction(Op, VT) == Legal ||
|
|
|
|
getOperationAction(Op, VT) == Custom);
|
|
|
|
}
|
|
|
|
|
2012-10-26 23:49:28 +00:00
|
|
|
/// isOperationExpand - Return true if the specified operation is illegal on
|
|
|
|
/// this target or unlikely to be made legal with custom lowering. This is
|
|
|
|
/// used to help guide high-level lowering decisions.
|
|
|
|
bool isOperationExpand(unsigned Op, EVT VT) const {
|
|
|
|
return (!isTypeLegal(VT) || getOperationAction(Op, VT) == Expand);
|
|
|
|
}
|
|
|
|
|
2005-08-24 16:34:59 +00:00
|
|
|
/// isOperationLegal - Return true if the specified operation is legal on this
|
|
|
|
/// target.
|
2009-08-10 22:56:29 +00:00
|
|
|
bool isOperationLegal(unsigned Op, EVT VT) const {
|
2009-08-11 20:47:22 +00:00
|
|
|
return (VT == MVT::Other || isTypeLegal(VT)) &&
|
2009-01-28 17:46:25 +00:00
|
|
|
getOperationAction(Op, VT) == Legal;
|
2005-01-16 07:27:49 +00:00
|
|
|
}
|
Disable some DAG combiner optimizations that may be
wrong for volatile loads and stores. In fact this
is almost all of them! There are three types of
problems: (1) it is wrong to change the width of
a volatile memory access. These may be used to
do memory mapped i/o, in which case a load can have
an effect even if the result is not used. Consider
loading an i32 but only using the lower 8 bits. It
is wrong to change this into a load of an i8, because
you are no longer tickling the other three bytes. It
is also unwise to make a load/store wider. For
example, changing an i16 load into an i32 load is
wrong no matter how aligned things are, since the
fact of loading an additional 2 bytes can have
i/o side-effects. (2) it is wrong to change the
number of volatile load/stores: they may be counted
by the hardware. (3) it is wrong to change a volatile
load/store that requires one memory access into one
that requires several. For example on x86-32, you
can store a double in one processor operation, but to
store an i64 requires two (two i32 stores). In a
multi-threaded program you may want to bitcast an i64
to a double and store as a double because that will
occur atomically, and be indivisible to other threads.
So it would be wrong to convert the store-of-double
into a store of an i64, because this will become two
i32 stores - no longer atomic. My policy here is
to say that the number of processor operations for
an illegal operation is undefined. So it is alright
to change a store of an i64 (requires at least two
stores; but could be validly lowered to memcpy for
example) into a store of double (one processor op).
In short, if the new store is legal and has the same
size then I say that the transform is ok. It would
also be possible to say that transforms are always
ok if before they were illegal, whether after they
are illegal or not, but that's more awkward to do
and I doubt it buys us anything much.
However this exposed an interesting thing - on x86-32
a store of i64 is considered legal! That is because
operations are marked legal by default, regardless of
whether the type is legal or not. In some ways this
is clever: before type legalization this means that
operations on illegal types are considered legal;
after type legalization there are no illegal types
so now operations are only legal if they really are.
But I consider this to be too cunning for mere mortals.
Better to do things explicitly by testing AfterLegalize.
So I have changed things so that operations with illegal
types are considered illegal - indeed they can never
map to a machine operation. However this means that
the DAG combiner is more conservative because before
it was "accidentally" performing transforms where the
type was illegal because the operation was nonetheless
marked legal. So in a few such places I added a check
on AfterLegalize, which I suppose was actually just
forgotten before. This causes the DAG combiner to do
slightly more than it used to, which resulted in the X86
backend blowing up because it got a slightly surprising
node it wasn't expecting, so I tweaked it.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52254 91177308-0d34-0410-b5e6-96231b3b80d8
2008-06-13 19:07:40 +00:00
|
|
|
|
2008-10-14 21:26:46 +00:00
|
|
|
/// getLoadExtAction - Return how this load with extension should be treated:
|
2006-10-04 00:56:09 +00:00
|
|
|
/// either it is legal, needs to be promoted to a larger size, needs to be
|
|
|
|
/// expanded to some other code sequence, or the target has a custom expander
|
|
|
|
/// for it.
|
2012-12-11 11:14:33 +00:00
|
|
|
LegalizeAction getLoadExtAction(unsigned ExtType, EVT VT) const {
|
|
|
|
assert(ExtType < ISD::LAST_LOADEXT_TYPE &&
|
|
|
|
VT.getSimpleVT() < MVT::LAST_VALUETYPE &&
|
2008-06-06 12:08:01 +00:00
|
|
|
"Table isn't big enough!");
|
2012-12-11 11:14:33 +00:00
|
|
|
return (LegalizeAction)LoadExtActions[VT.getSimpleVT().SimpleTy][ExtType];
|
2006-10-04 00:56:09 +00:00
|
|
|
}
|
2008-06-12 09:01:30 +00:00
|
|
|
|
2008-10-14 21:26:46 +00:00
|
|
|
/// isLoadExtLegal - Return true if the specified load with extension is legal
|
2006-10-13 21:14:26 +00:00
|
|
|
/// on this target.
|
2010-05-13 23:01:26 +00:00
|
|
|
bool isLoadExtLegal(unsigned ExtType, EVT VT) const {
|
2012-12-11 11:14:33 +00:00
|
|
|
return VT.isSimple() && getLoadExtAction(ExtType, VT) == Legal;
|
2006-10-04 00:56:09 +00:00
|
|
|
}
|
2008-06-12 09:01:30 +00:00
|
|
|
|
2008-01-17 19:59:44 +00:00
|
|
|
/// getTruncStoreAction - Return how this store with truncation should be
|
|
|
|
/// treated: either it is legal, needs to be promoted to a larger size, needs
|
|
|
|
/// to be expanded to some other code sequence, or the target has a custom
|
|
|
|
/// expander for it.
|
2012-12-11 11:14:33 +00:00
|
|
|
LegalizeAction getTruncStoreAction(EVT ValVT, EVT MemVT) const {
|
|
|
|
assert(ValVT.getSimpleVT() < MVT::LAST_VALUETYPE &&
|
|
|
|
MemVT.getSimpleVT() < MVT::LAST_VALUETYPE &&
|
2008-06-06 12:08:01 +00:00
|
|
|
"Table isn't big enough!");
|
2012-12-11 11:14:33 +00:00
|
|
|
return (LegalizeAction)TruncStoreActions[ValVT.getSimpleVT().SimpleTy]
|
|
|
|
[MemVT.getSimpleVT().SimpleTy];
|
2006-10-13 21:14:26 +00:00
|
|
|
}
|
2008-06-12 09:01:30 +00:00
|
|
|
|
2008-01-17 19:59:44 +00:00
|
|
|
/// isTruncStoreLegal - Return true if the specified store with truncation is
|
2006-10-13 21:14:26 +00:00
|
|
|
/// legal on this target.
|
2009-08-10 22:56:29 +00:00
|
|
|
bool isTruncStoreLegal(EVT ValVT, EVT MemVT) const {
|
Disable some DAG combiner optimizations that may be
wrong for volatile loads and stores. In fact this
is almost all of them! There are three types of
problems: (1) it is wrong to change the width of
a volatile memory access. These may be used to
do memory mapped i/o, in which case a load can have
an effect even if the result is not used. Consider
loading an i32 but only using the lower 8 bits. It
is wrong to change this into a load of an i8, because
you are no longer tickling the other three bytes. It
is also unwise to make a load/store wider. For
example, changing an i16 load into an i32 load is
wrong no matter how aligned things are, since the
fact of loading an additional 2 bytes can have
i/o side-effects. (2) it is wrong to change the
number of volatile load/stores: they may be counted
by the hardware. (3) it is wrong to change a volatile
load/store that requires one memory access into one
that requires several. For example on x86-32, you
can store a double in one processor operation, but to
store an i64 requires two (two i32 stores). In a
multi-threaded program you may want to bitcast an i64
to a double and store as a double because that will
occur atomically, and be indivisible to other threads.
So it would be wrong to convert the store-of-double
into a store of an i64, because this will become two
i32 stores - no longer atomic. My policy here is
to say that the number of processor operations for
an illegal operation is undefined. So it is alright
to change a store of an i64 (requires at least two
stores; but could be validly lowered to memcpy for
example) into a store of double (one processor op).
In short, if the new store is legal and has the same
size then I say that the transform is ok. It would
also be possible to say that transforms are always
ok if before they were illegal, whether after they
are illegal or not, but that's more awkward to do
and I doubt it buys us anything much.
However this exposed an interesting thing - on x86-32
a store of i64 is considered legal! That is because
operations are marked legal by default, regardless of
whether the type is legal or not. In some ways this
is clever: before type legalization this means that
operations on illegal types are considered legal;
after type legalization there are no illegal types
so now operations are only legal if they really are.
But I consider this to be too cunning for mere mortals.
Better to do things explicitly by testing AfterLegalize.
So I have changed things so that operations with illegal
types are considered illegal - indeed they can never
map to a machine operation. However this means that
the DAG combiner is more conservative because before
it was "accidentally" performing transforms where the
type was illegal because the operation was nonetheless
marked legal. So in a few such places I added a check
on AfterLegalize, which I suppose was actually just
forgotten before. This causes the DAG combiner to do
slightly more than it used to, which resulted in the X86
backend blowing up because it got a slightly surprising
node it wasn't expecting, so I tweaked it.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52254 91177308-0d34-0410-b5e6-96231b3b80d8
2008-06-13 19:07:40 +00:00
|
|
|
return isTypeLegal(ValVT) && MemVT.isSimple() &&
|
2012-12-11 11:14:33 +00:00
|
|
|
getTruncStoreAction(ValVT, MemVT) == Legal;
|
2006-10-13 21:14:26 +00:00
|
|
|
}
|
|
|
|
|
2006-11-09 18:56:43 +00:00
|
|
|
/// getIndexedLoadAction - Return how the indexed load should be treated:
|
|
|
|
/// either it is legal, needs to be promoted to a larger size, needs to be
|
|
|
|
/// expanded to some other code sequence, or the target has a custom expander
|
|
|
|
/// for it.
|
|
|
|
LegalizeAction
|
2012-12-11 11:14:33 +00:00
|
|
|
getIndexedLoadAction(unsigned IdxMode, EVT VT) const {
|
|
|
|
assert(IdxMode < ISD::LAST_INDEXED_MODE &&
|
|
|
|
VT.getSimpleVT() < MVT::LAST_VALUETYPE &&
|
2008-01-25 10:20:53 +00:00
|
|
|
"Table isn't big enough!");
|
2012-12-11 11:14:33 +00:00
|
|
|
unsigned Ty = (unsigned)VT.getSimpleVT().SimpleTy;
|
2010-05-13 23:01:26 +00:00
|
|
|
return (LegalizeAction)((IndexedModeActions[Ty][IdxMode] & 0xf0) >> 4);
|
2006-11-09 18:56:43 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// isIndexedLoadLegal - Return true if the specified indexed load is legal
|
|
|
|
/// on this target.
|
2009-08-10 22:56:29 +00:00
|
|
|
bool isIndexedLoadLegal(unsigned IdxMode, EVT VT) const {
|
2008-06-12 09:01:30 +00:00
|
|
|
return VT.isSimple() &&
|
2012-12-11 11:14:33 +00:00
|
|
|
(getIndexedLoadAction(IdxMode, VT) == Legal ||
|
|
|
|
getIndexedLoadAction(IdxMode, VT) == Custom);
|
2006-11-09 18:56:43 +00:00
|
|
|
}
|
2008-06-12 09:01:30 +00:00
|
|
|
|
2006-11-09 18:56:43 +00:00
|
|
|
/// getIndexedStoreAction - Return how the indexed store should be treated:
|
|
|
|
/// either it is legal, needs to be promoted to a larger size, needs to be
|
|
|
|
/// expanded to some other code sequence, or the target has a custom expander
|
|
|
|
/// for it.
|
|
|
|
LegalizeAction
|
2012-12-11 11:14:33 +00:00
|
|
|
getIndexedStoreAction(unsigned IdxMode, EVT VT) const {
|
|
|
|
assert(IdxMode < ISD::LAST_INDEXED_MODE &&
|
|
|
|
VT.getSimpleVT() < MVT::LAST_VALUETYPE &&
|
2008-01-25 10:20:53 +00:00
|
|
|
"Table isn't big enough!");
|
2012-12-11 11:14:33 +00:00
|
|
|
unsigned Ty = (unsigned)VT.getSimpleVT().SimpleTy;
|
2010-05-13 23:01:26 +00:00
|
|
|
return (LegalizeAction)(IndexedModeActions[Ty][IdxMode] & 0x0f);
|
2010-06-23 15:49:18 +00:00
|
|
|
}
|
Disable some DAG combiner optimizations that may be
wrong for volatile loads and stores. In fact this
is almost all of them! There are three types of
problems: (1) it is wrong to change the width of
a volatile memory access. These may be used to
do memory mapped i/o, in which case a load can have
an effect even if the result is not used. Consider
loading an i32 but only using the lower 8 bits. It
is wrong to change this into a load of an i8, because
you are no longer tickling the other three bytes. It
is also unwise to make a load/store wider. For
example, changing an i16 load into an i32 load is
wrong no matter how aligned things are, since the
fact of loading an additional 2 bytes can have
i/o side-effects. (2) it is wrong to change the
number of volatile load/stores: they may be counted
by the hardware. (3) it is wrong to change a volatile
load/store that requires one memory access into one
that requires several. For example on x86-32, you
can store a double in one processor operation, but to
store an i64 requires two (two i32 stores). In a
multi-threaded program you may want to bitcast an i64
to a double and store as a double because that will
occur atomically, and be indivisible to other threads.
So it would be wrong to convert the store-of-double
into a store of an i64, because this will become two
i32 stores - no longer atomic. My policy here is
to say that the number of processor operations for
an illegal operation is undefined. So it is alright
to change a store of an i64 (requires at least two
stores; but could be validly lowered to memcpy for
example) into a store of double (one processor op).
In short, if the new store is legal and has the same
size then I say that the transform is ok. It would
also be possible to say that transforms are always
ok if before they were illegal, whether after they
are illegal or not, but that's more awkward to do
and I doubt it buys us anything much.
However this exposed an interesting thing - on x86-32
a store of i64 is considered legal! That is because
operations are marked legal by default, regardless of
whether the type is legal or not. In some ways this
is clever: before type legalization this means that
operations on illegal types are considered legal;
after type legalization there are no illegal types
so now operations are only legal if they really are.
But I consider this to be too cunning for mere mortals.
Better to do things explicitly by testing AfterLegalize.
So I have changed things so that operations with illegal
types are considered illegal - indeed they can never
map to a machine operation. However this means that
the DAG combiner is more conservative because before
it was "accidentally" performing transforms where the
type was illegal because the operation was nonetheless
marked legal. So in a few such places I added a check
on AfterLegalize, which I suppose was actually just
forgotten before. This causes the DAG combiner to do
slightly more than it used to, which resulted in the X86
backend blowing up because it got a slightly surprising
node it wasn't expecting, so I tweaked it.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@52254 91177308-0d34-0410-b5e6-96231b3b80d8
2008-06-13 19:07:40 +00:00
|
|
|
|
2006-11-09 18:56:43 +00:00
|
|
|
/// isIndexedStoreLegal - Return true if the specified indexed load is legal
|
|
|
|
/// on this target.
|
2009-08-10 22:56:29 +00:00
|
|
|
bool isIndexedStoreLegal(unsigned IdxMode, EVT VT) const {
|
2008-06-12 09:01:30 +00:00
|
|
|
return VT.isSimple() &&
|
2012-12-11 11:14:33 +00:00
|
|
|
(getIndexedStoreAction(IdxMode, VT) == Legal ||
|
|
|
|
getIndexedStoreAction(IdxMode, VT) == Custom);
|
2006-11-09 18:56:43 +00:00
|
|
|
}
|
2008-06-12 09:01:30 +00:00
|
|
|
|
2008-10-15 02:05:31 +00:00
|
|
|
/// getCondCodeAction - Return how the condition code should be treated:
|
|
|
|
/// either it is legal, needs to be expanded to some other code sequence,
|
|
|
|
/// or the target has a custom expander for it.
|
|
|
|
LegalizeAction
|
2012-12-11 11:14:33 +00:00
|
|
|
getCondCodeAction(ISD::CondCode CC, EVT VT) const {
|
2008-10-15 02:05:31 +00:00
|
|
|
assert((unsigned)CC < array_lengthof(CondCodeActions) &&
|
2012-12-11 11:14:33 +00:00
|
|
|
(unsigned)VT.getSimpleVT().SimpleTy < sizeof(CondCodeActions[0])*4 &&
|
2008-10-15 02:05:31 +00:00
|
|
|
"Table isn't big enough!");
|
2012-09-14 15:36:50 +00:00
|
|
|
/// The lower 5 bits of the SimpleTy index into Nth 2bit set from the 64bit
|
|
|
|
/// value and the upper 27 bits index into the second dimension of the
|
|
|
|
/// array to select what 64bit value to use.
|
2008-10-15 02:05:31 +00:00
|
|
|
LegalizeAction Action = (LegalizeAction)
|
2012-12-11 11:14:33 +00:00
|
|
|
((CondCodeActions[CC][VT.getSimpleVT().SimpleTy >> 5]
|
|
|
|
>> (2*(VT.getSimpleVT().SimpleTy & 0x1F))) & 3);
|
2008-10-15 02:05:31 +00:00
|
|
|
assert(Action != Promote && "Can't promote condition code!");
|
|
|
|
return Action;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// isCondCodeLegal - Return true if the specified condition code is legal
|
|
|
|
/// on this target.
|
2012-12-11 11:14:33 +00:00
|
|
|
bool isCondCodeLegal(ISD::CondCode CC, EVT VT) const {
|
|
|
|
return getCondCodeAction(CC, VT) == Legal ||
|
|
|
|
getCondCodeAction(CC, VT) == Custom;
|
2008-10-15 02:05:31 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-01-16 07:27:49 +00:00
|
|
|
/// getTypeToPromoteTo - If the action for this operation is to promote, this
|
|
|
|
/// method returns the ValueType to promote to.
|
2012-12-11 11:14:33 +00:00
|
|
|
EVT getTypeToPromoteTo(unsigned Op, EVT VT) const {
|
2005-01-16 07:27:49 +00:00
|
|
|
assert(getOperationAction(Op, VT) == Promote &&
|
|
|
|
"This operation isn't promoted!");
|
2006-04-04 00:25:10 +00:00
|
|
|
|
|
|
|
// See if this has an explicit type specified.
|
2009-08-11 20:47:22 +00:00
|
|
|
std::map<std::pair<unsigned, MVT::SimpleValueType>,
|
|
|
|
MVT::SimpleValueType>::const_iterator PTTI =
|
2012-12-11 11:14:33 +00:00
|
|
|
PromoteToType.find(std::make_pair(Op, VT.getSimpleVT().SimpleTy));
|
2006-04-04 00:25:10 +00:00
|
|
|
if (PTTI != PromoteToType.end()) return PTTI->second;
|
2008-06-06 12:08:01 +00:00
|
|
|
|
|
|
|
assert((VT.isInteger() || VT.isFloatingPoint()) &&
|
2006-04-04 00:25:10 +00:00
|
|
|
"Cannot autopromote this type, add it with AddPromotedToType.");
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2012-12-11 11:14:33 +00:00
|
|
|
EVT NVT = VT;
|
2005-01-16 07:27:49 +00:00
|
|
|
do {
|
2012-12-11 11:14:33 +00:00
|
|
|
NVT = (MVT::SimpleValueType)(NVT.getSimpleVT().SimpleTy+1);
|
2009-08-11 20:47:22 +00:00
|
|
|
assert(NVT.isInteger() == VT.isInteger() && NVT != MVT::isVoid &&
|
2005-01-16 07:27:49 +00:00
|
|
|
"Didn't find type to promote to!");
|
2005-08-24 16:34:59 +00:00
|
|
|
} while (!isTypeLegal(NVT) ||
|
|
|
|
getOperationAction(Op, NVT) == Promote);
|
2005-01-16 07:27:49 +00:00
|
|
|
return NVT;
|
|
|
|
}
|
|
|
|
|
2009-08-10 22:56:29 +00:00
|
|
|
/// getValueType - Return the EVT corresponding to this LLVM type.
|
2007-04-28 05:42:38 +00:00
|
|
|
/// This is fixed by the LLVM operations except for the pointer size. If
|
2009-08-11 20:47:22 +00:00
|
|
|
/// AllowUnknown is true, this will return MVT::Other for types with no EVT
|
2007-04-28 05:42:38 +00:00
|
|
|
/// counterpart (e.g. structs), otherwise it will assert.
|
2011-07-18 04:54:35 +00:00
|
|
|
EVT getValueType(Type *Ty, bool AllowUnknown = false) const {
|
2011-12-05 06:29:09 +00:00
|
|
|
// Lower scalar pointers to native pointer types.
|
|
|
|
if (Ty->isPointerTy()) return PointerTy;
|
|
|
|
|
|
|
|
if (Ty->isVectorTy()) {
|
|
|
|
VectorType *VTy = cast<VectorType>(Ty);
|
|
|
|
Type *Elm = VTy->getElementType();
|
|
|
|
// Lower vectors of pointers to native pointer types.
|
|
|
|
if (Elm->isPointerTy())
|
|
|
|
Elm = EVT(PointerTy).getTypeForEVT(Ty->getContext());
|
|
|
|
return EVT::getVectorVT(Ty->getContext(), EVT::getEVT(Elm, false),
|
|
|
|
VTy->getNumElements());
|
|
|
|
}
|
|
|
|
return EVT::getEVT(Ty, AllowUnknown);
|
2007-03-31 04:05:24 +00:00
|
|
|
}
|
2012-12-13 06:34:11 +00:00
|
|
|
|
|
|
|
/// Return the MVT corresponding to this LLVM type. See getValueType.
|
|
|
|
MVT getSimpleValueType(Type *Ty, bool AllowUnknown = false) const {
|
|
|
|
return getValueType(Ty, AllowUnknown).getSimpleVT();
|
|
|
|
}
|
2005-04-21 20:59:05 +00:00
|
|
|
|
2008-01-23 23:17:41 +00:00
|
|
|
/// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
|
2008-02-28 22:31:51 +00:00
|
|
|
/// function arguments in the caller parameter area. This is the actual
|
|
|
|
/// alignment, not its logarithm.
|
2011-07-18 04:54:35 +00:00
|
|
|
virtual unsigned getByValTypeAlignment(Type *Ty) const;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2007-06-28 23:29:44 +00:00
|
|
|
/// getRegisterType - Return the type of registers that this ValueType will
|
|
|
|
/// eventually require.
|
2012-12-11 11:14:33 +00:00
|
|
|
EVT getRegisterType(MVT VT) const {
|
2009-08-12 00:36:31 +00:00
|
|
|
assert((unsigned)VT.SimpleTy < array_lengthof(RegisterTypeForVT));
|
|
|
|
return RegisterTypeForVT[VT.SimpleTy];
|
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2009-08-12 00:36:31 +00:00
|
|
|
/// getRegisterType - Return the type of registers that this ValueType will
|
|
|
|
/// eventually require.
|
2012-12-11 11:14:33 +00:00
|
|
|
EVT getRegisterType(LLVMContext &Context, EVT VT) const {
|
2008-06-06 12:08:01 +00:00
|
|
|
if (VT.isSimple()) {
|
2009-08-11 20:47:22 +00:00
|
|
|
assert((unsigned)VT.getSimpleVT().SimpleTy <
|
|
|
|
array_lengthof(RegisterTypeForVT));
|
|
|
|
return RegisterTypeForVT[VT.getSimpleVT().SimpleTy];
|
2008-01-25 10:20:53 +00:00
|
|
|
}
|
2008-06-06 12:08:01 +00:00
|
|
|
if (VT.isVector()) {
|
2012-12-11 11:14:33 +00:00
|
|
|
EVT VT1, RegisterVT;
|
2007-10-16 09:56:48 +00:00
|
|
|
unsigned NumIntermediates;
|
2009-08-12 00:36:31 +00:00
|
|
|
(void)getVectorTypeBreakdown(Context, VT, VT1,
|
|
|
|
NumIntermediates, RegisterVT);
|
2012-12-11 10:16:19 +00:00
|
|
|
return RegisterVT;
|
2007-10-16 09:56:48 +00:00
|
|
|
}
|
2008-06-06 12:08:01 +00:00
|
|
|
if (VT.isInteger()) {
|
2009-08-12 00:36:31 +00:00
|
|
|
return getRegisterType(Context, getTypeToTransformTo(Context, VT));
|
2008-02-11 11:09:23 +00:00
|
|
|
}
|
2012-02-05 22:14:15 +00:00
|
|
|
llvm_unreachable("Unsupported extended type!");
|
2007-06-28 23:29:44 +00:00
|
|
|
}
|
2007-10-16 09:56:48 +00:00
|
|
|
|
2007-06-21 14:42:22 +00:00
|
|
|
/// getNumRegisters - Return the number of registers that this ValueType will
|
2006-12-13 06:09:03 +00:00
|
|
|
/// eventually require. This is one for any types promoted to live in larger
|
|
|
|
/// registers, but may be more than one for types (like i64) that are split
|
2008-02-11 11:09:23 +00:00
|
|
|
/// into pieces. For types like i140, which are first promoted then expanded,
|
|
|
|
/// it is the number of registers needed to hold all the bits of the original
|
|
|
|
/// type. For an i140 on a 32 bit machine this means 5 registers.
|
2009-08-12 00:36:31 +00:00
|
|
|
unsigned getNumRegisters(LLVMContext &Context, EVT VT) const {
|
2008-06-06 12:08:01 +00:00
|
|
|
if (VT.isSimple()) {
|
2009-08-11 20:47:22 +00:00
|
|
|
assert((unsigned)VT.getSimpleVT().SimpleTy <
|
|
|
|
array_lengthof(NumRegistersForVT));
|
|
|
|
return NumRegistersForVT[VT.getSimpleVT().SimpleTy];
|
2008-01-25 10:20:53 +00:00
|
|
|
}
|
2008-06-06 12:08:01 +00:00
|
|
|
if (VT.isVector()) {
|
2012-12-11 11:14:33 +00:00
|
|
|
EVT VT1, VT2;
|
2007-10-16 09:56:48 +00:00
|
|
|
unsigned NumIntermediates;
|
2009-08-12 00:36:31 +00:00
|
|
|
return getVectorTypeBreakdown(Context, VT, VT1, NumIntermediates, VT2);
|
2007-10-16 09:56:48 +00:00
|
|
|
}
|
2008-06-06 12:08:01 +00:00
|
|
|
if (VT.isInteger()) {
|
|
|
|
unsigned BitWidth = VT.getSizeInBits();
|
2009-08-12 00:36:31 +00:00
|
|
|
unsigned RegWidth = getRegisterType(Context, VT).getSizeInBits();
|
2008-02-11 11:09:23 +00:00
|
|
|
return (BitWidth + RegWidth - 1) / RegWidth;
|
|
|
|
}
|
2012-02-05 22:14:15 +00:00
|
|
|
llvm_unreachable("Unsupported extended type!");
|
2005-01-07 07:44:22 +00:00
|
|
|
}
|
2008-01-23 23:17:41 +00:00
|
|
|
|
2008-03-05 01:30:59 +00:00
|
|
|
/// ShouldShrinkFPConstant - If true, then instruction selection should
|
|
|
|
/// seek to shrink the FP constant of the specified type to a smaller type
|
|
|
|
/// in order to save space and / or reduce runtime.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool ShouldShrinkFPConstant(EVT) const { return true; }
|
2008-03-05 01:30:59 +00:00
|
|
|
|
2006-03-01 04:52:55 +00:00
|
|
|
/// hasTargetDAGCombine - If true, the target has custom DAG combine
|
|
|
|
/// transformations that it can perform for the specified node.
|
|
|
|
bool hasTargetDAGCombine(ISD::NodeType NT) const {
|
2008-01-25 10:20:53 +00:00
|
|
|
assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray));
|
2006-03-01 04:52:55 +00:00
|
|
|
return TargetDAGCombineArray[NT >> 3] & (1 << (NT&7));
|
|
|
|
}
|
2005-01-07 07:44:22 +00:00
|
|
|
|
2005-07-19 04:52:44 +00:00
|
|
|
/// This function returns the maximum number of store operations permitted
|
|
|
|
/// to replace a call to llvm.memset. The value is set by the target at the
|
2011-01-06 06:52:41 +00:00
|
|
|
/// performance threshold for such a replacement. If OptSize is true,
|
|
|
|
/// return the limit for functions that have OptSize attribute.
|
2005-07-19 04:52:44 +00:00
|
|
|
/// @brief Get maximum # of store operations permitted for llvm.memset
|
2011-01-06 06:52:41 +00:00
|
|
|
unsigned getMaxStoresPerMemset(bool OptSize) const {
|
|
|
|
return OptSize ? maxStoresPerMemsetOptSize : maxStoresPerMemset;
|
|
|
|
}
|
2005-07-19 04:52:44 +00:00
|
|
|
|
|
|
|
/// This function returns the maximum number of store operations permitted
|
|
|
|
/// to replace a call to llvm.memcpy. The value is set by the target at the
|
2011-01-06 06:52:41 +00:00
|
|
|
/// performance threshold for such a replacement. If OptSize is true,
|
|
|
|
/// return the limit for functions that have OptSize attribute.
|
2005-07-19 04:52:44 +00:00
|
|
|
/// @brief Get maximum # of store operations permitted for llvm.memcpy
|
2011-01-06 06:52:41 +00:00
|
|
|
unsigned getMaxStoresPerMemcpy(bool OptSize) const {
|
|
|
|
return OptSize ? maxStoresPerMemcpyOptSize : maxStoresPerMemcpy;
|
|
|
|
}
|
2005-07-19 04:52:44 +00:00
|
|
|
|
|
|
|
/// This function returns the maximum number of store operations permitted
|
|
|
|
/// to replace a call to llvm.memmove. The value is set by the target at the
|
2011-01-06 06:52:41 +00:00
|
|
|
/// performance threshold for such a replacement. If OptSize is true,
|
|
|
|
/// return the limit for functions that have OptSize attribute.
|
2005-07-19 04:52:44 +00:00
|
|
|
/// @brief Get maximum # of store operations permitted for llvm.memmove
|
2011-01-06 06:52:41 +00:00
|
|
|
unsigned getMaxStoresPerMemmove(bool OptSize) const {
|
|
|
|
return OptSize ? maxStoresPerMemmoveOptSize : maxStoresPerMemmove;
|
|
|
|
}
|
2005-07-19 04:52:44 +00:00
|
|
|
|
2005-08-27 19:09:02 +00:00
|
|
|
/// This function returns true if the target allows unaligned memory accesses.
|
2012-12-10 23:21:26 +00:00
|
|
|
/// of the specified type. If true, it also returns whether the unaligned
|
|
|
|
/// memory access is "fast" in the second argument by reference. This is used,
|
|
|
|
/// for example, in situations where an array copy/move/set is converted to a
|
|
|
|
/// sequence of store operations. It's use helps to ensure that such
|
|
|
|
/// replacements don't generate code that causes an alignment error (trap) on
|
|
|
|
/// the target machine.
|
2005-08-27 19:09:02 +00:00
|
|
|
/// @brief Determine if the target supports unaligned memory accesses.
|
2012-12-10 23:21:26 +00:00
|
|
|
virtual bool allowsUnalignedMemoryAccesses(EVT, bool *Fast = 0) const {
|
2009-08-15 21:21:19 +00:00
|
|
|
return false;
|
2006-01-29 08:40:37 +00:00
|
|
|
}
|
2008-05-15 08:39:06 +00:00
|
|
|
|
2009-05-13 21:42:09 +00:00
|
|
|
/// This function returns true if the target would benefit from code placement
|
|
|
|
/// optimization.
|
|
|
|
/// @brief Determine if the target should perform code placement optimization.
|
|
|
|
bool shouldOptimizeCodePlacement() const {
|
|
|
|
return benefitFromCodePlacementOpt;
|
|
|
|
}
|
|
|
|
|
2008-05-15 08:39:06 +00:00
|
|
|
/// getOptimalMemOpType - Returns the target specific optimal type for load
|
2010-04-02 19:36:14 +00:00
|
|
|
/// and store operations as a result of memset, memcpy, and memmove
|
|
|
|
/// lowering. If DstAlign is zero that means it's safe to destination
|
|
|
|
/// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
|
|
|
|
/// means there isn't a need to check it against alignment requirement,
|
2012-12-12 02:34:41 +00:00
|
|
|
/// probably because the source does not need to be loaded. If 'IsMemset' is
|
|
|
|
/// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
|
|
|
|
/// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
|
|
|
|
/// source is constant so it does not need to be loaded.
|
2010-04-16 20:11:05 +00:00
|
|
|
/// It returns EVT::Other if the type should be determined using generic
|
|
|
|
/// target-independent logic.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual EVT getOptimalMemOpType(uint64_t /*Size*/,
|
|
|
|
unsigned /*DstAlign*/, unsigned /*SrcAlign*/,
|
2012-12-12 02:34:41 +00:00
|
|
|
bool /*IsMemset*/,
|
|
|
|
bool /*ZeroMemset*/,
|
2011-10-20 08:57:54 +00:00
|
|
|
bool /*MemcpyStrSrc*/,
|
|
|
|
MachineFunction &/*MF*/) const {
|
2010-03-07 07:45:08 +00:00
|
|
|
return MVT::Other;
|
2008-05-15 08:39:06 +00:00
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2012-12-12 01:32:07 +00:00
|
|
|
/// isSafeMemOpType - Returns true if it's safe to use load / store of the
|
2012-12-12 00:42:09 +00:00
|
|
|
/// specified type to expand memcpy / memset inline. This is mostly true
|
2012-12-12 01:32:07 +00:00
|
|
|
/// for all types except for some special cases. For example, on X86
|
2012-12-12 00:42:09 +00:00
|
|
|
/// targets without SSE2 f64 load / store are done with fldl / fstpl which
|
2012-12-12 01:32:07 +00:00
|
|
|
/// also does type conversion. Note the specified type doesn't have to be
|
|
|
|
/// legal as the hook is used before type legalization.
|
|
|
|
virtual bool isSafeMemOpType(MVT VT) const {
|
|
|
|
return true;
|
2012-12-12 00:42:09 +00:00
|
|
|
}
|
|
|
|
|
2006-12-10 23:12:42 +00:00
|
|
|
/// usesUnderscoreSetJmp - Determine if we should use _setjmp or setjmp
|
2005-09-27 22:13:36 +00:00
|
|
|
/// to implement llvm.setjmp.
|
2006-12-10 23:12:42 +00:00
|
|
|
bool usesUnderscoreSetJmp() const {
|
|
|
|
return UseUnderscoreSetJmp;
|
2005-09-27 22:13:36 +00:00
|
|
|
}
|
2006-12-10 23:12:42 +00:00
|
|
|
|
|
|
|
/// usesUnderscoreLongJmp - Determine if we should use _longjmp or longjmp
|
|
|
|
/// to implement llvm.longjmp.
|
|
|
|
bool usesUnderscoreLongJmp() const {
|
|
|
|
return UseUnderscoreLongJmp;
|
|
|
|
}
|
|
|
|
|
2012-07-02 22:39:56 +00:00
|
|
|
/// supportJumpTables - return whether the target can generate code for
|
|
|
|
/// jump tables.
|
|
|
|
bool supportJumpTables() const {
|
|
|
|
return SupportJumpTables;
|
|
|
|
}
|
|
|
|
|
2012-09-25 20:35:36 +00:00
|
|
|
/// getMinimumJumpTableEntries - return integer threshold on number of
|
|
|
|
/// blocks to use jump tables rather than if sequence.
|
|
|
|
int getMinimumJumpTableEntries() const {
|
|
|
|
return MinimumJumpTableEntries;
|
|
|
|
}
|
|
|
|
|
2006-01-13 17:47:52 +00:00
|
|
|
/// getStackPointerRegisterToSaveRestore - If a physical register, this
|
|
|
|
/// specifies the register that llvm.savestack/llvm.restorestack should save
|
|
|
|
/// and restore.
|
|
|
|
unsigned getStackPointerRegisterToSaveRestore() const {
|
|
|
|
return StackPointerRegisterToSaveRestore;
|
|
|
|
}
|
2005-07-19 04:52:44 +00:00
|
|
|
|
2012-02-14 04:45:49 +00:00
|
|
|
/// getExceptionPointerRegister - If a physical register, this returns
|
2007-02-22 15:37:19 +00:00
|
|
|
/// the register that receives the exception address on entry to a landing
|
|
|
|
/// pad.
|
2012-02-14 04:45:49 +00:00
|
|
|
unsigned getExceptionPointerRegister() const {
|
2007-02-22 15:37:19 +00:00
|
|
|
return ExceptionPointerRegister;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// getExceptionSelectorRegister - If a physical register, this returns
|
|
|
|
/// the register that receives the exception typeid on entry to a landing
|
|
|
|
/// pad.
|
|
|
|
unsigned getExceptionSelectorRegister() const {
|
|
|
|
return ExceptionSelectorRegister;
|
|
|
|
}
|
|
|
|
|
2006-09-04 06:21:35 +00:00
|
|
|
/// getJumpBufSize - returns the target's jmp_buf size in bytes (if never
|
|
|
|
/// set, the default is 200)
|
|
|
|
unsigned getJumpBufSize() const {
|
|
|
|
return JumpBufSize;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// getJumpBufAlignment - returns the target's jmp_buf alignment in bytes
|
|
|
|
/// (if never set, the default is 0)
|
|
|
|
unsigned getJumpBufAlignment() const {
|
|
|
|
return JumpBufAlignment;
|
|
|
|
}
|
2006-11-03 03:04:06 +00:00
|
|
|
|
2010-07-11 04:01:49 +00:00
|
|
|
/// getMinStackArgumentAlignment - return the minimum stack alignment of an
|
|
|
|
/// argument.
|
|
|
|
unsigned getMinStackArgumentAlignment() const {
|
|
|
|
return MinStackArgumentAlignment;
|
|
|
|
}
|
|
|
|
|
2011-05-06 20:34:06 +00:00
|
|
|
/// getMinFunctionAlignment - return the minimum function alignment.
|
|
|
|
///
|
|
|
|
unsigned getMinFunctionAlignment() const {
|
|
|
|
return MinFunctionAlignment;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// getPrefFunctionAlignment - return the preferred function alignment.
|
|
|
|
///
|
|
|
|
unsigned getPrefFunctionAlignment() const {
|
|
|
|
return PrefFunctionAlignment;
|
|
|
|
}
|
|
|
|
|
2008-02-28 00:43:03 +00:00
|
|
|
/// getPrefLoopAlignment - return the preferred loop alignment.
|
|
|
|
///
|
|
|
|
unsigned getPrefLoopAlignment() const {
|
|
|
|
return PrefLoopAlignment;
|
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2010-06-23 16:07:42 +00:00
|
|
|
/// getShouldFoldAtomicFences - return whether the combiner should fold
|
|
|
|
/// fence MEMBARRIER instructions into the atomic intrinsic instructions.
|
|
|
|
///
|
|
|
|
bool getShouldFoldAtomicFences() const {
|
|
|
|
return ShouldFoldAtomicFences;
|
|
|
|
}
|
|
|
|
|
2011-08-03 21:06:02 +00:00
|
|
|
/// getInsertFencesFor - return whether the DAG builder should automatically
|
|
|
|
/// insert fences and reduce ordering for atomics.
|
|
|
|
///
|
|
|
|
bool getInsertFencesForAtomic() const {
|
|
|
|
return InsertFencesForAtomic;
|
|
|
|
}
|
|
|
|
|
2006-11-03 07:21:16 +00:00
|
|
|
/// getPreIndexedAddressParts - returns true by value, base pointer and
|
2006-11-03 03:04:06 +00:00
|
|
|
/// offset pointer and addressing mode by reference if the node's address
|
|
|
|
/// can be legally represented as pre-indexed load / store address.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool getPreIndexedAddressParts(SDNode * /*N*/, SDValue &/*Base*/,
|
|
|
|
SDValue &/*Offset*/,
|
|
|
|
ISD::MemIndexedMode &/*AM*/,
|
|
|
|
SelectionDAG &/*DAG*/) const {
|
2006-11-03 03:04:06 +00:00
|
|
|
return false;
|
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-11-07 09:04:16 +00:00
|
|
|
/// getPostIndexedAddressParts - returns true by value, base pointer and
|
|
|
|
/// offset pointer and addressing mode by reference if this node can be
|
|
|
|
/// combined with a load / store to form a post-indexed load / store.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool getPostIndexedAddressParts(SDNode * /*N*/, SDNode * /*Op*/,
|
|
|
|
SDValue &/*Base*/, SDValue &/*Offset*/,
|
|
|
|
ISD::MemIndexedMode &/*AM*/,
|
|
|
|
SelectionDAG &/*DAG*/) const {
|
2006-11-07 09:04:16 +00:00
|
|
|
return false;
|
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2010-01-25 23:26:13 +00:00
|
|
|
/// getJumpTableEncoding - Return the entry encoding for a jump table in the
|
|
|
|
/// current function. The returned value is a member of the
|
|
|
|
/// MachineJumpTableInfo::JTEntryKind enum.
|
|
|
|
virtual unsigned getJumpTableEncoding() const;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2010-01-26 04:05:28 +00:00
|
|
|
virtual const MCExpr *
|
2011-10-20 08:57:54 +00:00
|
|
|
LowerCustomJumpTableEntry(const MachineJumpTableInfo * /*MJTI*/,
|
|
|
|
const MachineBasicBlock * /*MBB*/, unsigned /*uid*/,
|
|
|
|
MCContext &/*Ctx*/) const {
|
2012-02-05 22:14:15 +00:00
|
|
|
llvm_unreachable("Need to implement this hook if target has custom JTIs");
|
2010-01-26 04:05:28 +00:00
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
Much improved pic jumptable codegen:
Then:
call "L1$pb"
"L1$pb":
popl %eax
...
LBB1_1: # entry
imull $4, %ecx, %ecx
leal LJTI1_0-"L1$pb"(%eax), %edx
addl LJTI1_0-"L1$pb"(%ecx,%eax), %edx
jmpl *%edx
.align 2
.set L1_0_set_3,LBB1_3-LJTI1_0
.set L1_0_set_2,LBB1_2-LJTI1_0
.set L1_0_set_5,LBB1_5-LJTI1_0
.set L1_0_set_4,LBB1_4-LJTI1_0
LJTI1_0:
.long L1_0_set_3
.long L1_0_set_2
Now:
call "L1$pb"
"L1$pb":
popl %eax
...
LBB1_1: # entry
addl LJTI1_0-"L1$pb"(%eax,%ecx,4), %eax
jmpl *%eax
.align 2
.set L1_0_set_3,LBB1_3-"L1$pb"
.set L1_0_set_2,LBB1_2-"L1$pb"
.set L1_0_set_5,LBB1_5-"L1$pb"
.set L1_0_set_4,LBB1_4-"L1$pb"
LJTI1_0:
.long L1_0_set_3
.long L1_0_set_2
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43924 91177308-0d34-0410-b5e6-96231b3b80d8
2007-11-09 01:32:10 +00:00
|
|
|
/// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
|
|
|
|
/// jumptable.
|
2008-07-27 21:46:04 +00:00
|
|
|
virtual SDValue getPICJumpTableRelocBase(SDValue Table,
|
2010-01-26 04:05:28 +00:00
|
|
|
SelectionDAG &DAG) const;
|
2008-02-28 00:43:03 +00:00
|
|
|
|
2010-01-26 05:30:30 +00:00
|
|
|
/// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
|
|
|
|
/// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
|
|
|
|
/// MCExpr.
|
|
|
|
virtual const MCExpr *
|
2010-01-26 06:28:43 +00:00
|
|
|
getPICJumpTableRelocBaseExpr(const MachineFunction *MF,
|
2010-01-26 05:30:30 +00:00
|
|
|
unsigned JTI, MCContext &Ctx) const;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
Teach DAGCombine to fold constant offsets into GlobalAddress nodes,
and add a TargetLowering hook for it to use to determine when this
is legal (i.e. not in PIC mode, etc.)
This allows instruction selection to emit folded constant offsets
in more cases, such as the included testcase, eliminating the need
for explicit arithmetic instructions.
This eliminates the need for the C++ code in X86ISelDAGToDAG.cpp
that attempted to achieve the same effect, but wasn't as effective.
Also, fix handling of offsets in GlobalAddressSDNodes in several
places, including changing GlobalAddressSDNode's offset from
int to int64_t.
The Mips, Alpha, Sparc, and CellSPU targets appear to be
unaware of GlobalAddress offsets currently, so set the hook to
false on those targets.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@57748 91177308-0d34-0410-b5e6-96231b3b80d8
2008-10-18 02:06:02 +00:00
|
|
|
/// isOffsetFoldingLegal - Return true if folding a constant offset
|
|
|
|
/// with the given GlobalAddress is legal. It is frequently not legal in
|
|
|
|
/// PIC relocation models.
|
|
|
|
virtual bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const;
|
|
|
|
|
2010-07-06 05:18:56 +00:00
|
|
|
/// getStackCookieLocation - Return true if the target stores stack
|
|
|
|
/// protector cookies at a fixed offset in some non-standard address
|
|
|
|
/// space, and populates the address space and offset as
|
|
|
|
/// appropriate.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool getStackCookieLocation(unsigned &/*AddressSpace*/,
|
|
|
|
unsigned &/*Offset*/) const {
|
2010-07-06 05:18:56 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2010-07-24 21:52:08 +00:00
|
|
|
/// getMaximalGlobalOffset - Returns the maximal possible offset which can be
|
|
|
|
/// used for loads / stores from the global.
|
|
|
|
virtual unsigned getMaximalGlobalOffset() const {
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2006-01-30 04:08:18 +00:00
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// TargetLowering Optimization Methods
|
|
|
|
//
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-02-16 21:11:51 +00:00
|
|
|
/// TargetLoweringOpt - A convenience struct that encapsulates a DAG, and two
|
2008-07-27 21:46:04 +00:00
|
|
|
/// SDValues for returning information from TargetLowering to its clients
|
2010-06-23 15:49:18 +00:00
|
|
|
/// that want to combine
|
2006-02-16 21:11:51 +00:00
|
|
|
struct TargetLoweringOpt {
|
|
|
|
SelectionDAG &DAG;
|
2010-04-17 06:13:15 +00:00
|
|
|
bool LegalTys;
|
|
|
|
bool LegalOps;
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Old;
|
|
|
|
SDValue New;
|
2006-02-16 21:11:51 +00:00
|
|
|
|
2010-04-17 06:13:15 +00:00
|
|
|
explicit TargetLoweringOpt(SelectionDAG &InDAG,
|
2010-06-24 14:30:44 +00:00
|
|
|
bool LT, bool LO) :
|
|
|
|
DAG(InDAG), LegalTys(LT), LegalOps(LO) {}
|
2010-04-17 06:13:15 +00:00
|
|
|
|
|
|
|
bool LegalTypes() const { return LegalTys; }
|
|
|
|
bool LegalOperations() const { return LegalOps; }
|
2010-06-23 15:49:18 +00:00
|
|
|
|
|
|
|
bool CombineTo(SDValue O, SDValue N) {
|
|
|
|
Old = O;
|
|
|
|
New = N;
|
2006-02-16 21:11:51 +00:00
|
|
|
return true;
|
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
|
|
|
/// ShrinkDemandedConstant - Check to see if the specified operand of the
|
2007-10-15 05:30:27 +00:00
|
|
|
/// specified instruction is a constant integer. If so, check to see if
|
|
|
|
/// there are any bits set in the constant that are not demanded. If so,
|
|
|
|
/// shrink the constant and return true.
|
2008-07-27 21:46:04 +00:00
|
|
|
bool ShrinkDemandedConstant(SDValue Op, const APInt &Demanded);
|
Implement support for using modeling implicit-zero-extension on x86-64
with SUBREG_TO_REG, teach SimpleRegisterCoalescing to coalesce
SUBREG_TO_REG instructions (which are similar to INSERT_SUBREG
instructions), and teach the DAGCombiner to take advantage of this on
targets which support it. This eliminates many redundant
zero-extension operations on x86-64.
This adds a new TargetLowering hook, isZExtFree. It's similar to
isTruncateFree, except it only applies to actual definitions, and not
no-op truncates which may not zero the high bits.
Also, this adds a new optimization to SimplifyDemandedBits: transform
operations like x+y into (zext (add (trunc x), (trunc y))) on targets
where all the casts are no-ops. In contexts where the high part of the
add is explicitly masked off, this allows the mask operation to be
eliminated. Fix the DAGCombiner to avoid undoing these transformations
to eliminate casts on targets where the casts are no-ops.
Also, this adds a new two-address lowering heuristic. Since
two-address lowering runs before coalescing, it helps to be able to
look through copies when deciding whether commuting and/or
three-address conversion are profitable.
Also, fix a bug in LiveInterval::MergeInClobberRanges. It didn't handle
the case that a clobber range extended both before and beyond an
existing live range. In that case, multiple live ranges need to be
added. This was exposed by the new subreg coalescing code.
Remove 2008-05-06-SpillerBug.ll. It was bugpoint-reduced, and the
spiller behavior it was looking for no longer occurrs with the new
instruction selection.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@68576 91177308-0d34-0410-b5e6-96231b3b80d8
2009-04-08 00:15:30 +00:00
|
|
|
|
|
|
|
/// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the
|
|
|
|
/// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening
|
|
|
|
/// cast, but it could be generalized for targets with other types of
|
|
|
|
/// implicit widening casts.
|
|
|
|
bool ShrinkDemandedOp(SDValue Op, unsigned BitWidth, const APInt &Demanded,
|
|
|
|
DebugLoc dl);
|
2006-02-16 21:11:51 +00:00
|
|
|
};
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-02-16 21:11:51 +00:00
|
|
|
/// SimplifyDemandedBits - Look at Op. At this point, we know that only the
|
|
|
|
/// DemandedMask bits of the result of Op are ever used downstream. If we can
|
|
|
|
/// use this information to simplify Op, create a new simplified DAG node and
|
2010-06-23 15:49:18 +00:00
|
|
|
/// return true, returning the original and new nodes in Old and New.
|
|
|
|
/// Otherwise, analyze the expression and return a mask of KnownOne and
|
|
|
|
/// KnownZero bits for the expression (used to simplify the caller).
|
|
|
|
/// The KnownZero/One bits may only be accurate for those bits in the
|
2006-02-16 21:11:51 +00:00
|
|
|
/// DemandedMask.
|
2010-06-23 15:49:18 +00:00
|
|
|
bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedMask,
|
2008-02-27 00:25:32 +00:00
|
|
|
APInt &KnownZero, APInt &KnownOne,
|
2006-02-16 21:11:51 +00:00
|
|
|
TargetLoweringOpt &TLO, unsigned Depth = 0) const;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-03-01 04:52:55 +00:00
|
|
|
/// computeMaskedBitsForTargetNode - Determine which of the bits specified in
|
2010-06-23 15:49:18 +00:00
|
|
|
/// Mask are known to be either zero or one and return them in the
|
2006-03-01 04:52:55 +00:00
|
|
|
/// KnownZero/KnownOne bitsets.
|
2008-07-27 21:46:04 +00:00
|
|
|
virtual void computeMaskedBitsForTargetNode(const SDValue Op,
|
2010-06-23 15:49:18 +00:00
|
|
|
APInt &KnownZero,
|
2008-02-13 00:35:47 +00:00
|
|
|
APInt &KnownOne,
|
2007-06-22 14:59:07 +00:00
|
|
|
const SelectionDAG &DAG,
|
2006-03-01 04:52:55 +00:00
|
|
|
unsigned Depth = 0) const;
|
2006-03-13 23:18:16 +00:00
|
|
|
|
2006-05-06 09:26:22 +00:00
|
|
|
/// ComputeNumSignBitsForTargetNode - This method can be implemented by
|
|
|
|
/// targets that want to expose additional information about sign bits to the
|
|
|
|
/// DAG Combiner.
|
2008-07-27 21:46:04 +00:00
|
|
|
virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
|
2006-05-06 09:26:22 +00:00
|
|
|
unsigned Depth = 0) const;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-03-01 04:52:55 +00:00
|
|
|
struct DAGCombinerInfo {
|
|
|
|
void *DC; // The DAG Combiner object.
|
|
|
|
bool BeforeLegalize;
|
2009-07-24 18:22:59 +00:00
|
|
|
bool BeforeLegalizeOps;
|
2007-02-08 22:13:59 +00:00
|
|
|
bool CalledByLegalizer;
|
2006-03-01 04:52:55 +00:00
|
|
|
public:
|
|
|
|
SelectionDAG &DAG;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2009-07-24 18:22:59 +00:00
|
|
|
DAGCombinerInfo(SelectionDAG &dag, bool bl, bool blo, bool cl, void *dc)
|
|
|
|
: DC(dc), BeforeLegalize(bl), BeforeLegalizeOps(blo),
|
|
|
|
CalledByLegalizer(cl), DAG(dag) {}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-03-01 04:52:55 +00:00
|
|
|
bool isBeforeLegalize() const { return BeforeLegalize; }
|
2009-07-24 18:22:59 +00:00
|
|
|
bool isBeforeLegalizeOps() const { return BeforeLegalizeOps; }
|
2007-02-08 22:13:59 +00:00
|
|
|
bool isCalledByLegalizer() const { return CalledByLegalizer; }
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-03-01 04:52:55 +00:00
|
|
|
void AddToWorklist(SDNode *N);
|
2011-04-02 02:40:26 +00:00
|
|
|
void RemoveFromWorklist(SDNode *N);
|
Optimize some 64-bit multiplication by constants into two lea's or one lea + shl since imulq is slow (latency 5). e.g.
x * 40
=>
shlq $3, %rdi
leaq (%rdi,%rdi,4), %rax
This has the added benefit of allowing more multiply to be folded into addressing mode. e.g.
a * 24 + b
=>
leaq (%rdi,%rdi,2), %rax
leaq (%rsi,%rax,8), %rax
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@67917 91177308-0d34-0410-b5e6-96231b3b80d8
2009-03-28 05:57:29 +00:00
|
|
|
SDValue CombineTo(SDNode *N, const std::vector<SDValue> &To,
|
|
|
|
bool AddTo = true);
|
|
|
|
SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true);
|
|
|
|
SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo = true);
|
2009-01-29 01:59:02 +00:00
|
|
|
|
|
|
|
void CommitTargetLoweringOpt(const TargetLoweringOpt &TLO);
|
2006-03-01 04:52:55 +00:00
|
|
|
};
|
|
|
|
|
2010-06-23 15:49:18 +00:00
|
|
|
/// SimplifySetCC - Try to simplify a setcc built with the specified operands
|
2008-07-27 21:46:04 +00:00
|
|
|
/// and cc. If it is unable to simplify it, return a null SDValue.
|
2009-08-10 22:56:29 +00:00
|
|
|
SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
|
2007-02-08 22:13:59 +00:00
|
|
|
ISD::CondCode Cond, bool foldBooleans,
|
2009-02-03 00:47:48 +00:00
|
|
|
DAGCombinerInfo &DCI, DebugLoc dl) const;
|
2007-02-08 22:13:59 +00:00
|
|
|
|
2008-05-12 20:08:05 +00:00
|
|
|
/// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
|
|
|
|
/// node is a GlobalAddress + offset.
|
|
|
|
virtual bool
|
2010-04-15 01:51:59 +00:00
|
|
|
isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const;
|
2008-05-12 20:08:05 +00:00
|
|
|
|
2006-03-01 04:52:55 +00:00
|
|
|
/// PerformDAGCombine - This method will be invoked for all target nodes and
|
|
|
|
/// for any target-independent nodes that the target has registered with
|
|
|
|
/// invoke it for.
|
|
|
|
///
|
|
|
|
/// The semantics are as follows:
|
|
|
|
/// Return Value:
|
2008-07-27 21:46:04 +00:00
|
|
|
/// SDValue.Val == 0 - No change was made
|
|
|
|
/// SDValue.Val == N - N was replaced, is dead, and is already handled.
|
2008-11-24 14:53:14 +00:00
|
|
|
/// otherwise - N should be replaced by the returned Operand.
|
2006-03-01 04:52:55 +00:00
|
|
|
///
|
|
|
|
/// In addition, methods provided by DAGCombinerInfo may be used to perform
|
|
|
|
/// more complex transformations.
|
|
|
|
///
|
2008-07-27 21:46:04 +00:00
|
|
|
virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
|
2010-04-16 06:14:10 +00:00
|
|
|
|
2010-04-17 06:13:15 +00:00
|
|
|
/// isTypeDesirableForOp - Return true if the target has native support for
|
|
|
|
/// the specified value type and it is 'desirable' to use the type for the
|
|
|
|
/// given node type. e.g. On x86 i16 is legal, but undesirable since i16
|
|
|
|
/// instruction encodings are longer and some i16 instructions are slow.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool isTypeDesirableForOp(unsigned /*Opc*/, EVT VT) const {
|
2010-04-17 06:13:15 +00:00
|
|
|
// By default, assume all legal types are desirable.
|
|
|
|
return isTypeLegal(VT);
|
|
|
|
}
|
|
|
|
|
2011-02-02 01:06:55 +00:00
|
|
|
/// isDesirableToPromoteOp - Return true if it is profitable for dag combiner
|
|
|
|
/// to transform a floating point op of specified opcode to a equivalent op of
|
|
|
|
/// an integer type. e.g. f32 load -> i32 load can be profitable on ARM.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool isDesirableToTransformToIntegerOp(unsigned /*Opc*/,
|
|
|
|
EVT /*VT*/) const {
|
2011-02-02 01:06:55 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2010-04-17 06:13:15 +00:00
|
|
|
/// IsDesirableToPromoteOp - This method query the target whether it is
|
2010-04-16 06:14:10 +00:00
|
|
|
/// beneficial for dag combiner to promote the specified node. If true, it
|
|
|
|
/// should return the desired promotion type by reference.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool IsDesirableToPromoteOp(SDValue /*Op*/, EVT &/*PVT*/) const {
|
2010-04-16 06:14:10 +00:00
|
|
|
return false;
|
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2005-01-07 07:44:22 +00:00
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// TargetLowering Configuration Methods - These methods should be invoked by
|
|
|
|
// the derived class constructor to configure this object for the target.
|
|
|
|
//
|
|
|
|
|
|
|
|
protected:
|
2008-11-23 15:47:28 +00:00
|
|
|
/// setBooleanContents - Specify how the target extends the result of a
|
|
|
|
/// boolean value from i1 to a wider type. See getBooleanContents.
|
|
|
|
void setBooleanContents(BooleanContent Ty) { BooleanContents = Ty; }
|
2011-09-06 19:07:46 +00:00
|
|
|
/// setBooleanVectorContents - Specify how the target extends the result
|
|
|
|
/// of a vector boolean value from a vector of i1 to a wider type. See
|
|
|
|
/// getBooleanContents.
|
|
|
|
void setBooleanVectorContents(BooleanContent Ty) {
|
|
|
|
BooleanVectorContents = Ty;
|
|
|
|
}
|
2005-04-07 19:41:18 +00:00
|
|
|
|
2006-01-25 09:09:02 +00:00
|
|
|
/// setSchedulingPreference - Specify the target scheduling preference.
|
2010-05-19 20:19:50 +00:00
|
|
|
void setSchedulingPreference(Sched::Preference Pref) {
|
2006-01-25 09:09:02 +00:00
|
|
|
SchedPreferenceInfo = Pref;
|
|
|
|
}
|
|
|
|
|
2006-12-10 23:12:42 +00:00
|
|
|
/// setUseUnderscoreSetJmp - Indicate whether this target prefers to
|
|
|
|
/// use _setjmp to implement llvm.setjmp or the non _ version.
|
|
|
|
/// Defaults to false.
|
|
|
|
void setUseUnderscoreSetJmp(bool Val) {
|
|
|
|
UseUnderscoreSetJmp = Val;
|
2005-09-27 22:13:36 +00:00
|
|
|
}
|
2006-12-10 23:12:42 +00:00
|
|
|
|
|
|
|
/// setUseUnderscoreLongJmp - Indicate whether this target prefers to
|
|
|
|
/// use _longjmp to implement llvm.longjmp or the non _ version.
|
|
|
|
/// Defaults to false.
|
|
|
|
void setUseUnderscoreLongJmp(bool Val) {
|
|
|
|
UseUnderscoreLongJmp = Val;
|
|
|
|
}
|
|
|
|
|
2012-07-02 22:39:56 +00:00
|
|
|
/// setSupportJumpTables - Indicate whether the target can generate code for
|
|
|
|
/// jump tables.
|
|
|
|
void setSupportJumpTables(bool Val) {
|
|
|
|
SupportJumpTables = Val;
|
|
|
|
}
|
|
|
|
|
2012-09-25 20:35:36 +00:00
|
|
|
/// setMinimumJumpTableEntries - Indicate the number of blocks to generate
|
|
|
|
/// jump tables rather than if sequence.
|
|
|
|
void setMinimumJumpTableEntries(int Val) {
|
|
|
|
MinimumJumpTableEntries = Val;
|
|
|
|
}
|
|
|
|
|
2006-01-13 17:47:52 +00:00
|
|
|
/// setStackPointerRegisterToSaveRestore - If set to a physical register, this
|
|
|
|
/// specifies the register that llvm.savestack/llvm.restorestack should save
|
|
|
|
/// and restore.
|
|
|
|
void setStackPointerRegisterToSaveRestore(unsigned R) {
|
|
|
|
StackPointerRegisterToSaveRestore = R;
|
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2007-02-22 15:37:19 +00:00
|
|
|
/// setExceptionPointerRegister - If set to a physical register, this sets
|
|
|
|
/// the register that receives the exception address on entry to a landing
|
|
|
|
/// pad.
|
|
|
|
void setExceptionPointerRegister(unsigned R) {
|
|
|
|
ExceptionPointerRegister = R;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// setExceptionSelectorRegister - If set to a physical register, this sets
|
|
|
|
/// the register that receives the exception typeid on entry to a landing
|
|
|
|
/// pad.
|
|
|
|
void setExceptionSelectorRegister(unsigned R) {
|
|
|
|
ExceptionSelectorRegister = R;
|
|
|
|
}
|
|
|
|
|
2007-01-05 23:31:08 +00:00
|
|
|
/// SelectIsExpensive - Tells the code generator not to expand operations
|
|
|
|
/// into sequences that use the select operations if possible.
|
2011-02-25 21:41:48 +00:00
|
|
|
void setSelectIsExpensive(bool isExpensive = true) {
|
|
|
|
SelectIsExpensive = isExpensive;
|
2010-11-30 18:12:52 +00:00
|
|
|
}
|
|
|
|
|
2011-02-25 21:41:48 +00:00
|
|
|
/// JumpIsExpensive - Tells the code generator not to expand sequence of
|
2011-04-15 05:18:47 +00:00
|
|
|
/// operations into a separate sequences that increases the amount of
|
2010-11-30 18:12:52 +00:00
|
|
|
/// flow control.
|
|
|
|
void setJumpIsExpensive(bool isExpensive = true) {
|
|
|
|
JumpIsExpensive = isExpensive;
|
|
|
|
}
|
2005-05-12 02:02:21 +00:00
|
|
|
|
2005-10-21 00:02:42 +00:00
|
|
|
/// setIntDivIsCheap - Tells the code generator that integer divide is
|
2005-10-20 02:14:14 +00:00
|
|
|
/// expensive, and if possible, should be replaced by an alternate sequence
|
|
|
|
/// of instructions not containing an integer divide.
|
2005-10-21 00:02:42 +00:00
|
|
|
void setIntDivIsCheap(bool isCheap = true) { IntDivIsCheap = isCheap; }
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2012-10-04 21:33:40 +00:00
|
|
|
/// addBypassSlowDiv - Tells the code generator which bitwidths to bypass.
|
|
|
|
void addBypassSlowDiv(unsigned int SlowBitWidth, unsigned int FastBitWidth) {
|
|
|
|
BypassSlowDivWidths[SlowBitWidth] = FastBitWidth;
|
2012-09-04 18:22:17 +00:00
|
|
|
}
|
|
|
|
|
2005-10-21 00:02:42 +00:00
|
|
|
/// setPow2DivIsCheap - Tells the code generator that it shouldn't generate
|
|
|
|
/// srl/add/sra for a signed divide by power of two, and let the target handle
|
|
|
|
/// it.
|
|
|
|
void setPow2DivIsCheap(bool isCheap = true) { Pow2DivIsCheap = isCheap; }
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2005-01-07 07:44:22 +00:00
|
|
|
/// addRegisterClass - Add the specified register class as an available
|
|
|
|
/// regclass for the specified value type. This indicates the selector can
|
|
|
|
/// handle values of that class natively.
|
2012-12-11 11:14:33 +00:00
|
|
|
void addRegisterClass(EVT VT, const TargetRegisterClass *RC) {
|
|
|
|
assert((unsigned)VT.getSimpleVT().SimpleTy < array_lengthof(RegClassForVT));
|
2005-01-07 07:44:22 +00:00
|
|
|
AvailableRegClasses.push_back(std::make_pair(VT, RC));
|
2012-12-11 11:14:33 +00:00
|
|
|
RegClassForVT[VT.getSimpleVT().SimpleTy] = RC;
|
2005-01-07 07:44:22 +00:00
|
|
|
}
|
|
|
|
|
2010-07-19 22:15:08 +00:00
|
|
|
/// findRepresentativeClass - Return the largest legal super-reg register class
|
2010-07-21 06:09:07 +00:00
|
|
|
/// of the register class for the specified type and its associated "cost".
|
|
|
|
virtual std::pair<const TargetRegisterClass*, uint8_t>
|
2012-12-11 11:14:33 +00:00
|
|
|
findRepresentativeClass(EVT VT) const;
|
2010-07-19 22:15:08 +00:00
|
|
|
|
2005-01-07 07:44:22 +00:00
|
|
|
/// computeRegisterProperties - Once all of the register classes are added,
|
|
|
|
/// this allows us to compute derived properties we expose.
|
|
|
|
void computeRegisterProperties();
|
2005-04-21 20:59:05 +00:00
|
|
|
|
2005-01-16 07:27:49 +00:00
|
|
|
/// setOperationAction - Indicate that the specified operation does not work
|
|
|
|
/// with the specified type and indicate what to do about it.
|
2009-08-11 20:47:22 +00:00
|
|
|
void setOperationAction(unsigned Op, MVT VT,
|
2005-01-16 07:27:49 +00:00
|
|
|
LegalizeAction Action) {
|
2010-05-14 07:37:40 +00:00
|
|
|
assert(Op < array_lengthof(OpActions[0]) && "Table isn't big enough!");
|
|
|
|
OpActions[(unsigned)VT.SimpleTy][Op] = (uint8_t)Action;
|
2005-01-07 07:44:22 +00:00
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2008-10-14 21:26:46 +00:00
|
|
|
/// setLoadExtAction - Indicate that the specified load with extension does
|
2010-03-25 00:03:04 +00:00
|
|
|
/// not work with the specified type and indicate what to do about it.
|
2009-08-11 20:47:22 +00:00
|
|
|
void setLoadExtAction(unsigned ExtType, MVT VT,
|
2010-05-13 23:01:26 +00:00
|
|
|
LegalizeAction Action) {
|
2010-11-03 12:17:33 +00:00
|
|
|
assert(ExtType < ISD::LAST_LOADEXT_TYPE && VT < MVT::LAST_VALUETYPE &&
|
2006-10-04 00:56:09 +00:00
|
|
|
"Table isn't big enough!");
|
2010-05-13 23:01:26 +00:00
|
|
|
LoadExtActions[VT.SimpleTy][ExtType] = (uint8_t)Action;
|
2006-10-04 00:56:09 +00:00
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2008-01-17 19:59:44 +00:00
|
|
|
/// setTruncStoreAction - Indicate that the specified truncating store does
|
2010-03-25 00:03:04 +00:00
|
|
|
/// not work with the specified type and indicate what to do about it.
|
2009-08-11 20:47:22 +00:00
|
|
|
void setTruncStoreAction(MVT ValVT, MVT MemVT,
|
2008-01-17 19:59:44 +00:00
|
|
|
LegalizeAction Action) {
|
2010-11-03 12:17:33 +00:00
|
|
|
assert(ValVT < MVT::LAST_VALUETYPE && MemVT < MVT::LAST_VALUETYPE &&
|
2008-06-06 12:08:01 +00:00
|
|
|
"Table isn't big enough!");
|
2010-05-13 23:01:26 +00:00
|
|
|
TruncStoreActions[ValVT.SimpleTy][MemVT.SimpleTy] = (uint8_t)Action;
|
2006-10-13 21:14:26 +00:00
|
|
|
}
|
|
|
|
|
2006-11-09 18:56:43 +00:00
|
|
|
/// setIndexedLoadAction - Indicate that the specified indexed load does or
|
2010-03-25 00:03:04 +00:00
|
|
|
/// does not work with the specified type and indicate what to do abort
|
2006-11-09 18:56:43 +00:00
|
|
|
/// it. NOTE: All indexed mode loads are initialized to Expand in
|
|
|
|
/// TargetLowering.cpp
|
2009-08-11 20:47:22 +00:00
|
|
|
void setIndexedLoadAction(unsigned IdxMode, MVT VT,
|
2006-11-09 18:56:43 +00:00
|
|
|
LegalizeAction Action) {
|
2010-11-03 12:17:33 +00:00
|
|
|
assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE &&
|
|
|
|
(unsigned)Action < 0xf && "Table isn't big enough!");
|
2010-05-13 23:01:26 +00:00
|
|
|
// Load action are kept in the upper half.
|
|
|
|
IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0xf0;
|
|
|
|
IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action) <<4;
|
2006-11-09 18:56:43 +00:00
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-11-09 18:56:43 +00:00
|
|
|
/// setIndexedStoreAction - Indicate that the specified indexed store does or
|
2010-03-25 00:03:04 +00:00
|
|
|
/// does not work with the specified type and indicate what to do about
|
2006-11-09 18:56:43 +00:00
|
|
|
/// it. NOTE: All indexed mode stores are initialized to Expand in
|
|
|
|
/// TargetLowering.cpp
|
2009-08-11 20:47:22 +00:00
|
|
|
void setIndexedStoreAction(unsigned IdxMode, MVT VT,
|
2006-11-09 18:56:43 +00:00
|
|
|
LegalizeAction Action) {
|
2010-11-03 12:17:33 +00:00
|
|
|
assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE &&
|
|
|
|
(unsigned)Action < 0xf && "Table isn't big enough!");
|
2010-05-13 23:01:26 +00:00
|
|
|
// Store action are kept in the lower half.
|
|
|
|
IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0x0f;
|
|
|
|
IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action);
|
2006-11-09 18:56:43 +00:00
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2008-10-15 02:05:31 +00:00
|
|
|
/// setCondCodeAction - Indicate that the specified condition code is or isn't
|
|
|
|
/// supported on the target and indicate what to do about it.
|
2009-08-11 20:47:22 +00:00
|
|
|
void setCondCodeAction(ISD::CondCode CC, MVT VT,
|
2009-08-10 20:46:15 +00:00
|
|
|
LegalizeAction Action) {
|
2010-11-03 12:17:33 +00:00
|
|
|
assert(VT < MVT::LAST_VALUETYPE &&
|
2008-10-15 02:05:31 +00:00
|
|
|
(unsigned)CC < array_lengthof(CondCodeActions) &&
|
|
|
|
"Table isn't big enough!");
|
2012-09-14 15:36:50 +00:00
|
|
|
/// The lower 5 bits of the SimpleTy index into Nth 2bit set from the 64bit
|
|
|
|
/// value and the upper 27 bits index into the second dimension of the
|
|
|
|
/// array to select what 64bit value to use.
|
2012-09-13 15:24:43 +00:00
|
|
|
CondCodeActions[(unsigned)CC][VT.SimpleTy >> 5]
|
|
|
|
&= ~(uint64_t(3UL) << (VT.SimpleTy & 0x1F)*2);
|
|
|
|
CondCodeActions[(unsigned)CC][VT.SimpleTy >> 5]
|
|
|
|
|= (uint64_t)Action << (VT.SimpleTy & 0x1F)*2;
|
2008-10-15 02:05:31 +00:00
|
|
|
}
|
|
|
|
|
2006-04-04 00:25:10 +00:00
|
|
|
/// AddPromotedToType - If Opc/OrigVT is specified as being promoted, the
|
|
|
|
/// promotion code defaults to trying a larger integer/fp until it can find
|
|
|
|
/// one that works. If that default is insufficient, this method can be used
|
|
|
|
/// by the target to override the default.
|
2009-08-11 20:47:22 +00:00
|
|
|
void AddPromotedToType(unsigned Opc, MVT OrigVT, MVT DestVT) {
|
|
|
|
PromoteToType[std::make_pair(Opc, OrigVT.SimpleTy)] = DestVT.SimpleTy;
|
2006-04-04 00:25:10 +00:00
|
|
|
}
|
2005-01-07 07:44:22 +00:00
|
|
|
|
2006-03-01 04:52:55 +00:00
|
|
|
/// setTargetDAGCombine - Targets should invoke this method for each target
|
|
|
|
/// independent node that they want to provide a custom DAG combiner for by
|
|
|
|
/// implementing the PerformDAGCombine virtual method.
|
|
|
|
void setTargetDAGCombine(ISD::NodeType NT) {
|
2008-01-25 10:20:53 +00:00
|
|
|
assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray));
|
2006-03-01 04:52:55 +00:00
|
|
|
TargetDAGCombineArray[NT >> 3] |= 1 << (NT&7);
|
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-09-04 06:21:35 +00:00
|
|
|
/// setJumpBufSize - Set the target's required jmp_buf buffer size (in
|
|
|
|
/// bytes); default is 200
|
|
|
|
void setJumpBufSize(unsigned Size) {
|
|
|
|
JumpBufSize = Size;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// setJumpBufAlignment - Set the target's required jmp_buf buffer
|
|
|
|
/// alignment (in bytes); default is 0
|
|
|
|
void setJumpBufAlignment(unsigned Align) {
|
|
|
|
JumpBufAlignment = Align;
|
|
|
|
}
|
2007-05-16 23:45:53 +00:00
|
|
|
|
2011-10-17 18:23:13 +00:00
|
|
|
/// setMinFunctionAlignment - Set the target's minimum function alignment (in
|
|
|
|
/// log2(bytes))
|
2011-05-06 20:34:06 +00:00
|
|
|
void setMinFunctionAlignment(unsigned Align) {
|
|
|
|
MinFunctionAlignment = Align;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// setPrefFunctionAlignment - Set the target's preferred function alignment.
|
2011-05-07 19:23:14 +00:00
|
|
|
/// This should be set if there is a performance benefit to
|
2011-10-17 18:23:13 +00:00
|
|
|
/// higher-than-minimum alignment (in log2(bytes))
|
2011-05-06 20:34:06 +00:00
|
|
|
void setPrefFunctionAlignment(unsigned Align) {
|
|
|
|
PrefFunctionAlignment = Align;
|
|
|
|
}
|
|
|
|
|
2008-02-28 00:43:03 +00:00
|
|
|
/// setPrefLoopAlignment - Set the target's preferred loop alignment. Default
|
|
|
|
/// alignment is zero, it means the target does not care about loop alignment.
|
2011-10-17 18:23:13 +00:00
|
|
|
/// The alignment is specified in log2(bytes).
|
2008-02-28 00:43:03 +00:00
|
|
|
void setPrefLoopAlignment(unsigned Align) {
|
|
|
|
PrefLoopAlignment = Align;
|
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2010-07-11 04:01:49 +00:00
|
|
|
/// setMinStackArgumentAlignment - Set the minimum stack alignment of an
|
2011-10-17 18:23:13 +00:00
|
|
|
/// argument (in log2(bytes)).
|
2010-07-11 04:01:49 +00:00
|
|
|
void setMinStackArgumentAlignment(unsigned Align) {
|
|
|
|
MinStackArgumentAlignment = Align;
|
|
|
|
}
|
|
|
|
|
2010-06-23 16:07:42 +00:00
|
|
|
/// setShouldFoldAtomicFences - Set if the target's implementation of the
|
|
|
|
/// atomic operation intrinsics includes locking. Default is false.
|
|
|
|
void setShouldFoldAtomicFences(bool fold) {
|
|
|
|
ShouldFoldAtomicFences = fold;
|
|
|
|
}
|
|
|
|
|
2012-07-23 08:51:15 +00:00
|
|
|
/// setInsertFencesForAtomic - Set if the DAG builder should
|
2011-08-03 21:06:02 +00:00
|
|
|
/// automatically insert fences and reduce the order of atomic memory
|
|
|
|
/// operations to Monotonic.
|
|
|
|
void setInsertFencesForAtomic(bool fence) {
|
|
|
|
InsertFencesForAtomic = fence;
|
|
|
|
}
|
|
|
|
|
2005-01-07 07:44:22 +00:00
|
|
|
public:
|
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// Lowering methods - These methods must be implemented by targets so that
|
2012-11-05 02:59:23 +00:00
|
|
|
// the SelectionDAGBuilder code knows how to lower these.
|
2005-01-07 07:44:22 +00:00
|
|
|
//
|
|
|
|
|
Major calling convention code refactoring.
Instead of awkwardly encoding calling-convention information with ISD::CALL,
ISD::FORMAL_ARGUMENTS, ISD::RET, and ISD::ARG_FLAGS nodes, TargetLowering
provides three virtual functions for targets to override:
LowerFormalArguments, LowerCall, and LowerRet, which replace the custom
lowering done on the special nodes. They provide the same information, but
in a more immediately usable format.
This also reworks much of the target-independent tail call logic. The
decision of whether or not to perform a tail call is now cleanly split
between target-independent portions, and the target dependent portion
in IsEligibleForTailCallOptimization.
This also synchronizes all in-tree targets, to help enable future
refactoring and feature work.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@78142 91177308-0d34-0410-b5e6-96231b3b80d8
2009-08-05 01:29:28 +00:00
|
|
|
/// LowerFormalArguments - This hook must be implemented to lower the
|
|
|
|
/// incoming (formal) arguments, described by the Ins array, into the
|
|
|
|
/// specified DAG. The implementation should fill in the InVals array
|
|
|
|
/// with legal-type argument values, and return the resulting token
|
|
|
|
/// chain value.
|
|
|
|
///
|
|
|
|
virtual SDValue
|
2011-10-20 08:57:54 +00:00
|
|
|
LowerFormalArguments(SDValue /*Chain*/, CallingConv::ID /*CallConv*/,
|
|
|
|
bool /*isVarArg*/,
|
|
|
|
const SmallVectorImpl<ISD::InputArg> &/*Ins*/,
|
|
|
|
DebugLoc /*dl*/, SelectionDAG &/*DAG*/,
|
|
|
|
SmallVectorImpl<SDValue> &/*InVals*/) const {
|
2012-02-05 22:14:15 +00:00
|
|
|
llvm_unreachable("Not Implemented");
|
Major calling convention code refactoring.
Instead of awkwardly encoding calling-convention information with ISD::CALL,
ISD::FORMAL_ARGUMENTS, ISD::RET, and ISD::ARG_FLAGS nodes, TargetLowering
provides three virtual functions for targets to override:
LowerFormalArguments, LowerCall, and LowerRet, which replace the custom
lowering done on the special nodes. They provide the same information, but
in a more immediately usable format.
This also reworks much of the target-independent tail call logic. The
decision of whether or not to perform a tail call is now cleanly split
between target-independent portions, and the target dependent portion
in IsEligibleForTailCallOptimization.
This also synchronizes all in-tree targets, to help enable future
refactoring and feature work.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@78142 91177308-0d34-0410-b5e6-96231b3b80d8
2009-08-05 01:29:28 +00:00
|
|
|
}
|
2005-01-07 07:44:22 +00:00
|
|
|
|
2006-12-31 05:23:18 +00:00
|
|
|
struct ArgListEntry {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Node;
|
2011-07-18 04:54:35 +00:00
|
|
|
Type* Ty;
|
2008-06-30 20:33:57 +00:00
|
|
|
bool isSExt : 1;
|
|
|
|
bool isZExt : 1;
|
|
|
|
bool isInReg : 1;
|
|
|
|
bool isSRet : 1;
|
|
|
|
bool isNest : 1;
|
|
|
|
bool isByVal : 1;
|
2008-02-22 17:49:45 +00:00
|
|
|
uint16_t Alignment;
|
2007-02-01 08:39:52 +00:00
|
|
|
|
2007-07-27 12:58:54 +00:00
|
|
|
ArgListEntry() : isSExt(false), isZExt(false), isInReg(false),
|
2008-02-22 17:49:45 +00:00
|
|
|
isSRet(false), isNest(false), isByVal(false), Alignment(0) { }
|
2006-12-31 05:23:18 +00:00
|
|
|
};
|
|
|
|
typedef std::vector<ArgListEntry> ArgListTy;
|
2012-05-25 16:35:28 +00:00
|
|
|
|
|
|
|
/// CallLoweringInfo - This structure contains all information that is
|
|
|
|
/// necessary for lowering calls. It is passed to TLI::LowerCallTo when the
|
|
|
|
/// SelectionDAG builder needs to lower a call, and targets will see this
|
|
|
|
/// struct in their LowerCall implementation.
|
|
|
|
struct CallLoweringInfo {
|
|
|
|
SDValue Chain;
|
|
|
|
Type *RetTy;
|
|
|
|
bool RetSExt : 1;
|
|
|
|
bool RetZExt : 1;
|
|
|
|
bool IsVarArg : 1;
|
|
|
|
bool IsInReg : 1;
|
|
|
|
bool DoesNotReturn : 1;
|
|
|
|
bool IsReturnValueUsed : 1;
|
|
|
|
|
|
|
|
// IsTailCall should be modified by implementations of
|
|
|
|
// TargetLowering::LowerCall that perform tail call conversions.
|
|
|
|
bool IsTailCall;
|
|
|
|
|
|
|
|
unsigned NumFixedArgs;
|
|
|
|
CallingConv::ID CallConv;
|
|
|
|
SDValue Callee;
|
|
|
|
ArgListTy &Args;
|
|
|
|
SelectionDAG &DAG;
|
|
|
|
DebugLoc DL;
|
|
|
|
ImmutableCallSite *CS;
|
|
|
|
SmallVector<ISD::OutputArg, 32> Outs;
|
|
|
|
SmallVector<SDValue, 32> OutVals;
|
|
|
|
SmallVector<ISD::InputArg, 32> Ins;
|
|
|
|
|
|
|
|
|
|
|
|
/// CallLoweringInfo - Constructs a call lowering context based on the
|
|
|
|
/// ImmutableCallSite \p cs.
|
|
|
|
CallLoweringInfo(SDValue chain, Type *retTy,
|
|
|
|
FunctionType *FTy, bool isTailCall, SDValue callee,
|
|
|
|
ArgListTy &args, SelectionDAG &dag, DebugLoc dl,
|
|
|
|
ImmutableCallSite &cs)
|
2012-10-09 21:38:14 +00:00
|
|
|
: Chain(chain), RetTy(retTy), RetSExt(cs.paramHasAttr(0, Attributes::SExt)),
|
|
|
|
RetZExt(cs.paramHasAttr(0, Attributes::ZExt)), IsVarArg(FTy->isVarArg()),
|
|
|
|
IsInReg(cs.paramHasAttr(0, Attributes::InReg)),
|
2012-05-25 16:35:28 +00:00
|
|
|
DoesNotReturn(cs.doesNotReturn()),
|
|
|
|
IsReturnValueUsed(!cs.getInstruction()->use_empty()),
|
|
|
|
IsTailCall(isTailCall), NumFixedArgs(FTy->getNumParams()),
|
|
|
|
CallConv(cs.getCallingConv()), Callee(callee), Args(args), DAG(dag),
|
|
|
|
DL(dl), CS(&cs) {}
|
|
|
|
|
|
|
|
/// CallLoweringInfo - Constructs a call lowering context based on the
|
|
|
|
/// provided call information.
|
|
|
|
CallLoweringInfo(SDValue chain, Type *retTy, bool retSExt, bool retZExt,
|
|
|
|
bool isVarArg, bool isInReg, unsigned numFixedArgs,
|
|
|
|
CallingConv::ID callConv, bool isTailCall,
|
|
|
|
bool doesNotReturn, bool isReturnValueUsed, SDValue callee,
|
|
|
|
ArgListTy &args, SelectionDAG &dag, DebugLoc dl)
|
|
|
|
: Chain(chain), RetTy(retTy), RetSExt(retSExt), RetZExt(retZExt),
|
|
|
|
IsVarArg(isVarArg), IsInReg(isInReg), DoesNotReturn(doesNotReturn),
|
|
|
|
IsReturnValueUsed(isReturnValueUsed), IsTailCall(isTailCall),
|
|
|
|
NumFixedArgs(numFixedArgs), CallConv(callConv), Callee(callee),
|
|
|
|
Args(args), DAG(dag), DL(dl), CS(NULL) {}
|
|
|
|
};
|
|
|
|
|
|
|
|
/// LowerCallTo - This function lowers an abstract call to a function into an
|
|
|
|
/// actual call. This returns a pair of operands. The first element is the
|
|
|
|
/// return value for the function (if RetTy is not VoidTy). The second
|
|
|
|
/// element is the outgoing token chain. It calls LowerCall to do the actual
|
|
|
|
/// lowering.
|
|
|
|
std::pair<SDValue, SDValue> LowerCallTo(CallLoweringInfo &CLI) const;
|
Major calling convention code refactoring.
Instead of awkwardly encoding calling-convention information with ISD::CALL,
ISD::FORMAL_ARGUMENTS, ISD::RET, and ISD::ARG_FLAGS nodes, TargetLowering
provides three virtual functions for targets to override:
LowerFormalArguments, LowerCall, and LowerRet, which replace the custom
lowering done on the special nodes. They provide the same information, but
in a more immediately usable format.
This also reworks much of the target-independent tail call logic. The
decision of whether or not to perform a tail call is now cleanly split
between target-independent portions, and the target dependent portion
in IsEligibleForTailCallOptimization.
This also synchronizes all in-tree targets, to help enable future
refactoring and feature work.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@78142 91177308-0d34-0410-b5e6-96231b3b80d8
2009-08-05 01:29:28 +00:00
|
|
|
|
|
|
|
/// LowerCall - This hook must be implemented to lower calls into the
|
|
|
|
/// the specified DAG. The outgoing arguments to the call are described
|
|
|
|
/// by the Outs array, and the values to be returned by the call are
|
|
|
|
/// described by the Ins array. The implementation should fill in the
|
|
|
|
/// InVals array with legal-type return values from the call, and return
|
|
|
|
/// the resulting token chain value.
|
|
|
|
virtual SDValue
|
2012-05-25 16:35:28 +00:00
|
|
|
LowerCall(CallLoweringInfo &/*CLI*/,
|
2011-10-20 08:57:54 +00:00
|
|
|
SmallVectorImpl<SDValue> &/*InVals*/) const {
|
2012-02-05 22:14:15 +00:00
|
|
|
llvm_unreachable("Not Implemented");
|
Major calling convention code refactoring.
Instead of awkwardly encoding calling-convention information with ISD::CALL,
ISD::FORMAL_ARGUMENTS, ISD::RET, and ISD::ARG_FLAGS nodes, TargetLowering
provides three virtual functions for targets to override:
LowerFormalArguments, LowerCall, and LowerRet, which replace the custom
lowering done on the special nodes. They provide the same information, but
in a more immediately usable format.
This also reworks much of the target-independent tail call logic. The
decision of whether or not to perform a tail call is now cleanly split
between target-independent portions, and the target dependent portion
in IsEligibleForTailCallOptimization.
This also synchronizes all in-tree targets, to help enable future
refactoring and feature work.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@78142 91177308-0d34-0410-b5e6-96231b3b80d8
2009-08-05 01:29:28 +00:00
|
|
|
}
|
|
|
|
|
2011-02-28 17:17:53 +00:00
|
|
|
/// HandleByVal - Target-specific cleanup for formal ByVal parameters.
|
Issue:
Stack is formed improperly for long structures passed as byval arguments for
EABI mode.
If we took AAPCS reference, we can found the next statements:
A: "If the argument requires double-word alignment (8-byte), the NCRN (Next
Core Register Number) is rounded up to the next even register number." (5.5
Parameter Passing, Stage C, C.3).
B: "The alignment of an aggregate shall be the alignment of its most-aligned
component." (4.3 Composite Types, 4.3.1 Aggregates).
So if we have structure with doubles (9 double fields) and 3 Core unused
registers (r1, r2, r3): caller should use r2 and r3 registers only.
Currently r1,r2,r3 set is used, but it is invalid.
Callee VA routine should also use r2 and r3 regs only. All is ok here. This
behaviour is guessed by rounding up SP address with ADD+BFC operations.
Fix:
Main fix is in ARMTargetLowering::HandleByVal. If we detected AAPCS mode and
8 byte alignment, we waste odd registers then.
P.S.:
I also improved LDRB_POST_IMM regression test. Since ldrb instruction will
not generated by current regression test after this patch.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@166018 91177308-0d34-0410-b5e6-96231b3b80d8
2012-10-16 07:16:47 +00:00
|
|
|
virtual void HandleByVal(CCState *, unsigned &, unsigned) const {}
|
2011-02-28 17:17:53 +00:00
|
|
|
|
2009-11-07 02:11:54 +00:00
|
|
|
/// CanLowerReturn - This hook should be implemented to check whether the
|
|
|
|
/// return values described by the Outs array can fit into the return
|
|
|
|
/// registers. If false is returned, an sret-demotion is performed.
|
|
|
|
///
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool CanLowerReturn(CallingConv::ID /*CallConv*/,
|
2012-07-19 00:01:33 +00:00
|
|
|
MachineFunction &/*MF*/, bool /*isVarArg*/,
|
2011-10-20 08:57:54 +00:00
|
|
|
const SmallVectorImpl<ISD::OutputArg> &/*Outs*/,
|
|
|
|
LLVMContext &/*Context*/) const
|
2009-11-07 02:11:54 +00:00
|
|
|
{
|
|
|
|
// Return true by default to get preexisting behavior.
|
|
|
|
return true;
|
|
|
|
}
|
2010-04-17 15:26:15 +00:00
|
|
|
|
Major calling convention code refactoring.
Instead of awkwardly encoding calling-convention information with ISD::CALL,
ISD::FORMAL_ARGUMENTS, ISD::RET, and ISD::ARG_FLAGS nodes, TargetLowering
provides three virtual functions for targets to override:
LowerFormalArguments, LowerCall, and LowerRet, which replace the custom
lowering done on the special nodes. They provide the same information, but
in a more immediately usable format.
This also reworks much of the target-independent tail call logic. The
decision of whether or not to perform a tail call is now cleanly split
between target-independent portions, and the target dependent portion
in IsEligibleForTailCallOptimization.
This also synchronizes all in-tree targets, to help enable future
refactoring and feature work.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@78142 91177308-0d34-0410-b5e6-96231b3b80d8
2009-08-05 01:29:28 +00:00
|
|
|
/// LowerReturn - This hook must be implemented to lower outgoing
|
|
|
|
/// return values, described by the Outs array, into the specified
|
|
|
|
/// DAG. The implementation should return the resulting token chain
|
|
|
|
/// value.
|
|
|
|
///
|
|
|
|
virtual SDValue
|
2011-10-20 08:57:54 +00:00
|
|
|
LowerReturn(SDValue /*Chain*/, CallingConv::ID /*CallConv*/,
|
|
|
|
bool /*isVarArg*/,
|
|
|
|
const SmallVectorImpl<ISD::OutputArg> &/*Outs*/,
|
|
|
|
const SmallVectorImpl<SDValue> &/*OutVals*/,
|
|
|
|
DebugLoc /*dl*/, SelectionDAG &/*DAG*/) const {
|
2012-02-05 22:14:15 +00:00
|
|
|
llvm_unreachable("Not Implemented");
|
Major calling convention code refactoring.
Instead of awkwardly encoding calling-convention information with ISD::CALL,
ISD::FORMAL_ARGUMENTS, ISD::RET, and ISD::ARG_FLAGS nodes, TargetLowering
provides three virtual functions for targets to override:
LowerFormalArguments, LowerCall, and LowerRet, which replace the custom
lowering done on the special nodes. They provide the same information, but
in a more immediately usable format.
This also reworks much of the target-independent tail call logic. The
decision of whether or not to perform a tail call is now cleanly split
between target-independent portions, and the target dependent portion
in IsEligibleForTailCallOptimization.
This also synchronizes all in-tree targets, to help enable future
refactoring and feature work.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@78142 91177308-0d34-0410-b5e6-96231b3b80d8
2009-08-05 01:29:28 +00:00
|
|
|
}
|
2007-11-05 23:12:20 +00:00
|
|
|
|
2010-11-30 23:55:39 +00:00
|
|
|
/// isUsedByReturnOnly - Return true if result of the specified node is used
|
2012-04-10 01:51:00 +00:00
|
|
|
/// by a return node only. It also compute and return the input chain for the
|
|
|
|
/// tail call.
|
|
|
|
/// This is used to determine whether it is possible
|
2010-11-30 23:55:39 +00:00
|
|
|
/// to codegen a libcall as tail call at legalization time.
|
2012-04-10 01:51:00 +00:00
|
|
|
virtual bool isUsedByReturnOnly(SDNode *, SDValue &Chain) const {
|
2010-11-30 23:55:39 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2011-03-21 01:19:09 +00:00
|
|
|
/// mayBeEmittedAsTailCall - Return true if the target may be able emit the
|
|
|
|
/// call instruction as a tail call. This is used by optimization passes to
|
|
|
|
/// determine if it's profitable to duplicate return instructions to enable
|
|
|
|
/// tailcall optimization.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool mayBeEmittedAsTailCall(CallInst *) const {
|
2011-03-21 01:19:09 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2011-03-17 14:21:56 +00:00
|
|
|
/// getTypeForExtArgOrReturn - Return the type that should be used to zero or
|
2011-03-16 22:20:18 +00:00
|
|
|
/// sign extend a zeroext/signext integer argument or return value.
|
|
|
|
/// FIXME: Most C calling convention requires the return type to be promoted,
|
|
|
|
/// but this is not true all the time, e.g. i1 on x86-64. It is also not
|
|
|
|
/// necessary for non-C calling conventions. The frontend should handle this
|
|
|
|
/// and include all of the necessary information.
|
2012-12-11 11:14:33 +00:00
|
|
|
virtual EVT getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
|
2011-10-20 08:57:54 +00:00
|
|
|
ISD::NodeType /*ExtendKind*/) const {
|
2012-12-11 11:14:33 +00:00
|
|
|
EVT MinVT = getRegisterType(Context, MVT::i32);
|
2011-03-17 14:53:37 +00:00
|
|
|
return VT.bitsLT(MinVT) ? MinVT : VT;
|
2011-03-16 22:20:18 +00:00
|
|
|
}
|
|
|
|
|
2009-01-21 09:00:29 +00:00
|
|
|
/// LowerOperationWrapper - This callback is invoked by the type legalizer
|
|
|
|
/// to legalize nodes with an illegal operand type but legal result types.
|
2009-01-21 04:48:39 +00:00
|
|
|
/// It replaces the LowerOperation callback in the type Legalizer.
|
2009-01-21 09:00:29 +00:00
|
|
|
/// The reason we can not do away with LowerOperation entirely is that
|
2009-01-21 04:48:39 +00:00
|
|
|
/// LegalizeDAG isn't yet ready to use this callback.
|
2009-01-21 09:00:29 +00:00
|
|
|
/// TODO: Consider merging with ReplaceNodeResults.
|
|
|
|
|
|
|
|
/// The target places new result values for the node in Results (their number
|
|
|
|
/// and types must exactly match those of the original return values of
|
|
|
|
/// the node), or leaves Results empty, which indicates that the node is not
|
|
|
|
/// to be custom lowered after all.
|
|
|
|
/// The default implementation calls LowerOperation.
|
|
|
|
virtual void LowerOperationWrapper(SDNode *N,
|
2009-01-21 04:48:39 +00:00
|
|
|
SmallVectorImpl<SDValue> &Results,
|
2010-04-17 15:26:15 +00:00
|
|
|
SelectionDAG &DAG) const;
|
2009-01-21 04:48:39 +00:00
|
|
|
|
2010-06-23 15:49:18 +00:00
|
|
|
/// LowerOperation - This callback is invoked for operations that are
|
2006-01-28 03:37:03 +00:00
|
|
|
/// unsupported by the target, which are registered to use 'custom' lowering,
|
|
|
|
/// and whose defined values are all legal.
|
2005-01-16 07:27:49 +00:00
|
|
|
/// If the target has no operations that require custom lowering, it need not
|
|
|
|
/// implement this. The default implementation of this aborts.
|
2010-04-17 15:26:15 +00:00
|
|
|
virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
|
2005-01-19 03:36:03 +00:00
|
|
|
|
2008-12-01 11:39:25 +00:00
|
|
|
/// ReplaceNodeResults - This callback is invoked when a node result type is
|
|
|
|
/// illegal for the target, and the operation was registered to use 'custom'
|
|
|
|
/// lowering for that result type. The target places new result values for
|
|
|
|
/// the node in Results (their number and types must exactly match those of
|
|
|
|
/// the original return values of the node), or leaves Results empty, which
|
|
|
|
/// indicates that the node is not to be custom lowered after all.
|
2007-10-19 03:31:45 +00:00
|
|
|
///
|
|
|
|
/// If the target has no operations that require custom lowering, it need not
|
2008-07-04 11:47:58 +00:00
|
|
|
/// implement this. The default implementation aborts.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual void ReplaceNodeResults(SDNode * /*N*/,
|
|
|
|
SmallVectorImpl<SDValue> &/*Results*/,
|
|
|
|
SelectionDAG &/*DAG*/) const {
|
2012-02-05 22:14:15 +00:00
|
|
|
llvm_unreachable("ReplaceNodeResults not implemented for this target!");
|
2007-11-24 07:07:01 +00:00
|
|
|
}
|
2008-07-04 11:47:58 +00:00
|
|
|
|
2005-12-21 23:15:41 +00:00
|
|
|
/// getTargetNodeName() - This method returns the name of a target specific
|
|
|
|
/// DAG node.
|
2005-12-20 06:22:03 +00:00
|
|
|
virtual const char *getTargetNodeName(unsigned Opcode) const;
|
|
|
|
|
2008-08-19 21:26:29 +00:00
|
|
|
/// createFastISel - This method returns a target specific FastISel object,
|
|
|
|
/// or null if the target does not support "fast" ISel.
|
2012-08-03 04:06:28 +00:00
|
|
|
virtual FastISel *createFastISel(FunctionLoweringInfo &,
|
|
|
|
const TargetLibraryInfo *) const {
|
2008-09-03 23:12:08 +00:00
|
|
|
return 0;
|
|
|
|
}
|
2008-08-19 21:26:29 +00:00
|
|
|
|
2006-01-26 20:27:33 +00:00
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// Inline Asm Support hooks
|
|
|
|
//
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2009-07-20 17:51:36 +00:00
|
|
|
/// ExpandInlineAsm - This hook allows the target to expand an inline asm
|
|
|
|
/// call to be explicit llvm code if it wants to. This is useful for
|
|
|
|
/// turning simple inline asms into LLVM intrinsics, which gives the
|
|
|
|
/// compiler more information about the behavior of the code.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool ExpandInlineAsm(CallInst *) const {
|
2009-07-20 17:51:36 +00:00
|
|
|
return false;
|
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-02-04 02:12:09 +00:00
|
|
|
enum ConstraintType {
|
2008-11-13 21:52:36 +00:00
|
|
|
C_Register, // Constraint represents specific register(s).
|
|
|
|
C_RegisterClass, // Constraint represents any of register(s) in class.
|
2006-02-24 01:10:14 +00:00
|
|
|
C_Memory, // Memory constraint.
|
2006-02-04 02:12:09 +00:00
|
|
|
C_Other, // Something else.
|
|
|
|
C_Unknown // Unsupported constraint.
|
|
|
|
};
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2010-10-29 17:29:13 +00:00
|
|
|
enum ConstraintWeight {
|
|
|
|
// Generic weights.
|
|
|
|
CW_Invalid = -1, // No match.
|
|
|
|
CW_Okay = 0, // Acceptable.
|
|
|
|
CW_Good = 1, // Good weight.
|
|
|
|
CW_Better = 2, // Better weight.
|
|
|
|
CW_Best = 3, // Best weight.
|
2011-02-25 21:41:48 +00:00
|
|
|
|
2010-10-29 17:29:13 +00:00
|
|
|
// Well-known weights.
|
|
|
|
CW_SpecificReg = CW_Okay, // Specific register operands.
|
|
|
|
CW_Register = CW_Good, // Register operands.
|
|
|
|
CW_Memory = CW_Better, // Memory operands.
|
|
|
|
CW_Constant = CW_Best, // Constant operand.
|
|
|
|
CW_Default = CW_Okay // Default or don't know type.
|
|
|
|
};
|
|
|
|
|
2008-02-26 02:33:44 +00:00
|
|
|
/// AsmOperandInfo - This contains information for each constraint that we are
|
|
|
|
/// lowering.
|
|
|
|
struct AsmOperandInfo : public InlineAsm::ConstraintInfo {
|
|
|
|
/// ConstraintCode - This contains the actual string for the code, like "m".
|
2008-10-17 16:21:11 +00:00
|
|
|
/// TargetLowering picks the 'best' code from ConstraintInfo::Codes that
|
|
|
|
/// most closely matches the operand.
|
2008-02-26 02:33:44 +00:00
|
|
|
std::string ConstraintCode;
|
|
|
|
|
|
|
|
/// ConstraintType - Information about the constraint code, e.g. Register,
|
|
|
|
/// RegisterClass, Memory, Other, Unknown.
|
|
|
|
TargetLowering::ConstraintType ConstraintType;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2008-02-26 02:33:44 +00:00
|
|
|
/// CallOperandval - If this is the result output operand or a
|
|
|
|
/// clobber, this is null, otherwise it is the incoming operand to the
|
|
|
|
/// CallInst. This gets modified as the asm is processed.
|
|
|
|
Value *CallOperandVal;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2008-02-26 02:33:44 +00:00
|
|
|
/// ConstraintVT - The ValueType for the operand value.
|
2009-08-10 22:56:29 +00:00
|
|
|
EVT ConstraintVT;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2008-10-17 16:47:46 +00:00
|
|
|
/// isMatchingInputConstraint - Return true of this is an input operand that
|
|
|
|
/// is a matching constraint like "4".
|
|
|
|
bool isMatchingInputConstraint() const;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2008-10-17 16:21:11 +00:00
|
|
|
/// getMatchedOperand - If this is an input matching constraint, this method
|
|
|
|
/// returns the output operand it matches.
|
|
|
|
unsigned getMatchedOperand() const;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2010-09-13 18:15:37 +00:00
|
|
|
/// Copy constructor for copying from an AsmOperandInfo.
|
|
|
|
AsmOperandInfo(const AsmOperandInfo &info)
|
|
|
|
: InlineAsm::ConstraintInfo(info),
|
|
|
|
ConstraintCode(info.ConstraintCode),
|
|
|
|
ConstraintType(info.ConstraintType),
|
|
|
|
CallOperandVal(info.CallOperandVal),
|
|
|
|
ConstraintVT(info.ConstraintVT) {
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Copy constructor for copying from a ConstraintInfo.
|
2008-02-26 02:33:44 +00:00
|
|
|
AsmOperandInfo(const InlineAsm::ConstraintInfo &info)
|
2010-06-23 15:49:18 +00:00
|
|
|
: InlineAsm::ConstraintInfo(info),
|
2008-02-26 02:33:44 +00:00
|
|
|
ConstraintType(TargetLowering::C_Unknown),
|
2009-08-11 20:47:22 +00:00
|
|
|
CallOperandVal(0), ConstraintVT(MVT::Other) {
|
2008-02-26 02:33:44 +00:00
|
|
|
}
|
|
|
|
};
|
2011-02-25 21:41:48 +00:00
|
|
|
|
2010-10-29 23:37:38 +00:00
|
|
|
typedef std::vector<AsmOperandInfo> AsmOperandInfoVector;
|
2011-02-25 21:41:48 +00:00
|
|
|
|
2010-09-13 18:15:37 +00:00
|
|
|
/// ParseConstraints - Split up the constraint string from the inline
|
|
|
|
/// assembly value into the specific constraints and their prefixes,
|
|
|
|
/// and also tie in the associated operand values.
|
|
|
|
/// If this returns an empty vector, and if the constraint string itself
|
|
|
|
/// isn't empty, there was an error parsing.
|
2010-10-29 17:29:13 +00:00
|
|
|
virtual AsmOperandInfoVector ParseConstraints(ImmutableCallSite CS) const;
|
2011-02-25 21:41:48 +00:00
|
|
|
|
2010-10-29 17:29:13 +00:00
|
|
|
/// Examine constraint type and operand type and determine a weight value.
|
2010-09-13 18:15:37 +00:00
|
|
|
/// The operand object must already have been set up with the operand type.
|
2010-10-29 17:29:13 +00:00
|
|
|
virtual ConstraintWeight getMultipleConstraintMatchWeight(
|
2010-09-13 18:15:37 +00:00
|
|
|
AsmOperandInfo &info, int maIndex) const;
|
2011-02-25 21:41:48 +00:00
|
|
|
|
2010-10-29 17:29:13 +00:00
|
|
|
/// Examine constraint string and operand type and determine a weight value.
|
2010-09-13 18:15:37 +00:00
|
|
|
/// The operand object must already have been set up with the operand type.
|
2010-10-29 17:29:13 +00:00
|
|
|
virtual ConstraintWeight getSingleConstraintMatchWeight(
|
2010-09-13 18:15:37 +00:00
|
|
|
AsmOperandInfo &info, const char *constraint) const;
|
2008-02-26 02:33:44 +00:00
|
|
|
|
2008-04-27 00:09:47 +00:00
|
|
|
/// ComputeConstraintToUse - Determines the constraint code and constraint
|
|
|
|
/// type to use for the specific AsmOperandInfo, setting
|
2008-04-27 00:37:18 +00:00
|
|
|
/// OpInfo.ConstraintCode and OpInfo.ConstraintType. If the actual operand
|
|
|
|
/// being passed in is available, it can be passed in as Op, otherwise an
|
2011-02-25 21:41:48 +00:00
|
|
|
/// empty SDValue can be passed.
|
2008-04-27 00:37:18 +00:00
|
|
|
virtual void ComputeConstraintToUse(AsmOperandInfo &OpInfo,
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op,
|
2008-04-27 00:37:18 +00:00
|
|
|
SelectionDAG *DAG = 0) const;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2007-03-25 02:14:49 +00:00
|
|
|
/// getConstraintType - Given a constraint, return the type of constraint it
|
|
|
|
/// is for this target.
|
|
|
|
virtual ConstraintType getConstraintType(const std::string &Constraint) const;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-02-22 00:56:39 +00:00
|
|
|
/// getRegForInlineAsmConstraint - Given a physical register constraint (e.g.
|
|
|
|
/// {edx}), return the register number and the register class for the
|
2006-11-02 01:39:10 +00:00
|
|
|
/// register.
|
|
|
|
///
|
|
|
|
/// Given a register class constraint, like 'r', if this corresponds directly
|
|
|
|
/// to an LLVM register class, return a register of 0 and the register class
|
|
|
|
/// pointer.
|
|
|
|
///
|
|
|
|
/// This should only be used for C_Register constraints. On error,
|
|
|
|
/// this returns a register number of 0 and a null register class pointer..
|
2010-06-23 15:49:18 +00:00
|
|
|
virtual std::pair<unsigned, const TargetRegisterClass*>
|
2006-02-22 00:56:39 +00:00
|
|
|
getRegForInlineAsmConstraint(const std::string &Constraint,
|
2009-08-10 22:56:29 +00:00
|
|
|
EVT VT) const;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2008-01-29 02:21:21 +00:00
|
|
|
/// LowerXConstraint - try to replace an X constraint, which matches anything,
|
|
|
|
/// with another that has more specific requirements based on the type of the
|
2008-04-26 23:02:14 +00:00
|
|
|
/// corresponding operand. This returns null if there is no replacement to
|
|
|
|
/// make.
|
2009-08-10 22:56:29 +00:00
|
|
|
virtual const char *LowerXConstraint(EVT ConstraintVT) const;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2007-08-25 00:47:38 +00:00
|
|
|
/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
|
2010-06-25 21:55:36 +00:00
|
|
|
/// vector. If it is invalid, don't add anything to Ops.
|
2011-06-02 23:16:42 +00:00
|
|
|
virtual void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint,
|
2008-07-27 21:46:04 +00:00
|
|
|
std::vector<SDValue> &Ops,
|
2008-04-26 23:02:14 +00:00
|
|
|
SelectionDAG &DAG) const;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2005-08-26 20:53:09 +00:00
|
|
|
//===--------------------------------------------------------------------===//
|
2009-10-29 18:10:34 +00:00
|
|
|
// Instruction Emitting Hooks
|
2005-08-26 20:53:09 +00:00
|
|
|
//
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2008-01-30 18:18:23 +00:00
|
|
|
// EmitInstrWithCustomInserter - This method should be implemented by targets
|
2009-10-29 18:10:34 +00:00
|
|
|
// that mark instructions with the 'usesCustomInserter' flag. These
|
2005-08-26 20:53:09 +00:00
|
|
|
// instructions are special in various ways, which require special support to
|
|
|
|
// insert. The specified MachineInstr is created but not inserted into any
|
2009-10-29 18:10:34 +00:00
|
|
|
// basic blocks, and this method is called to expand it into a sequence of
|
|
|
|
// instructions, potentially also creating new basic blocks and control flow.
|
2010-05-01 00:01:06 +00:00
|
|
|
virtual MachineBasicBlock *
|
|
|
|
EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const;
|
2005-04-21 20:59:05 +00:00
|
|
|
|
2011-08-30 19:09:48 +00:00
|
|
|
/// AdjustInstrPostInstrSelection - This method should be implemented by
|
|
|
|
/// targets that mark instructions with the 'hasPostISelHook' flag. These
|
|
|
|
/// instructions must be adjusted after instruction selection by target hooks.
|
|
|
|
/// e.g. To fill in optional defs for ARM 's' setting instructions.
|
|
|
|
virtual void
|
|
|
|
AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const;
|
|
|
|
|
2006-03-16 21:47:42 +00:00
|
|
|
//===--------------------------------------------------------------------===//
|
2007-03-30 23:14:50 +00:00
|
|
|
// Addressing mode description hooks (used by LSR etc).
|
2006-03-16 21:47:42 +00:00
|
|
|
//
|
2007-03-16 08:42:32 +00:00
|
|
|
|
2012-03-13 20:59:56 +00:00
|
|
|
/// GetAddrModeArguments - CodeGenPrepare sinks address calculations into the
|
|
|
|
/// same BB as Load/Store instructions reading the address. This allows as
|
|
|
|
/// much computation as possible to be done in the address mode for that
|
|
|
|
/// operand. This hook lets targets also pass back when this should be done
|
|
|
|
/// on intrinsics which load/store.
|
|
|
|
virtual bool GetAddrModeArguments(IntrinsicInst *I,
|
|
|
|
SmallVectorImpl<Value*> &Ops,
|
|
|
|
Type *&AccessTy) const {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2007-03-30 23:14:50 +00:00
|
|
|
/// isLegalAddressingMode - Return true if the addressing mode represented by
|
|
|
|
/// AM is legal for this target, for a load/store of the specified type.
|
2009-05-11 19:11:53 +00:00
|
|
|
/// The type may be VoidTy, in which case only return true if the addressing
|
|
|
|
/// mode is legal for a load/store of any legal type.
|
2007-03-30 23:14:50 +00:00
|
|
|
/// TODO: Handle pre/postinc as well.
|
2011-07-18 04:54:35 +00:00
|
|
|
virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const;
|
2006-03-16 21:47:42 +00:00
|
|
|
|
2011-10-06 02:47:18 +00:00
|
|
|
/// isLegalICmpImmediate - Return true if the specified immediate is legal
|
|
|
|
/// icmp immediate, that is the target has icmp instructions which can compare
|
|
|
|
/// a register against the immediate without having to materialize the
|
|
|
|
/// immediate into a register.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool isLegalICmpImmediate(int64_t) const {
|
2011-10-06 02:47:18 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// isLegalAddImmediate - Return true if the specified immediate is legal
|
|
|
|
/// add immediate, that is the target has add instructions which can add
|
|
|
|
/// a register with the immediate without having to materialize the
|
|
|
|
/// immediate into a register.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool isLegalAddImmediate(int64_t) const {
|
2011-10-06 02:47:18 +00:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
Loosen up iv reuse to allow reuse of the same stride but a larger type when truncating from the larger type to smaller type is free.
e.g.
Turns this loop:
LBB1_1: # entry.bb_crit_edge
xorl %ecx, %ecx
xorw %dx, %dx
movw %dx, %si
LBB1_2: # bb
movl L_X$non_lazy_ptr, %edi
movw %si, (%edi)
movl L_Y$non_lazy_ptr, %edi
movw %dx, (%edi)
addw $4, %dx
incw %si
incl %ecx
cmpl %eax, %ecx
jne LBB1_2 # bb
into
LBB1_1: # entry.bb_crit_edge
xorl %ecx, %ecx
xorw %dx, %dx
LBB1_2: # bb
movl L_X$non_lazy_ptr, %esi
movw %cx, (%esi)
movl L_Y$non_lazy_ptr, %esi
movw %dx, (%esi)
addw $4, %dx
incl %ecx
cmpl %eax, %ecx
jne LBB1_2 # bb
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43375 91177308-0d34-0410-b5e6-96231b3b80d8
2007-10-26 01:56:11 +00:00
|
|
|
/// isTruncateFree - Return true if it's free to truncate a value of
|
|
|
|
/// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
|
|
|
|
/// register EAX to i16 by referencing its sub-register AX.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool isTruncateFree(Type * /*Ty1*/, Type * /*Ty2*/) const {
|
Loosen up iv reuse to allow reuse of the same stride but a larger type when truncating from the larger type to smaller type is free.
e.g.
Turns this loop:
LBB1_1: # entry.bb_crit_edge
xorl %ecx, %ecx
xorw %dx, %dx
movw %dx, %si
LBB1_2: # bb
movl L_X$non_lazy_ptr, %edi
movw %si, (%edi)
movl L_Y$non_lazy_ptr, %edi
movw %dx, (%edi)
addw $4, %dx
incw %si
incl %ecx
cmpl %eax, %ecx
jne LBB1_2 # bb
into
LBB1_1: # entry.bb_crit_edge
xorl %ecx, %ecx
xorw %dx, %dx
LBB1_2: # bb
movl L_X$non_lazy_ptr, %esi
movw %cx, (%esi)
movl L_Y$non_lazy_ptr, %esi
movw %dx, (%esi)
addw $4, %dx
incl %ecx
cmpl %eax, %ecx
jne LBB1_2 # bb
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@43375 91177308-0d34-0410-b5e6-96231b3b80d8
2007-10-26 01:56:11 +00:00
|
|
|
return false;
|
|
|
|
}
|
2007-10-29 19:58:20 +00:00
|
|
|
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool isTruncateFree(EVT /*VT1*/, EVT /*VT2*/) const {
|
2007-10-29 19:58:20 +00:00
|
|
|
return false;
|
|
|
|
}
|
Implement support for using modeling implicit-zero-extension on x86-64
with SUBREG_TO_REG, teach SimpleRegisterCoalescing to coalesce
SUBREG_TO_REG instructions (which are similar to INSERT_SUBREG
instructions), and teach the DAGCombiner to take advantage of this on
targets which support it. This eliminates many redundant
zero-extension operations on x86-64.
This adds a new TargetLowering hook, isZExtFree. It's similar to
isTruncateFree, except it only applies to actual definitions, and not
no-op truncates which may not zero the high bits.
Also, this adds a new optimization to SimplifyDemandedBits: transform
operations like x+y into (zext (add (trunc x), (trunc y))) on targets
where all the casts are no-ops. In contexts where the high part of the
add is explicitly masked off, this allows the mask operation to be
eliminated. Fix the DAGCombiner to avoid undoing these transformations
to eliminate casts on targets where the casts are no-ops.
Also, this adds a new two-address lowering heuristic. Since
two-address lowering runs before coalescing, it helps to be able to
look through copies when deciding whether commuting and/or
three-address conversion are profitable.
Also, fix a bug in LiveInterval::MergeInClobberRanges. It didn't handle
the case that a clobber range extended both before and beyond an
existing live range. In that case, multiple live ranges need to be
added. This was exposed by the new subreg coalescing code.
Remove 2008-05-06-SpillerBug.ll. It was bugpoint-reduced, and the
spiller behavior it was looking for no longer occurrs with the new
instruction selection.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@68576 91177308-0d34-0410-b5e6-96231b3b80d8
2009-04-08 00:15:30 +00:00
|
|
|
|
|
|
|
/// isZExtFree - Return true if any actual instruction that defines a
|
2010-01-06 19:38:29 +00:00
|
|
|
/// value of type Ty1 implicitly zero-extends the value to Ty2 in the result
|
Implement support for using modeling implicit-zero-extension on x86-64
with SUBREG_TO_REG, teach SimpleRegisterCoalescing to coalesce
SUBREG_TO_REG instructions (which are similar to INSERT_SUBREG
instructions), and teach the DAGCombiner to take advantage of this on
targets which support it. This eliminates many redundant
zero-extension operations on x86-64.
This adds a new TargetLowering hook, isZExtFree. It's similar to
isTruncateFree, except it only applies to actual definitions, and not
no-op truncates which may not zero the high bits.
Also, this adds a new optimization to SimplifyDemandedBits: transform
operations like x+y into (zext (add (trunc x), (trunc y))) on targets
where all the casts are no-ops. In contexts where the high part of the
add is explicitly masked off, this allows the mask operation to be
eliminated. Fix the DAGCombiner to avoid undoing these transformations
to eliminate casts on targets where the casts are no-ops.
Also, this adds a new two-address lowering heuristic. Since
two-address lowering runs before coalescing, it helps to be able to
look through copies when deciding whether commuting and/or
three-address conversion are profitable.
Also, fix a bug in LiveInterval::MergeInClobberRanges. It didn't handle
the case that a clobber range extended both before and beyond an
existing live range. In that case, multiple live ranges need to be
added. This was exposed by the new subreg coalescing code.
Remove 2008-05-06-SpillerBug.ll. It was bugpoint-reduced, and the
spiller behavior it was looking for no longer occurrs with the new
instruction selection.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@68576 91177308-0d34-0410-b5e6-96231b3b80d8
2009-04-08 00:15:30 +00:00
|
|
|
/// register. This does not necessarily include registers defined in
|
|
|
|
/// unknown ways, such as incoming arguments, or copies from unknown
|
|
|
|
/// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this
|
|
|
|
/// does not necessarily apply to truncate instructions. e.g. on x86-64,
|
|
|
|
/// all instructions that define 32-bit values implicit zero-extend the
|
|
|
|
/// result out to 64 bits.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool isZExtFree(Type * /*Ty1*/, Type * /*Ty2*/) const {
|
Implement support for using modeling implicit-zero-extension on x86-64
with SUBREG_TO_REG, teach SimpleRegisterCoalescing to coalesce
SUBREG_TO_REG instructions (which are similar to INSERT_SUBREG
instructions), and teach the DAGCombiner to take advantage of this on
targets which support it. This eliminates many redundant
zero-extension operations on x86-64.
This adds a new TargetLowering hook, isZExtFree. It's similar to
isTruncateFree, except it only applies to actual definitions, and not
no-op truncates which may not zero the high bits.
Also, this adds a new optimization to SimplifyDemandedBits: transform
operations like x+y into (zext (add (trunc x), (trunc y))) on targets
where all the casts are no-ops. In contexts where the high part of the
add is explicitly masked off, this allows the mask operation to be
eliminated. Fix the DAGCombiner to avoid undoing these transformations
to eliminate casts on targets where the casts are no-ops.
Also, this adds a new two-address lowering heuristic. Since
two-address lowering runs before coalescing, it helps to be able to
look through copies when deciding whether commuting and/or
three-address conversion are profitable.
Also, fix a bug in LiveInterval::MergeInClobberRanges. It didn't handle
the case that a clobber range extended both before and beyond an
existing live range. In that case, multiple live ranges need to be
added. This was exposed by the new subreg coalescing code.
Remove 2008-05-06-SpillerBug.ll. It was bugpoint-reduced, and the
spiller behavior it was looking for no longer occurrs with the new
instruction selection.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@68576 91177308-0d34-0410-b5e6-96231b3b80d8
2009-04-08 00:15:30 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool isZExtFree(EVT /*VT1*/, EVT /*VT2*/) const {
|
Implement support for using modeling implicit-zero-extension on x86-64
with SUBREG_TO_REG, teach SimpleRegisterCoalescing to coalesce
SUBREG_TO_REG instructions (which are similar to INSERT_SUBREG
instructions), and teach the DAGCombiner to take advantage of this on
targets which support it. This eliminates many redundant
zero-extension operations on x86-64.
This adds a new TargetLowering hook, isZExtFree. It's similar to
isTruncateFree, except it only applies to actual definitions, and not
no-op truncates which may not zero the high bits.
Also, this adds a new optimization to SimplifyDemandedBits: transform
operations like x+y into (zext (add (trunc x), (trunc y))) on targets
where all the casts are no-ops. In contexts where the high part of the
add is explicitly masked off, this allows the mask operation to be
eliminated. Fix the DAGCombiner to avoid undoing these transformations
to eliminate casts on targets where the casts are no-ops.
Also, this adds a new two-address lowering heuristic. Since
two-address lowering runs before coalescing, it helps to be able to
look through copies when deciding whether commuting and/or
three-address conversion are profitable.
Also, fix a bug in LiveInterval::MergeInClobberRanges. It didn't handle
the case that a clobber range extended both before and beyond an
existing live range. In that case, multiple live ranges need to be
added. This was exposed by the new subreg coalescing code.
Remove 2008-05-06-SpillerBug.ll. It was bugpoint-reduced, and the
spiller behavior it was looking for no longer occurrs with the new
instruction selection.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@68576 91177308-0d34-0410-b5e6-96231b3b80d8
2009-04-08 00:15:30 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2012-12-06 19:13:27 +00:00
|
|
|
/// isZExtFree - Return true if zero-extending the specific node Val to type
|
|
|
|
/// VT2 is free (either because it's implicitly zero-extended such as ARM
|
|
|
|
/// ldrb / ldrh or because it's folded such as X86 zero-extending loads).
|
|
|
|
virtual bool isZExtFree(SDValue Val, EVT VT2) const {
|
|
|
|
return isZExtFree(Val.getValueType(), VT2);
|
|
|
|
}
|
|
|
|
|
2012-04-02 22:10:29 +00:00
|
|
|
/// isFNegFree - Return true if an fneg operation is free to the point where
|
|
|
|
/// it is never worthwhile to replace it with a bitwise operation.
|
|
|
|
virtual bool isFNegFree(EVT) const {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// isFAbsFree - Return true if an fneg operation is free to the point where
|
|
|
|
/// it is never worthwhile to replace it with a bitwise operation.
|
|
|
|
virtual bool isFAbsFree(EVT) const {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2012-06-05 19:07:46 +00:00
|
|
|
/// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than
|
|
|
|
/// a pair of mul and add instructions. fmuladd intrinsics will be expanded to
|
|
|
|
/// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd
|
|
|
|
/// is expanded to mul + add.
|
|
|
|
virtual bool isFMAFasterThanMulAndAdd(EVT) const {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2009-05-28 00:35:15 +00:00
|
|
|
/// isNarrowingProfitable - Return true if it's profitable to narrow
|
|
|
|
/// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow
|
|
|
|
/// from i32 to i8 but not from i32 to i16.
|
2011-10-20 08:57:54 +00:00
|
|
|
virtual bool isNarrowingProfitable(EVT /*VT1*/, EVT /*VT2*/) const {
|
2009-05-29 05:31:40 +00:00
|
|
|
return false;
|
2009-05-28 00:35:15 +00:00
|
|
|
}
|
|
|
|
|
2006-05-16 17:42:15 +00:00
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// Div utility functions
|
|
|
|
//
|
Emit a more efficient magic number multiplication for exact sdivs.
We have to do this in DAGBuilder instead of DAGCombiner, because the exact bit is lost after building.
struct foo { char x[24]; };
long bar(struct foo *a, struct foo *b) { return a-b; }
is now compiled into
movl 4(%esp), %eax
subl 8(%esp), %eax
sarl $3, %eax
imull $-1431655765, %eax, %eax
instead of
movl 4(%esp), %eax
subl 8(%esp), %eax
movl $715827883, %ecx
imull %ecx
movl %edx, %eax
shrl $31, %eax
sarl $2, %edx
addl %eax, %edx
movl %edx, %eax
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@134695 91177308-0d34-0410-b5e6-96231b3b80d8
2011-07-08 10:31:30 +00:00
|
|
|
SDValue BuildExactSDIV(SDValue Op1, SDValue Op2, DebugLoc dl,
|
|
|
|
SelectionDAG &DAG) const;
|
2011-11-07 17:09:05 +00:00
|
|
|
SDValue BuildSDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization,
|
2012-12-10 22:00:20 +00:00
|
|
|
std::vector<SDNode*> *Created) const;
|
2011-11-07 17:09:05 +00:00
|
|
|
SDValue BuildUDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization,
|
2012-12-10 22:00:20 +00:00
|
|
|
std::vector<SDNode*> *Created) const;
|
2006-05-16 17:42:15 +00:00
|
|
|
|
|
|
|
|
2007-01-12 02:11:51 +00:00
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// Runtime Library hooks
|
|
|
|
//
|
|
|
|
|
|
|
|
/// setLibcallName - Rename the default libcall routine name for the specified
|
|
|
|
/// libcall.
|
2007-01-12 22:49:32 +00:00
|
|
|
void setLibcallName(RTLIB::Libcall Call, const char *Name) {
|
2007-01-12 02:11:51 +00:00
|
|
|
LibcallRoutineNames[Call] = Name;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// getLibcallName - Get the libcall routine name for the specified libcall.
|
|
|
|
///
|
|
|
|
const char *getLibcallName(RTLIB::Libcall Call) const {
|
2007-01-12 22:49:32 +00:00
|
|
|
return LibcallRoutineNames[Call];
|
2007-01-12 02:11:51 +00:00
|
|
|
}
|
|
|
|
|
2007-01-31 09:29:11 +00:00
|
|
|
/// setCmpLibcallCC - Override the default CondCode to be used to test the
|
|
|
|
/// result of the comparison libcall against zero.
|
|
|
|
void setCmpLibcallCC(RTLIB::Libcall Call, ISD::CondCode CC) {
|
|
|
|
CmpLibcallCCs[Call] = CC;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// getCmpLibcallCC - Get the CondCode that's to be used to test the result of
|
|
|
|
/// the comparison libcall against zero.
|
|
|
|
ISD::CondCode getCmpLibcallCC(RTLIB::Libcall Call) const {
|
|
|
|
return CmpLibcallCCs[Call];
|
|
|
|
}
|
|
|
|
|
2009-08-14 20:10:52 +00:00
|
|
|
/// setLibcallCallingConv - Set the CallingConv that should be used for the
|
|
|
|
/// specified libcall.
|
|
|
|
void setLibcallCallingConv(RTLIB::Libcall Call, CallingConv::ID CC) {
|
|
|
|
LibcallCallingConvs[Call] = CC;
|
|
|
|
}
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2009-08-14 20:10:52 +00:00
|
|
|
/// getLibcallCallingConv - Get the CallingConv that should be used for the
|
|
|
|
/// specified libcall.
|
|
|
|
CallingConv::ID getLibcallCallingConv(RTLIB::Libcall Call) const {
|
|
|
|
return LibcallCallingConvs[Call];
|
|
|
|
}
|
|
|
|
|
2006-03-16 21:47:42 +00:00
|
|
|
private:
|
2010-04-21 01:34:56 +00:00
|
|
|
const TargetMachine &TM;
|
2012-10-08 16:38:25 +00:00
|
|
|
const DataLayout *TD;
|
2010-04-21 01:34:56 +00:00
|
|
|
const TargetLoweringObjectFile &TLOF;
|
2005-01-19 03:36:03 +00:00
|
|
|
|
2012-10-09 16:06:12 +00:00
|
|
|
/// PointerTy - The type to use for pointers for the default address space,
|
|
|
|
/// usually i32 or i64.
|
2005-01-19 03:36:03 +00:00
|
|
|
///
|
2009-08-11 20:47:22 +00:00
|
|
|
MVT PointerTy;
|
2005-01-19 03:36:03 +00:00
|
|
|
|
2008-06-30 20:36:26 +00:00
|
|
|
/// IsLittleEndian - True if this is a little endian target.
|
|
|
|
///
|
|
|
|
bool IsLittleEndian;
|
|
|
|
|
2007-01-05 23:31:08 +00:00
|
|
|
/// SelectIsExpensive - Tells the code generator not to expand operations
|
|
|
|
/// into sequences that use the select operations if possible.
|
|
|
|
bool SelectIsExpensive;
|
2005-05-12 02:02:21 +00:00
|
|
|
|
2005-10-21 00:02:42 +00:00
|
|
|
/// IntDivIsCheap - Tells the code generator not to expand integer divides by
|
|
|
|
/// constants into a sequence of muls, adds, and shifts. This is a hack until
|
|
|
|
/// a real cost model is in place. If we ever optimize for size, this will be
|
|
|
|
/// set to true unconditionally.
|
|
|
|
bool IntDivIsCheap;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2012-10-04 21:33:40 +00:00
|
|
|
/// BypassSlowDivMap - Tells the code generator to bypass slow divide or
|
|
|
|
/// remainder instructions. For example, BypassSlowDivWidths[32,8] tells the
|
|
|
|
/// code generator to bypass 32-bit integer div/rem with an 8-bit unsigned
|
2012-09-04 18:22:17 +00:00
|
|
|
/// integer div/rem when the operands are positive and less than 256.
|
2012-10-04 21:33:40 +00:00
|
|
|
DenseMap <unsigned int, unsigned int> BypassSlowDivWidths;
|
2012-09-04 18:22:17 +00:00
|
|
|
|
2005-10-21 00:02:42 +00:00
|
|
|
/// Pow2DivIsCheap - Tells the code generator that it shouldn't generate
|
|
|
|
/// srl/add/sra for a signed divide by power of two, and let the target handle
|
|
|
|
/// it.
|
|
|
|
bool Pow2DivIsCheap;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2010-11-30 18:12:52 +00:00
|
|
|
/// JumpIsExpensive - Tells the code generator that it shouldn't generate
|
|
|
|
/// extra flow control instructions and should attempt to combine flow
|
|
|
|
/// control instructions via predication.
|
|
|
|
bool JumpIsExpensive;
|
|
|
|
|
2006-12-10 23:12:42 +00:00
|
|
|
/// UseUnderscoreSetJmp - This target prefers to use _setjmp to implement
|
|
|
|
/// llvm.setjmp. Defaults to false.
|
|
|
|
bool UseUnderscoreSetJmp;
|
|
|
|
|
|
|
|
/// UseUnderscoreLongJmp - This target prefers to use _longjmp to implement
|
|
|
|
/// llvm.longjmp. Defaults to false.
|
|
|
|
bool UseUnderscoreLongJmp;
|
|
|
|
|
2012-07-02 22:39:56 +00:00
|
|
|
/// SupportJumpTables - Whether the target can generate code for jumptables.
|
|
|
|
/// If it's not true, then each jumptable must be lowered into if-then-else's.
|
|
|
|
bool SupportJumpTables;
|
|
|
|
|
2012-09-25 20:35:36 +00:00
|
|
|
/// MinimumJumpTableEntries - Number of blocks threshold to use jump tables.
|
|
|
|
int MinimumJumpTableEntries;
|
|
|
|
|
2008-11-23 15:47:28 +00:00
|
|
|
/// BooleanContents - Information about the contents of the high-bits in
|
|
|
|
/// boolean values held in a type wider than i1. See getBooleanContents.
|
|
|
|
BooleanContent BooleanContents;
|
2011-09-06 19:07:46 +00:00
|
|
|
/// BooleanVectorContents - Information about the contents of the high-bits
|
|
|
|
/// in boolean vector values when the element type is wider than i1. See
|
|
|
|
/// getBooleanContents.
|
|
|
|
BooleanContent BooleanVectorContents;
|
2008-06-30 20:36:26 +00:00
|
|
|
|
|
|
|
/// SchedPreferenceInfo - The target scheduling preference: shortest possible
|
|
|
|
/// total cycles or lowest register usage.
|
2010-05-19 20:19:50 +00:00
|
|
|
Sched::Preference SchedPreferenceInfo;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-09-04 06:21:35 +00:00
|
|
|
/// JumpBufSize - The size, in bytes, of the target's jmp_buf buffers
|
|
|
|
unsigned JumpBufSize;
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-09-04 06:21:35 +00:00
|
|
|
/// JumpBufAlignment - The alignment, in bytes, of the target's jmp_buf
|
|
|
|
/// buffers
|
|
|
|
unsigned JumpBufAlignment;
|
2007-05-16 23:45:53 +00:00
|
|
|
|
2010-07-12 18:11:17 +00:00
|
|
|
/// MinStackArgumentAlignment - The minimum alignment that any argument
|
2010-07-11 04:01:49 +00:00
|
|
|
/// on the stack needs to have.
|
|
|
|
///
|
|
|
|
unsigned MinStackArgumentAlignment;
|
|
|
|
|
2011-05-06 20:34:06 +00:00
|
|
|
/// MinFunctionAlignment - The minimum function alignment (used when
|
|
|
|
/// optimizing for size, and to prevent explicitly provided alignment
|
|
|
|
/// from leading to incorrect code).
|
|
|
|
///
|
|
|
|
unsigned MinFunctionAlignment;
|
|
|
|
|
2011-05-07 19:23:14 +00:00
|
|
|
/// PrefFunctionAlignment - The preferred function alignment (used when
|
2011-05-06 20:34:06 +00:00
|
|
|
/// alignment unspecified and optimizing for speed).
|
|
|
|
///
|
|
|
|
unsigned PrefFunctionAlignment;
|
|
|
|
|
2011-05-07 19:23:14 +00:00
|
|
|
/// PrefLoopAlignment - The preferred loop alignment.
|
2008-02-28 00:43:03 +00:00
|
|
|
///
|
|
|
|
unsigned PrefLoopAlignment;
|
|
|
|
|
2010-06-23 16:07:42 +00:00
|
|
|
/// ShouldFoldAtomicFences - Whether fencing MEMBARRIER instructions should
|
|
|
|
/// be folded into the enclosed atomic intrinsic instruction by the
|
|
|
|
/// combiner.
|
|
|
|
bool ShouldFoldAtomicFences;
|
|
|
|
|
2011-08-03 21:06:02 +00:00
|
|
|
/// InsertFencesForAtomic - Whether the DAG builder should automatically
|
|
|
|
/// insert fences and reduce ordering for atomics. (This will be set for
|
|
|
|
/// for most architectures with weak memory ordering.)
|
|
|
|
bool InsertFencesForAtomic;
|
|
|
|
|
2006-01-13 17:47:52 +00:00
|
|
|
/// StackPointerRegisterToSaveRestore - If set to a physical register, this
|
|
|
|
/// specifies the register that llvm.savestack/llvm.restorestack should save
|
|
|
|
/// and restore.
|
|
|
|
unsigned StackPointerRegisterToSaveRestore;
|
2005-04-07 19:41:18 +00:00
|
|
|
|
2007-02-22 15:37:19 +00:00
|
|
|
/// ExceptionPointerRegister - If set to a physical register, this specifies
|
|
|
|
/// the register that receives the exception address on entry to a landing
|
|
|
|
/// pad.
|
|
|
|
unsigned ExceptionPointerRegister;
|
|
|
|
|
|
|
|
/// ExceptionSelectorRegister - If set to a physical register, this specifies
|
|
|
|
/// the register that receives the exception typeid on entry to a landing
|
|
|
|
/// pad.
|
|
|
|
unsigned ExceptionSelectorRegister;
|
|
|
|
|
2005-01-19 03:36:03 +00:00
|
|
|
/// RegClassForVT - This indicates the default register class to use for
|
|
|
|
/// each ValueType the target supports natively.
|
2012-02-22 05:59:10 +00:00
|
|
|
const TargetRegisterClass *RegClassForVT[MVT::LAST_VALUETYPE];
|
2009-08-11 20:47:22 +00:00
|
|
|
unsigned char NumRegistersForVT[MVT::LAST_VALUETYPE];
|
2012-12-11 11:14:33 +00:00
|
|
|
EVT RegisterTypeForVT[MVT::LAST_VALUETYPE];
|
2005-01-19 03:36:03 +00:00
|
|
|
|
2010-07-19 18:47:01 +00:00
|
|
|
/// RepRegClassForVT - This indicates the "representative" register class to
|
|
|
|
/// use for each ValueType the target supports natively. This information is
|
2010-07-21 06:09:07 +00:00
|
|
|
/// used by the scheduler to track register pressure. By default, the
|
|
|
|
/// representative register class is the largest legal super-reg register
|
|
|
|
/// class of the register class of the specified type. e.g. On x86, i8, i16,
|
2010-07-19 18:47:01 +00:00
|
|
|
/// and i32's representative class would be GR32.
|
|
|
|
const TargetRegisterClass *RepRegClassForVT[MVT::LAST_VALUETYPE];
|
|
|
|
|
2010-07-21 06:09:07 +00:00
|
|
|
/// RepRegClassCostForVT - This indicates the "cost" of the "representative"
|
|
|
|
/// register class for each ValueType. The cost is used by the scheduler to
|
|
|
|
/// approximate register pressure.
|
|
|
|
uint8_t RepRegClassCostForVT[MVT::LAST_VALUETYPE];
|
|
|
|
|
2005-01-19 03:36:03 +00:00
|
|
|
/// TransformToType - For any value types we are promoting or expanding, this
|
|
|
|
/// contains the value type that we are changing to. For Expanded types, this
|
|
|
|
/// contains one step of the expand (e.g. i64 -> i32), even if there are
|
|
|
|
/// multiple steps required (e.g. i64 -> i16). For types natively supported
|
|
|
|
/// by the system, this holds the same type (e.g. i32 -> i32).
|
2012-12-11 11:14:33 +00:00
|
|
|
EVT TransformToType[MVT::LAST_VALUETYPE];
|
2005-01-19 03:36:03 +00:00
|
|
|
|
|
|
|
/// OpActions - For each operation and each value type, keep a LegalizeAction
|
|
|
|
/// that indicates how instruction selection should deal with the operation.
|
|
|
|
/// Most operations are Legal (aka, supported natively by the target), but
|
|
|
|
/// operations that are not should be described. Note that operations on
|
|
|
|
/// non-legal value types are not described here.
|
2010-05-14 07:37:40 +00:00
|
|
|
uint8_t OpActions[MVT::LAST_VALUETYPE][ISD::BUILTIN_OP_END];
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2010-05-13 23:01:26 +00:00
|
|
|
/// LoadExtActions - For each load extension type and each value type,
|
2006-10-04 00:56:09 +00:00
|
|
|
/// keep a LegalizeAction that indicates how instruction selection should deal
|
2010-05-13 23:01:26 +00:00
|
|
|
/// with a load of a specific value type and extension type.
|
|
|
|
uint8_t LoadExtActions[MVT::LAST_VALUETYPE][ISD::LAST_LOADEXT_TYPE];
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2010-05-13 23:01:26 +00:00
|
|
|
/// TruncStoreActions - For each value type pair keep a LegalizeAction that
|
|
|
|
/// indicates whether a truncating store of a specific value type and
|
|
|
|
/// truncating type is legal.
|
|
|
|
uint8_t TruncStoreActions[MVT::LAST_VALUETYPE][MVT::LAST_VALUETYPE];
|
2006-11-09 17:55:04 +00:00
|
|
|
|
2009-06-09 00:19:10 +00:00
|
|
|
/// IndexedModeActions - For each indexed mode and each value type,
|
|
|
|
/// keep a pair of LegalizeAction that indicates how instruction
|
2010-05-13 23:01:26 +00:00
|
|
|
/// selection should deal with the load / store. The first dimension is the
|
|
|
|
/// value_type for the reference. The second dimension represents the various
|
|
|
|
/// modes for load store.
|
|
|
|
uint8_t IndexedModeActions[MVT::LAST_VALUETYPE][ISD::LAST_INDEXED_MODE];
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2008-10-15 02:05:31 +00:00
|
|
|
/// CondCodeActions - For each condition code (ISD::CondCode) keep a
|
|
|
|
/// LegalizeAction that indicates how instruction selection should
|
|
|
|
/// deal with the condition code.
|
2012-09-14 15:36:50 +00:00
|
|
|
/// Because each CC action takes up 2 bits, we need to have the array size
|
|
|
|
/// be large enough to fit all of the value types. This can be done by
|
|
|
|
/// dividing the MVT::LAST_VALUETYPE by 32 and adding one.
|
|
|
|
uint64_t CondCodeActions[ISD::SETCC_INVALID][(MVT::LAST_VALUETYPE / 32) + 1];
|
2008-10-15 02:05:31 +00:00
|
|
|
|
2006-01-29 08:40:37 +00:00
|
|
|
ValueTypeActionImpl ValueTypeActions;
|
2005-04-21 20:59:05 +00:00
|
|
|
|
2012-10-24 23:47:38 +00:00
|
|
|
public:
|
2011-05-27 21:03:13 +00:00
|
|
|
LegalizeKind
|
|
|
|
getTypeConversion(LLVMContext &Context, EVT VT) const {
|
|
|
|
// If this is a simple type, use the ComputeRegisterProp mechanism.
|
|
|
|
if (VT.isSimple()) {
|
2012-12-11 11:14:33 +00:00
|
|
|
assert((unsigned)VT.getSimpleVT().SimpleTy <
|
|
|
|
array_lengthof(TransformToType));
|
|
|
|
EVT NVT = TransformToType[VT.getSimpleVT().SimpleTy];
|
|
|
|
LegalizeTypeAction LA = ValueTypeActions.getTypeAction(VT.getSimpleVT());
|
2011-05-28 17:57:14 +00:00
|
|
|
|
2011-06-01 12:51:46 +00:00
|
|
|
assert(
|
2012-12-11 11:14:33 +00:00
|
|
|
(!(NVT.isSimple() && LA != TypeLegal) ||
|
|
|
|
ValueTypeActions.getTypeAction(NVT.getSimpleVT()) != TypePromoteInteger)
|
2011-06-01 12:51:46 +00:00
|
|
|
&& "Promote may not follow Expand or Promote");
|
2011-05-28 17:57:14 +00:00
|
|
|
|
2012-10-27 04:11:32 +00:00
|
|
|
if (LA == TypeSplitVector)
|
2012-12-11 11:14:33 +00:00
|
|
|
NVT = EVT::getVectorVT(Context, VT.getVectorElementType(),
|
|
|
|
VT.getVectorNumElements() / 2);
|
2011-05-27 21:03:13 +00:00
|
|
|
return LegalizeKind(LA, NVT);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Handle Extended Scalar Types.
|
|
|
|
if (!VT.isVector()) {
|
|
|
|
assert(VT.isInteger() && "Float types must be simple");
|
|
|
|
unsigned BitSize = VT.getSizeInBits();
|
|
|
|
// First promote to a power-of-two size, then expand if necessary.
|
|
|
|
if (BitSize < 8 || !isPowerOf2_32(BitSize)) {
|
|
|
|
EVT NVT = VT.getRoundIntegerType(Context);
|
|
|
|
assert(NVT != VT && "Unable to round integer VT");
|
|
|
|
LegalizeKind NextStep = getTypeConversion(Context, NVT);
|
|
|
|
// Avoid multi-step promotion.
|
2011-05-28 17:57:14 +00:00
|
|
|
if (NextStep.first == TypePromoteInteger) return NextStep;
|
2011-05-27 21:03:13 +00:00
|
|
|
// Return rounded integer type.
|
2011-05-28 17:57:14 +00:00
|
|
|
return LegalizeKind(TypePromoteInteger, NVT);
|
2011-05-27 21:03:13 +00:00
|
|
|
}
|
|
|
|
|
2011-05-28 17:57:14 +00:00
|
|
|
return LegalizeKind(TypeExpandInteger,
|
2011-05-27 21:03:13 +00:00
|
|
|
EVT::getIntegerVT(Context, VT.getSizeInBits()/2));
|
|
|
|
}
|
|
|
|
|
|
|
|
// Handle vector types.
|
|
|
|
unsigned NumElts = VT.getVectorNumElements();
|
|
|
|
EVT EltVT = VT.getVectorElementType();
|
|
|
|
|
|
|
|
// Vectors with only one element are always scalarized.
|
|
|
|
if (NumElts == 1)
|
2011-05-28 17:57:14 +00:00
|
|
|
return LegalizeKind(TypeScalarizeVector, EltVT);
|
2011-05-27 21:03:13 +00:00
|
|
|
|
2012-06-04 11:27:21 +00:00
|
|
|
// Try to widen vector elements until a legal type is found.
|
|
|
|
if (EltVT.isInteger()) {
|
2011-06-01 12:51:46 +00:00
|
|
|
// Vectors with a number of elements that is not a power of two are always
|
|
|
|
// widened, for example <3 x float> -> <4 x float>.
|
|
|
|
if (!VT.isPow2VectorType()) {
|
|
|
|
NumElts = (unsigned)NextPowerOf2(NumElts);
|
|
|
|
EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts);
|
|
|
|
return LegalizeKind(TypeWidenVector, NVT);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Examine the element type.
|
|
|
|
LegalizeKind LK = getTypeConversion(Context, EltVT);
|
|
|
|
|
|
|
|
// If type is to be expanded, split the vector.
|
|
|
|
// <4 x i140> -> <2 x i140>
|
|
|
|
if (LK.first == TypeExpandInteger)
|
|
|
|
return LegalizeKind(TypeSplitVector,
|
|
|
|
EVT::getVectorVT(Context, EltVT, NumElts / 2));
|
|
|
|
|
|
|
|
// Promote the integer element types until a legal vector type is found
|
|
|
|
// or until the element integer type is too big. If a legal type was not
|
|
|
|
// found, fallback to the usual mechanism of widening/splitting the
|
|
|
|
// vector.
|
|
|
|
while (1) {
|
|
|
|
// Increase the bitwidth of the element to the next pow-of-two
|
|
|
|
// (which is greater than 8 bits).
|
|
|
|
EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()
|
|
|
|
).getRoundIntegerType(Context);
|
|
|
|
|
|
|
|
// Stop trying when getting a non-simple element type.
|
|
|
|
// Note that vector elements may be greater than legal vector element
|
|
|
|
// types. Example: X86 XMM registers hold 64bit element on 32bit systems.
|
|
|
|
if (!EltVT.isSimple()) break;
|
|
|
|
|
|
|
|
// Build a new vector type and check if it is legal.
|
|
|
|
MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
|
|
|
|
// Found a legal promoted vector type.
|
2011-06-17 20:54:12 +00:00
|
|
|
if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal)
|
2011-06-01 12:51:46 +00:00
|
|
|
return LegalizeKind(TypePromoteInteger,
|
|
|
|
EVT::getVectorVT(Context, EltVT, NumElts));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2011-05-27 21:03:13 +00:00
|
|
|
// Try to widen the vector until a legal type is found.
|
|
|
|
// If there is no wider legal type, split the vector.
|
|
|
|
while (1) {
|
|
|
|
// Round up to the next power of 2.
|
|
|
|
NumElts = (unsigned)NextPowerOf2(NumElts);
|
|
|
|
|
|
|
|
// If there is no simple vector type with this many elements then there
|
|
|
|
// cannot be a larger legal vector type. Note that this assumes that
|
|
|
|
// there are no skipped intermediate vector types in the simple types.
|
2011-06-12 14:40:30 +00:00
|
|
|
if (!EltVT.isSimple()) break;
|
2011-05-27 21:03:13 +00:00
|
|
|
MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts);
|
|
|
|
if (LargerVector == MVT()) break;
|
|
|
|
|
|
|
|
// If this type is legal then widen the vector.
|
2011-05-28 17:57:14 +00:00
|
|
|
if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal)
|
|
|
|
return LegalizeKind(TypeWidenVector, LargerVector);
|
2011-05-27 21:03:13 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Widen odd vectors to next power of two.
|
|
|
|
if (!VT.isPow2VectorType()) {
|
|
|
|
EVT NVT = VT.getPow2VectorType(Context);
|
2011-05-28 17:57:14 +00:00
|
|
|
return LegalizeKind(TypeWidenVector, NVT);
|
2011-05-27 21:03:13 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Vectors with illegal element types are expanded.
|
|
|
|
EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2);
|
2011-05-28 17:57:14 +00:00
|
|
|
return LegalizeKind(TypeSplitVector, NVT);
|
2011-05-27 21:03:13 +00:00
|
|
|
}
|
|
|
|
|
2012-10-24 23:47:38 +00:00
|
|
|
private:
|
2012-12-11 11:14:33 +00:00
|
|
|
std::vector<std::pair<EVT, const TargetRegisterClass*> > AvailableRegClasses;
|
2005-07-19 04:52:44 +00:00
|
|
|
|
2006-03-01 04:52:55 +00:00
|
|
|
/// TargetDAGCombineArray - Targets can specify ISD nodes that they would
|
|
|
|
/// like PerformDAGCombine callbacks for by calling setTargetDAGCombine(),
|
|
|
|
/// which sets a bit in this array.
|
2008-05-09 21:50:23 +00:00
|
|
|
unsigned char
|
2009-01-05 19:40:39 +00:00
|
|
|
TargetDAGCombineArray[(ISD::BUILTIN_OP_END+CHAR_BIT-1)/CHAR_BIT];
|
2010-06-23 15:49:18 +00:00
|
|
|
|
2006-04-04 00:25:10 +00:00
|
|
|
/// PromoteToType - For operations that must be promoted to a specific type,
|
|
|
|
/// this holds the destination type. This map should be sparse, so don't hold
|
|
|
|
/// it as an array.
|
|
|
|
///
|
|
|
|
/// Targets add entries to this map with AddPromotedToType(..), clients access
|
|
|
|
/// this with getTypeToPromoteTo(..).
|
2009-08-11 20:47:22 +00:00
|
|
|
std::map<std::pair<unsigned, MVT::SimpleValueType>, MVT::SimpleValueType>
|
2008-06-08 20:54:56 +00:00
|
|
|
PromoteToType;
|
2007-01-12 02:11:51 +00:00
|
|
|
|
|
|
|
/// LibcallRoutineNames - Stores the name each libcall.
|
|
|
|
///
|
2007-01-12 22:49:32 +00:00
|
|
|
const char *LibcallRoutineNames[RTLIB::UNKNOWN_LIBCALL];
|
2007-01-12 02:11:51 +00:00
|
|
|
|
2007-01-31 09:29:11 +00:00
|
|
|
/// CmpLibcallCCs - The ISD::CondCode that should be used to test the result
|
|
|
|
/// of each of the comparison libcall against zero.
|
|
|
|
ISD::CondCode CmpLibcallCCs[RTLIB::UNKNOWN_LIBCALL];
|
|
|
|
|
2009-08-14 20:10:52 +00:00
|
|
|
/// LibcallCallingConvs - Stores the CallingConv that should be used for each
|
|
|
|
/// libcall.
|
|
|
|
CallingConv::ID LibcallCallingConvs[RTLIB::UNKNOWN_LIBCALL];
|
|
|
|
|
2005-07-19 04:52:44 +00:00
|
|
|
protected:
|
2009-03-03 02:55:14 +00:00
|
|
|
/// When lowering \@llvm.memset this field specifies the maximum number of
|
2005-07-19 04:52:44 +00:00
|
|
|
/// store operations that may be substituted for the call to memset. Targets
|
|
|
|
/// must set this value based on the cost threshold for that target. Targets
|
|
|
|
/// should assume that the memset will be done using as many of the largest
|
|
|
|
/// store operations first, followed by smaller ones, if necessary, per
|
|
|
|
/// alignment restrictions. For example, storing 9 bytes on a 32-bit machine
|
2005-07-27 05:53:44 +00:00
|
|
|
/// with 16-bit alignment would result in four 2-byte stores and one 1-byte
|
2005-07-19 04:52:44 +00:00
|
|
|
/// store. This only applies to setting a constant array of a constant size.
|
|
|
|
/// @brief Specify maximum number of store instructions per memset call.
|
2006-02-14 08:38:30 +00:00
|
|
|
unsigned maxStoresPerMemset;
|
2005-07-19 04:52:44 +00:00
|
|
|
|
2011-01-06 06:52:41 +00:00
|
|
|
/// Maximum number of stores operations that may be substituted for the call
|
|
|
|
/// to memset, used for functions with OptSize attribute.
|
|
|
|
unsigned maxStoresPerMemsetOptSize;
|
|
|
|
|
2009-03-03 02:55:14 +00:00
|
|
|
/// When lowering \@llvm.memcpy this field specifies the maximum number of
|
2005-07-19 04:52:44 +00:00
|
|
|
/// store operations that may be substituted for a call to memcpy. Targets
|
|
|
|
/// must set this value based on the cost threshold for that target. Targets
|
|
|
|
/// should assume that the memcpy will be done using as many of the largest
|
|
|
|
/// store operations first, followed by smaller ones, if necessary, per
|
|
|
|
/// alignment restrictions. For example, storing 7 bytes on a 32-bit machine
|
|
|
|
/// with 32-bit alignment would result in one 4-byte store, a one 2-byte store
|
|
|
|
/// and one 1-byte store. This only applies to copying a constant array of
|
|
|
|
/// constant size.
|
|
|
|
/// @brief Specify maximum bytes of store instructions per memcpy call.
|
2006-02-14 08:38:30 +00:00
|
|
|
unsigned maxStoresPerMemcpy;
|
2005-07-19 04:52:44 +00:00
|
|
|
|
2011-01-06 06:52:41 +00:00
|
|
|
/// Maximum number of store operations that may be substituted for a call
|
|
|
|
/// to memcpy, used for functions with OptSize attribute.
|
|
|
|
unsigned maxStoresPerMemcpyOptSize;
|
|
|
|
|
2009-03-03 02:55:14 +00:00
|
|
|
/// When lowering \@llvm.memmove this field specifies the maximum number of
|
2005-07-19 04:52:44 +00:00
|
|
|
/// store instructions that may be substituted for a call to memmove. Targets
|
|
|
|
/// must set this value based on the cost threshold for that target. Targets
|
|
|
|
/// should assume that the memmove will be done using as many of the largest
|
|
|
|
/// store operations first, followed by smaller ones, if necessary, per
|
2005-07-27 05:53:44 +00:00
|
|
|
/// alignment restrictions. For example, moving 9 bytes on a 32-bit machine
|
|
|
|
/// with 8-bit alignment would result in nine 1-byte stores. This only
|
2005-07-19 04:52:44 +00:00
|
|
|
/// applies to copying a constant array of constant size.
|
|
|
|
/// @brief Specify maximum bytes of store instructions per memmove call.
|
2006-02-14 08:38:30 +00:00
|
|
|
unsigned maxStoresPerMemmove;
|
2005-07-19 04:52:44 +00:00
|
|
|
|
2011-01-06 06:52:41 +00:00
|
|
|
/// Maximum number of store instructions that may be substituted for a call
|
|
|
|
/// to memmove, used for functions with OpSize attribute.
|
|
|
|
unsigned maxStoresPerMemmoveOptSize;
|
|
|
|
|
2009-05-13 21:42:09 +00:00
|
|
|
/// This field specifies whether the target can benefit from code placement
|
|
|
|
/// optimization.
|
|
|
|
bool benefitFromCodePlacementOpt;
|
2010-07-19 18:47:01 +00:00
|
|
|
|
2012-05-05 12:49:14 +00:00
|
|
|
/// predictableSelectIsExpensive - Tells the code generator that select is
|
|
|
|
/// more expensive than a branch if the branch is usually predicted right.
|
|
|
|
bool predictableSelectIsExpensive;
|
|
|
|
|
2010-07-19 18:47:01 +00:00
|
|
|
private:
|
|
|
|
/// isLegalRC - Return true if the value types that can be represented by the
|
|
|
|
/// specified register class are all legal.
|
|
|
|
bool isLegalRC(const TargetRegisterClass *RC) const;
|
2005-01-07 07:44:22 +00:00
|
|
|
};
|
2010-07-10 09:00:22 +00:00
|
|
|
|
|
|
|
/// GetReturnInfo - Given an LLVM IR type and return type attributes,
|
|
|
|
/// compute the return value EVTs and flags, and optionally also
|
|
|
|
/// the offsets, if the return value is being lowered to memory.
|
2011-07-18 04:54:35 +00:00
|
|
|
void GetReturnInfo(Type* ReturnType, Attributes attr,
|
2010-07-10 09:00:22 +00:00
|
|
|
SmallVectorImpl<ISD::OutputArg> &Outs,
|
2012-05-25 00:09:29 +00:00
|
|
|
const TargetLowering &TLI);
|
2010-07-10 09:00:22 +00:00
|
|
|
|
2005-01-07 07:44:22 +00:00
|
|
|
} // end llvm namespace
|
|
|
|
|
|
|
|
#endif
|