2005-01-07 08:44:22 +01:00
|
|
|
//===-- llvm/Target/TargetLowering.h - Target Lowering Info -----*- C++ -*-===//
|
2005-04-21 22:59:05 +02:00
|
|
|
//
|
2005-01-07 08:44:22 +01:00
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
2007-12-29 20:59:42 +01:00
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
2005-04-21 22:59:05 +02:00
|
|
|
//
|
2005-01-07 08:44:22 +01:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// 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 06:52:44 +02:00
|
|
|
// 3. Cost thresholds for alternative implementations of certain operations.
|
2005-01-07 08:44:22 +01:00
|
|
|
//
|
|
|
|
// In addition it has a few other components, like information about FP
|
|
|
|
// immediates.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#ifndef LLVM_TARGET_TARGETLOWERING_H
|
|
|
|
#define LLVM_TARGET_TARGETLOWERING_H
|
|
|
|
|
2008-02-26 03:33:44 +01:00
|
|
|
#include "llvm/InlineAsm.h"
|
2006-02-16 22:11:51 +01:00
|
|
|
#include "llvm/CodeGen/SelectionDAGNodes.h"
|
2007-01-12 23:49:32 +01:00
|
|
|
#include "llvm/CodeGen/RuntimeLibcalls.h"
|
2007-08-30 02:23:21 +02:00
|
|
|
#include "llvm/ADT/APFloat.h"
|
2008-09-04 01:12:08 +02:00
|
|
|
#include "llvm/ADT/DenseMap.h"
|
2008-10-15 01:54:11 +02:00
|
|
|
#include "llvm/ADT/SmallSet.h"
|
2009-04-27 20:41:29 +02:00
|
|
|
#include "llvm/ADT/SmallVector.h"
|
2007-09-07 06:06:50 +02:00
|
|
|
#include "llvm/ADT/STLExtras.h"
|
2009-06-20 00:08:58 +02:00
|
|
|
#include "llvm/Support/DebugLoc.h"
|
2009-02-27 14:37:18 +01:00
|
|
|
#include "llvm/Target/TargetMachine.h"
|
2009-01-05 20:47:30 +01:00
|
|
|
#include <climits>
|
2006-04-04 02:25:10 +02:00
|
|
|
#include <map>
|
2007-01-13 00:21:42 +01:00
|
|
|
#include <vector>
|
2005-01-07 08:44:22 +01:00
|
|
|
|
|
|
|
namespace llvm {
|
2008-09-10 22:11:02 +02:00
|
|
|
class AllocaInst;
|
2009-01-05 18:59:02 +01:00
|
|
|
class CallInst;
|
2005-01-07 08:44:22 +01:00
|
|
|
class Function;
|
2008-08-19 23:26:29 +02:00
|
|
|
class FastISel;
|
2008-05-12 22:08:05 +02:00
|
|
|
class MachineBasicBlock;
|
2008-08-19 23:26:29 +02:00
|
|
|
class MachineFunction;
|
2008-05-12 22:08:05 +02:00
|
|
|
class MachineFrameInfo;
|
|
|
|
class MachineInstr;
|
2008-09-23 23:53:34 +02:00
|
|
|
class MachineModuleInfo;
|
2009-01-13 01:35:13 +01:00
|
|
|
class DwarfWriter;
|
2005-01-07 08:44:22 +01:00
|
|
|
class SDNode;
|
2008-07-27 23:46:04 +02:00
|
|
|
class SDValue;
|
2005-01-07 08:44:22 +01:00
|
|
|
class SelectionDAG;
|
2008-05-12 22:08:05 +02:00
|
|
|
class TargetData;
|
|
|
|
class TargetMachine;
|
|
|
|
class TargetRegisterClass;
|
2007-11-06 00:12:20 +01:00
|
|
|
class TargetSubtarget;
|
2008-05-12 22:08:05 +02:00
|
|
|
class Value;
|
2005-01-07 08:44:22 +01:00
|
|
|
|
2009-02-27 14:37:18 +01:00
|
|
|
// FIXME: should this be here?
|
|
|
|
namespace TLSModel {
|
|
|
|
enum Model {
|
|
|
|
GeneralDynamic,
|
|
|
|
LocalDynamic,
|
|
|
|
InitialExec,
|
|
|
|
LocalExec
|
|
|
|
};
|
|
|
|
}
|
|
|
|
TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc);
|
|
|
|
|
|
|
|
|
2005-01-07 08:44:22 +01:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
/// 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 {
|
|
|
|
public:
|
2005-01-16 08:27:49 +01:00
|
|
|
/// 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 17:23:43 +01:00
|
|
|
Custom // Use the LowerOperation hook to implement custom lowering.
|
2005-01-16 08:27:49 +01:00
|
|
|
};
|
|
|
|
|
2005-01-19 04:36:03 +01:00
|
|
|
enum OutOfRangeShiftAmount {
|
|
|
|
Undefined, // Oversized shift amounts are undefined (default).
|
|
|
|
Mask, // Shift amounts are auto masked (anded) to value size.
|
2006-02-22 17:23:43 +01:00
|
|
|
Extend // Oversized shift pulls in zeros or sign bits.
|
2005-01-19 04:36:03 +01:00
|
|
|
};
|
|
|
|
|
2008-11-23 16:47:28 +01:00
|
|
|
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 21:41:18 +02:00
|
|
|
};
|
|
|
|
|
2006-01-25 10:09:02 +01:00
|
|
|
enum SchedPreference {
|
|
|
|
SchedulingForLatency, // Scheduling for shortest total latency.
|
2006-02-22 17:23:43 +01:00
|
|
|
SchedulingForRegPressure // Scheduling for lowest register pressure.
|
2006-01-25 10:09:02 +01:00
|
|
|
};
|
|
|
|
|
2007-08-01 17:32:29 +02:00
|
|
|
explicit TargetLowering(TargetMachine &TM);
|
2005-01-16 08:27:49 +01:00
|
|
|
virtual ~TargetLowering();
|
2005-01-07 08:44:22 +01:00
|
|
|
|
|
|
|
TargetMachine &getTargetMachine() const { return TM; }
|
2006-05-03 03:29:57 +02:00
|
|
|
const TargetData *getTargetData() const { return TD; }
|
2005-04-21 22:59:05 +02:00
|
|
|
|
2008-02-11 11:37:04 +01:00
|
|
|
bool isBigEndian() const { return !IsLittleEndian; }
|
2005-01-07 08:44:22 +01:00
|
|
|
bool isLittleEndian() const { return IsLittleEndian; }
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT getPointerTy() const { return PointerTy; }
|
|
|
|
MVT getShiftAmountTy() const { return ShiftAmountTy; }
|
2005-01-19 04:36:03 +01:00
|
|
|
OutOfRangeShiftAmount getShiftAmountFlavor() const {return ShiftAmtHandling; }
|
2005-01-17 00:59:30 +01:00
|
|
|
|
2006-10-07 00:46:34 +02:00
|
|
|
/// usesGlobalOffsetTable - Return true if this target uses a GOT for PIC
|
|
|
|
/// codegen.
|
|
|
|
bool usesGlobalOffsetTable() const { return UsesGlobalOffsetTable; }
|
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
llvm-svn: 43924
2007-11-09 02:32:10 +01:00
|
|
|
|
2007-01-06 00:31:08 +01:00
|
|
|
/// isSelectExpensive - Return true if the select operation is expensive for
|
2005-05-12 04:02:21 +02:00
|
|
|
/// this target.
|
2007-01-06 00:31:08 +01:00
|
|
|
bool isSelectExpensive() const { return SelectIsExpensive; }
|
2005-10-20 04:14:14 +02:00
|
|
|
|
2005-10-21 02:02:42 +02:00
|
|
|
/// isIntDivCheap() - Return true if integer divide is usually cheaper than
|
2005-10-20 04:14:14 +02:00
|
|
|
/// a sequence of several shifts, adds, and multiplies for this target.
|
2005-10-21 02:02:42 +02:00
|
|
|
bool isIntDivCheap() const { return IntDivIsCheap; }
|
2005-05-12 04:02:21 +02:00
|
|
|
|
2005-10-21 02:02:42 +02:00
|
|
|
/// 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
llvm-svn: 43375
2007-10-26 03:56:11 +02:00
|
|
|
|
2009-01-05 22:33:06 +01:00
|
|
|
/// 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
|
|
|
|
/// BRCOND the argument passed is MVT::Other since there are no other
|
|
|
|
/// operands to get a type hint from.
|
2009-01-01 16:52:00 +01:00
|
|
|
virtual MVT getSetCCResultType(MVT VT) const;
|
2005-04-07 21:41:18 +02:00
|
|
|
|
2008-11-23 16:47:28 +01:00
|
|
|
/// 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.
|
|
|
|
BooleanContent getBooleanContents() const { return BooleanContents;}
|
2005-04-07 21:41:18 +02:00
|
|
|
|
2006-01-25 10:09:02 +01:00
|
|
|
/// getSchedulingPreference - Return target scheduling preference.
|
|
|
|
SchedPreference getSchedulingPreference() const {
|
|
|
|
return SchedPreferenceInfo;
|
|
|
|
}
|
|
|
|
|
2005-01-17 00:59:30 +01:00
|
|
|
/// getRegClassFor - Return the register class that should be used for the
|
|
|
|
/// specified value type. This may only be called on legal types.
|
2008-06-06 14:08:01 +02:00
|
|
|
TargetRegisterClass *getRegClassFor(MVT VT) const {
|
|
|
|
assert((unsigned)VT.getSimpleVT() < array_lengthof(RegClassForVT));
|
|
|
|
TargetRegisterClass *RC = RegClassForVT[VT.getSimpleVT()];
|
2005-01-07 08:44:22 +01:00
|
|
|
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.
llvm-svn: 52254
2008-06-13 21:07:40 +02:00
|
|
|
|
2005-08-24 18:34:59 +02:00
|
|
|
/// isTypeLegal - Return true if the target has native support for the
|
2005-01-16 02:10:58 +01:00
|
|
|
/// specified value type. This means that it has a register that directly
|
|
|
|
/// holds it without promotions or expansions.
|
2008-06-06 14:08:01 +02:00
|
|
|
bool isTypeLegal(MVT VT) const {
|
|
|
|
assert(!VT.isSimple() ||
|
|
|
|
(unsigned)VT.getSimpleVT() < array_lengthof(RegClassForVT));
|
|
|
|
return VT.isSimple() && RegClassForVT[VT.getSimpleVT()] != 0;
|
2005-01-07 08:44:22 +01:00
|
|
|
}
|
2005-01-16 02:10:58 +01:00
|
|
|
|
2006-01-29 09:40:37 +01:00
|
|
|
class ValueTypeActionImpl {
|
|
|
|
/// ValueTypeActions - This is a bitvector that contains two bits for each
|
|
|
|
/// value type, where the two bits correspond to the LegalizeAction enum.
|
|
|
|
/// This can be queried with "getTypeAction(VT)".
|
2009-06-25 00:22:02 +02:00
|
|
|
/// dimension by (MVT::MAX_ALLOWED_VALUETYPE/32) * 2
|
|
|
|
uint32_t ValueTypeActions[(MVT::MAX_ALLOWED_VALUETYPE/32)*2];
|
2006-01-29 09:40:37 +01:00
|
|
|
public:
|
|
|
|
ValueTypeActionImpl() {
|
|
|
|
ValueTypeActions[0] = ValueTypeActions[1] = 0;
|
2009-06-24 21:41:55 +02:00
|
|
|
ValueTypeActions[2] = ValueTypeActions[3] = 0;
|
2006-01-29 09:40:37 +01:00
|
|
|
}
|
|
|
|
ValueTypeActionImpl(const ValueTypeActionImpl &RHS) {
|
|
|
|
ValueTypeActions[0] = RHS.ValueTypeActions[0];
|
|
|
|
ValueTypeActions[1] = RHS.ValueTypeActions[1];
|
2009-06-24 21:41:55 +02:00
|
|
|
ValueTypeActions[2] = RHS.ValueTypeActions[2];
|
|
|
|
ValueTypeActions[3] = RHS.ValueTypeActions[3];
|
2006-01-29 09:40:37 +01:00
|
|
|
}
|
|
|
|
|
2008-06-06 14:08:01 +02:00
|
|
|
LegalizeAction getTypeAction(MVT VT) const {
|
|
|
|
if (VT.isExtended()) {
|
2008-10-30 09:01:45 +01:00
|
|
|
if (VT.isVector()) {
|
2008-12-18 21:03:17 +01:00
|
|
|
return VT.isPow2VectorType() ? Expand : Promote;
|
2008-10-30 09:01:45 +01:00
|
|
|
}
|
2008-06-06 14:08:01 +02:00
|
|
|
if (VT.isInteger())
|
2007-10-16 11:56:48 +02:00
|
|
|
// First promote to a power-of-two size, then expand if necessary.
|
2008-06-06 14:08:01 +02:00
|
|
|
return VT == VT.getRoundIntegerType() ? Expand : Promote;
|
2007-10-16 11:56:48 +02:00
|
|
|
assert(0 && "Unsupported extended type!");
|
2008-06-06 14:08:01 +02:00
|
|
|
return Legal;
|
2007-10-16 11:56:48 +02:00
|
|
|
}
|
2008-06-06 14:08:01 +02:00
|
|
|
unsigned I = VT.getSimpleVT();
|
|
|
|
assert(I<4*array_lengthof(ValueTypeActions)*sizeof(ValueTypeActions[0]));
|
|
|
|
return (LegalizeAction)((ValueTypeActions[I>>4] >> ((2*I) & 31)) & 3);
|
2006-01-29 09:40:37 +01:00
|
|
|
}
|
2008-06-06 14:08:01 +02:00
|
|
|
void setTypeAction(MVT VT, LegalizeAction Action) {
|
|
|
|
unsigned I = VT.getSimpleVT();
|
|
|
|
assert(I<4*array_lengthof(ValueTypeActions)*sizeof(ValueTypeActions[0]));
|
|
|
|
ValueTypeActions[I>>4] |= Action << ((I*2) & 31);
|
2006-01-29 09:40:37 +01:00
|
|
|
}
|
|
|
|
};
|
|
|
|
|
|
|
|
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.
llvm-svn: 52254
2008-06-13 21:07:40 +02:00
|
|
|
|
2005-01-16 02:10:58 +01:00
|
|
|
/// getTypeAction - Return how we should legalize values of this type, either
|
2005-01-16 08:27:49 +01:00
|
|
|
/// 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.
|
2008-06-06 14:08:01 +02:00
|
|
|
LegalizeAction getTypeAction(MVT VT) const {
|
2006-01-29 09:40:37 +01:00
|
|
|
return ValueTypeActions.getTypeAction(VT);
|
2005-01-16 02:10:58 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/// 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 07:12:35 +01:00
|
|
|
/// 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.
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT getTypeToTransformTo(MVT VT) const {
|
|
|
|
if (VT.isSimple()) {
|
|
|
|
assert((unsigned)VT.getSimpleVT() < array_lengthof(TransformToType));
|
|
|
|
MVT NVT = TransformToType[VT.getSimpleVT()];
|
2007-10-16 11:56:48 +02:00
|
|
|
assert(getTypeAction(NVT) != Promote &&
|
|
|
|
"Promote may not follow Expand or Promote");
|
|
|
|
return NVT;
|
|
|
|
}
|
|
|
|
|
2008-10-20 18:24:25 +02:00
|
|
|
if (VT.isVector()) {
|
2008-12-18 21:03:17 +01:00
|
|
|
MVT NVT = VT.getPow2VectorType();
|
|
|
|
if (NVT == VT) {
|
|
|
|
// Vector length is a power of 2 - split to half the size.
|
|
|
|
unsigned NumElts = VT.getVectorNumElements();
|
|
|
|
MVT EltVT = VT.getVectorElementType();
|
|
|
|
return (NumElts == 1) ? EltVT : MVT::getVectorVT(EltVT, NumElts / 2);
|
|
|
|
}
|
|
|
|
// Promote to a power of two size, avoiding multi-step promotion.
|
|
|
|
return getTypeAction(NVT) == Promote ? getTypeToTransformTo(NVT) : NVT;
|
2008-10-20 18:24:25 +02:00
|
|
|
} else if (VT.isInteger()) {
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT NVT = VT.getRoundIntegerType();
|
2007-10-16 11:56:48 +02:00
|
|
|
if (NVT == VT)
|
|
|
|
// Size is a power of two - expand to half the size.
|
2008-06-06 14:08:01 +02:00
|
|
|
return MVT::getIntegerVT(VT.getSizeInBits() / 2);
|
2007-10-16 11:56:48 +02:00
|
|
|
else
|
|
|
|
// Promote to a power of two size, avoiding multi-step promotion.
|
|
|
|
return getTypeAction(NVT) == Promote ? getTypeToTransformTo(NVT) : NVT;
|
|
|
|
}
|
|
|
|
assert(0 && "Unsupported extended type!");
|
2009-07-03 18:00:23 +02:00
|
|
|
return MVT(MVT::Other); // Not reached
|
2005-01-16 02:10:58 +01:00
|
|
|
}
|
2007-10-16 11:56:48 +02:00
|
|
|
|
2006-12-13 21:52:00 +01:00
|
|
|
/// 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.
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT getTypeToExpandTo(MVT VT) const {
|
|
|
|
assert(!VT.isVector());
|
2006-12-13 21:52:00 +01:00
|
|
|
while (true) {
|
|
|
|
switch (getTypeAction(VT)) {
|
|
|
|
case Legal:
|
|
|
|
return VT;
|
|
|
|
case Expand:
|
2007-06-25 18:23:39 +02:00
|
|
|
VT = getTypeToTransformTo(VT);
|
2006-12-13 21:52:00 +01:00
|
|
|
break;
|
|
|
|
default:
|
|
|
|
assert(false && "Type is not legal nor is it to be expanded!");
|
|
|
|
return VT;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return VT;
|
|
|
|
}
|
|
|
|
|
2007-02-15 04:39:18 +01:00
|
|
|
/// getVectorTypeBreakdown - Vector types are broken down into some number of
|
2007-06-25 18:23:39 +02:00
|
|
|
/// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32
|
2006-03-31 02:28:23 +02:00
|
|
|
/// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack.
|
2007-06-25 18:23:39 +02:00
|
|
|
/// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86.
|
2006-03-31 02:28:23 +02:00
|
|
|
///
|
2006-03-31 02:46:36 +02:00
|
|
|
/// This method returns the number of registers needed, and the VT for each
|
2007-06-29 01:29:44 +02:00
|
|
|
/// register. It also returns the VT and quantity of the intermediate values
|
|
|
|
/// before they are promoted/expanded.
|
2006-03-31 02:28:23 +02:00
|
|
|
///
|
2008-06-06 14:08:01 +02:00
|
|
|
unsigned getVectorTypeBreakdown(MVT VT,
|
|
|
|
MVT &IntermediateVT,
|
2007-06-29 01:29:44 +02:00
|
|
|
unsigned &NumIntermediates,
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT &RegisterVT) const;
|
2008-11-01 21:24:53 +01:00
|
|
|
|
|
|
|
/// 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.
|
|
|
|
typedef struct IntrinsicInfo {
|
|
|
|
unsigned opc; // target opcode
|
|
|
|
MVT memVT; // memory VT
|
|
|
|
const Value* ptrVal; // value representing memory location
|
|
|
|
int offset; // offset off of ptrVal
|
|
|
|
unsigned align; // alignment
|
|
|
|
bool vol; // is volatile?
|
|
|
|
bool readMem; // reads memory?
|
|
|
|
bool writeMem; // writes memory?
|
|
|
|
} IntrinisicInfo;
|
|
|
|
|
|
|
|
virtual bool getTgtMemIntrinsic(IntrinsicInfo& Info,
|
|
|
|
CallInst &I, unsigned Intrinsic) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2008-10-30 09:01:45 +01:00
|
|
|
/// getWidenVectorType: given a vector type, returns the type to widen to
|
|
|
|
/// (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
|
|
|
|
/// If there is no vector type that we want to widen to, returns MVT::Other
|
|
|
|
/// When and were to widen is target dependent based on the cost of
|
|
|
|
/// scalarizing vs using the wider vector type.
|
2009-01-15 18:34:08 +01:00
|
|
|
virtual MVT getWidenVectorType(MVT VT) const;
|
2008-10-30 09:01:45 +01:00
|
|
|
|
2007-08-30 02:23:21 +02:00
|
|
|
typedef std::vector<APFloat>::const_iterator legal_fpimm_iterator;
|
2005-01-07 08:44:22 +01:00
|
|
|
legal_fpimm_iterator legal_fpimm_begin() const {
|
|
|
|
return LegalFPImmediates.begin();
|
|
|
|
}
|
|
|
|
legal_fpimm_iterator legal_fpimm_end() const {
|
|
|
|
return LegalFPImmediates.end();
|
|
|
|
}
|
2006-04-04 19:21:22 +02:00
|
|
|
|
|
|
|
/// 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.
|
2009-04-29 07:20:52 +02:00
|
|
|
virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &Mask,
|
|
|
|
MVT VT) const {
|
2006-04-04 19:21:22 +02:00
|
|
|
return true;
|
|
|
|
}
|
2005-01-16 08:27:49 +01:00
|
|
|
|
2006-04-20 10:54:13 +02:00
|
|
|
/// 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.
|
2009-04-29 07:20:52 +02:00
|
|
|
virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
|
|
|
|
MVT VT) const {
|
2006-04-20 10:54:13 +02:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2005-08-24 18:34:59 +02:00
|
|
|
/// 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.
|
2008-06-06 14:08:01 +02:00
|
|
|
LegalizeAction getOperationAction(unsigned Op, MVT VT) const {
|
|
|
|
if (VT.isExtended()) return Expand;
|
2009-06-25 00:22:02 +02:00
|
|
|
assert(Op < array_lengthof(OpActions[0]) &&
|
|
|
|
(unsigned)VT.getSimpleVT() < sizeof(OpActions[0][0])*8 &&
|
2008-06-06 14:08:01 +02:00
|
|
|
"Table isn't big enough!");
|
2009-06-25 00:22:02 +02:00
|
|
|
unsigned I = (unsigned) VT.getSimpleVT();
|
|
|
|
unsigned J = I & 31;
|
|
|
|
I = I >> 5;
|
|
|
|
return (LegalizeAction)((OpActions[I][Op] >> (J*2) ) & 3);
|
2005-01-07 08:44:22 +01:00
|
|
|
}
|
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.
llvm-svn: 52254
2008-06-13 21:07:40 +02:00
|
|
|
|
2009-01-28 18:46:25 +01:00
|
|
|
/// 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.
|
|
|
|
bool isOperationLegalOrCustom(unsigned Op, MVT VT) const {
|
|
|
|
return (VT == MVT::Other || isTypeLegal(VT)) &&
|
|
|
|
(getOperationAction(Op, VT) == Legal ||
|
|
|
|
getOperationAction(Op, VT) == Custom);
|
|
|
|
}
|
|
|
|
|
2005-08-24 18:34:59 +02:00
|
|
|
/// isOperationLegal - Return true if the specified operation is legal on this
|
|
|
|
/// target.
|
2008-06-06 14:08:01 +02:00
|
|
|
bool isOperationLegal(unsigned Op, MVT VT) 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.
llvm-svn: 52254
2008-06-13 21:07:40 +02:00
|
|
|
return (VT == MVT::Other || isTypeLegal(VT)) &&
|
2009-01-28 18:46:25 +01:00
|
|
|
getOperationAction(Op, VT) == Legal;
|
2005-01-16 08:27:49 +01:00
|
|
|
}
|
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.
llvm-svn: 52254
2008-06-13 21:07:40 +02:00
|
|
|
|
2008-10-14 23:26:46 +02:00
|
|
|
/// getLoadExtAction - Return how this load with extension should be treated:
|
2006-10-04 02:56:09 +02:00
|
|
|
/// 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.
|
2008-10-14 23:26:46 +02:00
|
|
|
LegalizeAction getLoadExtAction(unsigned LType, MVT VT) const {
|
|
|
|
assert(LType < array_lengthof(LoadExtActions) &&
|
|
|
|
(unsigned)VT.getSimpleVT() < sizeof(LoadExtActions[0])*4 &&
|
2008-06-06 14:08:01 +02:00
|
|
|
"Table isn't big enough!");
|
2008-10-14 23:26:46 +02:00
|
|
|
return (LegalizeAction)((LoadExtActions[LType] >> (2*VT.getSimpleVT())) & 3);
|
2006-10-04 02:56:09 +02:00
|
|
|
}
|
2008-06-12 11:01:30 +02:00
|
|
|
|
2008-10-14 23:26:46 +02:00
|
|
|
/// isLoadExtLegal - Return true if the specified load with extension is legal
|
2006-10-13 23:14:26 +02:00
|
|
|
/// on this target.
|
2008-10-14 23:26:46 +02:00
|
|
|
bool isLoadExtLegal(unsigned LType, MVT VT) const {
|
2008-06-06 14:08:01 +02:00
|
|
|
return VT.isSimple() &&
|
2008-10-14 23:26:46 +02:00
|
|
|
(getLoadExtAction(LType, VT) == Legal ||
|
|
|
|
getLoadExtAction(LType, VT) == Custom);
|
2006-10-04 02:56:09 +02:00
|
|
|
}
|
2008-06-12 11:01:30 +02:00
|
|
|
|
2008-01-17 20:59:44 +01:00
|
|
|
/// 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.
|
2008-06-06 14:08:01 +02:00
|
|
|
LegalizeAction getTruncStoreAction(MVT ValVT,
|
|
|
|
MVT MemVT) const {
|
|
|
|
assert((unsigned)ValVT.getSimpleVT() < array_lengthof(TruncStoreActions) &&
|
|
|
|
(unsigned)MemVT.getSimpleVT() < sizeof(TruncStoreActions[0])*4 &&
|
|
|
|
"Table isn't big enough!");
|
|
|
|
return (LegalizeAction)((TruncStoreActions[ValVT.getSimpleVT()] >>
|
|
|
|
(2*MemVT.getSimpleVT())) & 3);
|
2006-10-13 23:14:26 +02:00
|
|
|
}
|
2008-06-12 11:01:30 +02:00
|
|
|
|
2008-01-17 20:59:44 +01:00
|
|
|
/// isTruncStoreLegal - Return true if the specified store with truncation is
|
2006-10-13 23:14:26 +02:00
|
|
|
/// legal on this target.
|
2008-06-06 14:08:01 +02:00
|
|
|
bool isTruncStoreLegal(MVT ValVT, MVT 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.
llvm-svn: 52254
2008-06-13 21:07:40 +02:00
|
|
|
return isTypeLegal(ValVT) && MemVT.isSimple() &&
|
2008-01-25 11:20:53 +01:00
|
|
|
(getTruncStoreAction(ValVT, MemVT) == Legal ||
|
|
|
|
getTruncStoreAction(ValVT, MemVT) == Custom);
|
2006-10-13 23:14:26 +02:00
|
|
|
}
|
|
|
|
|
2006-11-09 19:56:43 +01:00
|
|
|
/// 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
|
2008-06-06 14:08:01 +02:00
|
|
|
getIndexedLoadAction(unsigned IdxMode, MVT VT) const {
|
2009-06-09 02:19:10 +02:00
|
|
|
assert( IdxMode < array_lengthof(IndexedModeActions[0][0]) &&
|
|
|
|
((unsigned)VT.getSimpleVT()) < MVT::LAST_VALUETYPE &&
|
2008-01-25 11:20:53 +01:00
|
|
|
"Table isn't big enough!");
|
2009-06-09 02:19:10 +02:00
|
|
|
return (LegalizeAction)((IndexedModeActions[(unsigned)VT.getSimpleVT()][0][IdxMode]));
|
2006-11-09 19:56:43 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/// isIndexedLoadLegal - Return true if the specified indexed load is legal
|
|
|
|
/// on this target.
|
2008-06-06 14:08:01 +02:00
|
|
|
bool isIndexedLoadLegal(unsigned IdxMode, MVT VT) const {
|
2008-06-12 11:01:30 +02:00
|
|
|
return VT.isSimple() &&
|
|
|
|
(getIndexedLoadAction(IdxMode, VT) == Legal ||
|
|
|
|
getIndexedLoadAction(IdxMode, VT) == Custom);
|
2006-11-09 19:56:43 +01:00
|
|
|
}
|
2008-06-12 11:01:30 +02:00
|
|
|
|
2006-11-09 19:56:43 +01:00
|
|
|
/// 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
|
2008-06-06 14:08:01 +02:00
|
|
|
getIndexedStoreAction(unsigned IdxMode, MVT VT) const {
|
2009-06-09 02:19:10 +02:00
|
|
|
assert(IdxMode < array_lengthof(IndexedModeActions[0][1]) &&
|
|
|
|
(unsigned)VT.getSimpleVT() < MVT::LAST_VALUETYPE &&
|
2008-01-25 11:20:53 +01:00
|
|
|
"Table isn't big enough!");
|
2009-06-09 02:19:10 +02:00
|
|
|
return (LegalizeAction)((IndexedModeActions[(unsigned)VT.getSimpleVT()][1][IdxMode]));
|
2006-11-09 19:56:43 +01:00
|
|
|
}
|
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.
llvm-svn: 52254
2008-06-13 21:07:40 +02:00
|
|
|
|
2006-11-09 19:56:43 +01:00
|
|
|
/// isIndexedStoreLegal - Return true if the specified indexed load is legal
|
|
|
|
/// on this target.
|
2008-06-06 14:08:01 +02:00
|
|
|
bool isIndexedStoreLegal(unsigned IdxMode, MVT VT) const {
|
2008-06-12 11:01:30 +02:00
|
|
|
return VT.isSimple() &&
|
|
|
|
(getIndexedStoreAction(IdxMode, VT) == Legal ||
|
|
|
|
getIndexedStoreAction(IdxMode, VT) == Custom);
|
2006-11-09 19:56:43 +01:00
|
|
|
}
|
2008-06-12 11:01:30 +02:00
|
|
|
|
2007-08-09 03:04:01 +02:00
|
|
|
/// getConvertAction - Return how the conversion 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
|
2008-06-06 14:08:01 +02:00
|
|
|
getConvertAction(MVT FromVT, MVT ToVT) const {
|
|
|
|
assert((unsigned)FromVT.getSimpleVT() < array_lengthof(ConvertActions) &&
|
|
|
|
(unsigned)ToVT.getSimpleVT() < sizeof(ConvertActions[0])*4 &&
|
|
|
|
"Table isn't big enough!");
|
|
|
|
return (LegalizeAction)((ConvertActions[FromVT.getSimpleVT()] >>
|
|
|
|
(2*ToVT.getSimpleVT())) & 3);
|
2007-08-09 03:04:01 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/// isConvertLegal - Return true if the specified conversion is legal
|
|
|
|
/// on this target.
|
2008-06-06 14:08:01 +02:00
|
|
|
bool isConvertLegal(MVT FromVT, MVT ToVT) 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.
llvm-svn: 52254
2008-06-13 21:07:40 +02:00
|
|
|
return isTypeLegal(FromVT) && isTypeLegal(ToVT) &&
|
2008-06-12 11:01:30 +02:00
|
|
|
(getConvertAction(FromVT, ToVT) == Legal ||
|
|
|
|
getConvertAction(FromVT, ToVT) == Custom);
|
2007-08-09 03:04:01 +02:00
|
|
|
}
|
|
|
|
|
2008-10-15 04:05:31 +02:00
|
|
|
/// 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
|
|
|
|
getCondCodeAction(ISD::CondCode CC, MVT VT) const {
|
|
|
|
assert((unsigned)CC < array_lengthof(CondCodeActions) &&
|
|
|
|
(unsigned)VT.getSimpleVT() < sizeof(CondCodeActions[0])*4 &&
|
|
|
|
"Table isn't big enough!");
|
|
|
|
LegalizeAction Action = (LegalizeAction)
|
|
|
|
((CondCodeActions[CC] >> (2*VT.getSimpleVT())) & 3);
|
|
|
|
assert(Action != Promote && "Can't promote condition code!");
|
|
|
|
return Action;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// isCondCodeLegal - Return true if the specified condition code is legal
|
|
|
|
/// on this target.
|
|
|
|
bool isCondCodeLegal(ISD::CondCode CC, MVT VT) const {
|
|
|
|
return getCondCodeAction(CC, VT) == Legal ||
|
|
|
|
getCondCodeAction(CC, VT) == Custom;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2005-01-16 08:27:49 +01:00
|
|
|
/// getTypeToPromoteTo - If the action for this operation is to promote, this
|
|
|
|
/// method returns the ValueType to promote to.
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT getTypeToPromoteTo(unsigned Op, MVT VT) const {
|
2005-01-16 08:27:49 +01:00
|
|
|
assert(getOperationAction(Op, VT) == Promote &&
|
|
|
|
"This operation isn't promoted!");
|
2006-04-04 02:25:10 +02:00
|
|
|
|
|
|
|
// See if this has an explicit type specified.
|
2008-06-08 22:54:56 +02:00
|
|
|
std::map<std::pair<unsigned, MVT::SimpleValueType>,
|
|
|
|
MVT::SimpleValueType>::const_iterator PTTI =
|
|
|
|
PromoteToType.find(std::make_pair(Op, VT.getSimpleVT()));
|
2006-04-04 02:25:10 +02:00
|
|
|
if (PTTI != PromoteToType.end()) return PTTI->second;
|
2008-06-06 14:08:01 +02:00
|
|
|
|
|
|
|
assert((VT.isInteger() || VT.isFloatingPoint()) &&
|
2006-04-04 02:25:10 +02:00
|
|
|
"Cannot autopromote this type, add it with AddPromotedToType.");
|
|
|
|
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT NVT = VT;
|
2005-01-16 08:27:49 +01:00
|
|
|
do {
|
2008-06-06 14:08:01 +02:00
|
|
|
NVT = (MVT::SimpleValueType)(NVT.getSimpleVT()+1);
|
|
|
|
assert(NVT.isInteger() == VT.isInteger() && NVT != MVT::isVoid &&
|
2005-01-16 08:27:49 +01:00
|
|
|
"Didn't find type to promote to!");
|
2005-08-24 18:34:59 +02:00
|
|
|
} while (!isTypeLegal(NVT) ||
|
|
|
|
getOperationAction(Op, NVT) == Promote);
|
2005-01-16 08:27:49 +01:00
|
|
|
return NVT;
|
|
|
|
}
|
|
|
|
|
2008-06-06 14:08:01 +02:00
|
|
|
/// getValueType - Return the MVT corresponding to this LLVM type.
|
2007-04-28 07:42:38 +02:00
|
|
|
/// This is fixed by the LLVM operations except for the pointer size. If
|
|
|
|
/// AllowUnknown is true, this will return MVT::Other for types with no MVT
|
|
|
|
/// counterpart (e.g. structs), otherwise it will assert.
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT getValueType(const Type *Ty, bool AllowUnknown = false) const {
|
|
|
|
MVT VT = MVT::getMVT(Ty, AllowUnknown);
|
2007-03-31 06:05:24 +02:00
|
|
|
return VT == MVT::iPTR ? PointerTy : VT;
|
|
|
|
}
|
2005-04-21 22:59:05 +02:00
|
|
|
|
2008-01-24 00:17:41 +01:00
|
|
|
/// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
|
2008-02-28 23:31:51 +01:00
|
|
|
/// function arguments in the caller parameter area. This is the actual
|
|
|
|
/// alignment, not its logarithm.
|
2008-01-24 00:17:41 +01:00
|
|
|
virtual unsigned getByValTypeAlignment(const Type *Ty) const;
|
|
|
|
|
2007-06-29 01:29:44 +02:00
|
|
|
/// getRegisterType - Return the type of registers that this ValueType will
|
|
|
|
/// eventually require.
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT getRegisterType(MVT VT) const {
|
|
|
|
if (VT.isSimple()) {
|
|
|
|
assert((unsigned)VT.getSimpleVT() < array_lengthof(RegisterTypeForVT));
|
|
|
|
return RegisterTypeForVT[VT.getSimpleVT()];
|
2008-01-25 11:20:53 +01:00
|
|
|
}
|
2008-06-06 14:08:01 +02:00
|
|
|
if (VT.isVector()) {
|
|
|
|
MVT VT1, RegisterVT;
|
2007-10-16 11:56:48 +02:00
|
|
|
unsigned NumIntermediates;
|
|
|
|
(void)getVectorTypeBreakdown(VT, VT1, NumIntermediates, RegisterVT);
|
|
|
|
return RegisterVT;
|
|
|
|
}
|
2008-06-06 14:08:01 +02:00
|
|
|
if (VT.isInteger()) {
|
2008-02-11 12:09:23 +01:00
|
|
|
return getRegisterType(getTypeToTransformTo(VT));
|
|
|
|
}
|
2007-10-16 11:56:48 +02:00
|
|
|
assert(0 && "Unsupported extended type!");
|
2009-07-03 18:00:23 +02:00
|
|
|
return MVT(MVT::Other); // Not reached
|
2007-06-29 01:29:44 +02:00
|
|
|
}
|
2007-10-16 11:56:48 +02:00
|
|
|
|
2007-06-21 16:42:22 +02:00
|
|
|
/// getNumRegisters - Return the number of registers that this ValueType will
|
2006-12-13 07:09:03 +01:00
|
|
|
/// 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 12:09:23 +01:00
|
|
|
/// 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.
|
2008-06-06 14:08:01 +02:00
|
|
|
unsigned getNumRegisters(MVT VT) const {
|
|
|
|
if (VT.isSimple()) {
|
|
|
|
assert((unsigned)VT.getSimpleVT() < array_lengthof(NumRegistersForVT));
|
|
|
|
return NumRegistersForVT[VT.getSimpleVT()];
|
2008-01-25 11:20:53 +01:00
|
|
|
}
|
2008-06-06 14:08:01 +02:00
|
|
|
if (VT.isVector()) {
|
|
|
|
MVT VT1, VT2;
|
2007-10-16 11:56:48 +02:00
|
|
|
unsigned NumIntermediates;
|
|
|
|
return getVectorTypeBreakdown(VT, VT1, NumIntermediates, VT2);
|
|
|
|
}
|
2008-06-06 14:08:01 +02:00
|
|
|
if (VT.isInteger()) {
|
|
|
|
unsigned BitWidth = VT.getSizeInBits();
|
|
|
|
unsigned RegWidth = getRegisterType(VT).getSizeInBits();
|
2008-02-11 12:09:23 +01:00
|
|
|
return (BitWidth + RegWidth - 1) / RegWidth;
|
|
|
|
}
|
2007-10-16 11:56:48 +02:00
|
|
|
assert(0 && "Unsupported extended type!");
|
2008-03-30 20:22:13 +02:00
|
|
|
return 0; // Not reached
|
2005-01-07 08:44:22 +01:00
|
|
|
}
|
2008-01-24 00:17:41 +01:00
|
|
|
|
2008-03-05 02:30:59 +01:00
|
|
|
/// 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.
|
2008-06-06 14:08:01 +02:00
|
|
|
virtual bool ShouldShrinkFPConstant(MVT VT) const { return true; }
|
2008-03-05 02:30:59 +01:00
|
|
|
|
2006-03-01 05:52:55 +01:00
|
|
|
/// 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 11:20:53 +01:00
|
|
|
assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray));
|
2006-03-01 05:52:55 +01:00
|
|
|
return TargetDAGCombineArray[NT >> 3] & (1 << (NT&7));
|
|
|
|
}
|
2005-01-07 08:44:22 +01:00
|
|
|
|
2005-07-19 06:52:44 +02:00
|
|
|
/// 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
|
|
|
|
/// performance threshold for such a replacement.
|
|
|
|
/// @brief Get maximum # of store operations permitted for llvm.memset
|
2006-02-14 09:38:30 +01:00
|
|
|
unsigned getMaxStoresPerMemset() const { return maxStoresPerMemset; }
|
2005-07-19 06:52:44 +02:00
|
|
|
|
|
|
|
/// 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
|
|
|
|
/// performance threshold for such a replacement.
|
|
|
|
/// @brief Get maximum # of store operations permitted for llvm.memcpy
|
2006-02-14 09:38:30 +01:00
|
|
|
unsigned getMaxStoresPerMemcpy() const { return maxStoresPerMemcpy; }
|
2005-07-19 06:52:44 +02:00
|
|
|
|
|
|
|
/// 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
|
|
|
|
/// performance threshold for such a replacement.
|
|
|
|
/// @brief Get maximum # of store operations permitted for llvm.memmove
|
2006-02-14 09:38:30 +01:00
|
|
|
unsigned getMaxStoresPerMemmove() const { return maxStoresPerMemmove; }
|
2005-07-19 06:52:44 +02:00
|
|
|
|
2005-08-27 21:09:02 +02:00
|
|
|
/// This function returns true if the target allows unaligned memory accesses.
|
|
|
|
/// 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.
|
|
|
|
/// @brief Determine if the target supports unaligned memory accesses.
|
2006-01-29 09:40:37 +01:00
|
|
|
bool allowsUnalignedMemoryAccesses() const {
|
|
|
|
return allowUnalignedMemoryAccesses;
|
|
|
|
}
|
2008-05-15 10:39:06 +02:00
|
|
|
|
2009-05-13 23:42:09 +02:00
|
|
|
/// 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 10:39:06 +02:00
|
|
|
/// getOptimalMemOpType - Returns the target specific optimal type for load
|
2008-05-16 00:13:02 +02:00
|
|
|
/// and store operations as a result of memset, memcpy, and memmove lowering.
|
2008-05-15 10:39:06 +02:00
|
|
|
/// It returns MVT::iAny if SelectionDAG should be responsible for
|
|
|
|
/// determining it.
|
2008-06-06 14:08:01 +02:00
|
|
|
virtual MVT getOptimalMemOpType(uint64_t Size, unsigned Align,
|
2009-06-05 23:57:13 +02:00
|
|
|
bool isSrcConst, bool isSrcStr,
|
|
|
|
SelectionDAG &DAG) const {
|
2008-05-15 10:39:06 +02:00
|
|
|
return MVT::iAny;
|
|
|
|
}
|
2005-09-28 00:13:36 +02:00
|
|
|
|
2006-12-11 00:12:42 +01:00
|
|
|
/// usesUnderscoreSetJmp - Determine if we should use _setjmp or setjmp
|
2005-09-28 00:13:36 +02:00
|
|
|
/// to implement llvm.setjmp.
|
2006-12-11 00:12:42 +01:00
|
|
|
bool usesUnderscoreSetJmp() const {
|
|
|
|
return UseUnderscoreSetJmp;
|
2005-09-28 00:13:36 +02:00
|
|
|
}
|
2006-12-11 00:12:42 +01:00
|
|
|
|
|
|
|
/// usesUnderscoreLongJmp - Determine if we should use _longjmp or longjmp
|
|
|
|
/// to implement llvm.longjmp.
|
|
|
|
bool usesUnderscoreLongJmp() const {
|
|
|
|
return UseUnderscoreLongJmp;
|
|
|
|
}
|
|
|
|
|
2006-01-13 18:47:52 +01:00
|
|
|
/// 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 06:52:44 +02:00
|
|
|
|
2007-02-22 16:37:19 +01:00
|
|
|
/// getExceptionAddressRegister - If a physical register, this returns
|
|
|
|
/// the register that receives the exception address on entry to a landing
|
|
|
|
/// pad.
|
|
|
|
unsigned getExceptionAddressRegister() const {
|
|
|
|
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 08:21:35 +02:00
|
|
|
/// 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 04:04:06 +01:00
|
|
|
|
2007-05-17 01:45:53 +02:00
|
|
|
/// getIfCvtBlockLimit - returns the target specific if-conversion block size
|
|
|
|
/// limit. Any block whose size is greater should not be predicated.
|
2008-02-28 01:43:03 +01:00
|
|
|
unsigned getIfCvtBlockSizeLimit() const {
|
2007-05-17 01:45:53 +02:00
|
|
|
return IfCvtBlockSizeLimit;
|
|
|
|
}
|
|
|
|
|
2007-06-01 10:25:24 +02:00
|
|
|
/// getIfCvtDupBlockLimit - returns the target specific size limit for a
|
|
|
|
/// block to be considered for duplication. Any block whose size is greater
|
|
|
|
/// should not be duplicated to facilitate its predication.
|
2008-02-28 01:43:03 +01:00
|
|
|
unsigned getIfCvtDupBlockSizeLimit() const {
|
2007-06-01 10:25:24 +02:00
|
|
|
return IfCvtDupBlockSizeLimit;
|
|
|
|
}
|
|
|
|
|
2008-02-28 01:43:03 +01:00
|
|
|
/// getPrefLoopAlignment - return the preferred loop alignment.
|
|
|
|
///
|
|
|
|
unsigned getPrefLoopAlignment() const {
|
|
|
|
return PrefLoopAlignment;
|
|
|
|
}
|
|
|
|
|
2006-11-03 08:21:16 +01:00
|
|
|
/// getPreIndexedAddressParts - returns true by value, base pointer and
|
2006-11-03 04:04:06 +01:00
|
|
|
/// offset pointer and addressing mode by reference if the node's address
|
|
|
|
/// can be legally represented as pre-indexed load / store address.
|
2008-07-27 23:46:04 +02:00
|
|
|
virtual bool getPreIndexedAddressParts(SDNode *N, SDValue &Base,
|
|
|
|
SDValue &Offset,
|
2006-11-09 18:55:04 +01:00
|
|
|
ISD::MemIndexedMode &AM,
|
2009-01-15 17:29:45 +01:00
|
|
|
SelectionDAG &DAG) const {
|
2006-11-03 04:04:06 +01:00
|
|
|
return false;
|
|
|
|
}
|
2006-09-04 08:21:35 +02:00
|
|
|
|
2006-11-07 10:04:16 +01:00
|
|
|
/// 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.
|
2006-11-09 05:29:46 +01:00
|
|
|
virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op,
|
2008-07-27 23:46:04 +02:00
|
|
|
SDValue &Base, SDValue &Offset,
|
2006-11-09 18:55:04 +01:00
|
|
|
ISD::MemIndexedMode &AM,
|
2009-01-15 17:29:45 +01:00
|
|
|
SelectionDAG &DAG) const {
|
2006-11-07 10:04:16 +01:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
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
llvm-svn: 43924
2007-11-09 02:32:10 +01:00
|
|
|
/// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
|
|
|
|
/// jumptable.
|
2008-07-27 23:46:04 +02:00
|
|
|
virtual SDValue getPICJumpTableRelocBase(SDValue Table,
|
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
llvm-svn: 43924
2007-11-09 02:32:10 +01:00
|
|
|
SelectionDAG &DAG) const;
|
2008-02-28 01:43:03 +01:00
|
|
|
|
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.
llvm-svn: 57748
2008-10-18 04:06:02 +02:00
|
|
|
/// 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;
|
|
|
|
|
2009-07-01 20:50:55 +02:00
|
|
|
/// getFunctionAlignment - Return the Log2 alignment of this function.
|
2009-07-01 00:38:32 +02:00
|
|
|
virtual unsigned getFunctionAlignment(const Function *) const = 0;
|
|
|
|
|
2006-01-30 05:08:18 +01:00
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// TargetLowering Optimization Methods
|
|
|
|
//
|
|
|
|
|
2006-02-16 22:11:51 +01:00
|
|
|
/// TargetLoweringOpt - A convenience struct that encapsulates a DAG, and two
|
2008-07-27 23:46:04 +02:00
|
|
|
/// SDValues for returning information from TargetLowering to its clients
|
2006-02-16 22:11:51 +01:00
|
|
|
/// that want to combine
|
|
|
|
struct TargetLoweringOpt {
|
|
|
|
SelectionDAG &DAG;
|
2008-07-27 23:46:04 +02:00
|
|
|
SDValue Old;
|
|
|
|
SDValue New;
|
2006-02-16 22:11:51 +01:00
|
|
|
|
2008-11-24 15:53:14 +01:00
|
|
|
explicit TargetLoweringOpt(SelectionDAG &InDAG) : DAG(InDAG) {}
|
2006-02-16 22:11:51 +01:00
|
|
|
|
2008-07-27 23:46:04 +02:00
|
|
|
bool CombineTo(SDValue O, SDValue N) {
|
2006-02-16 22:11:51 +01:00
|
|
|
Old = O;
|
|
|
|
New = N;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// ShrinkDemandedConstant - Check to see if the specified operand of the
|
2007-10-15 07:30:27 +02:00
|
|
|
/// 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 23:46:04 +02:00
|
|
|
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.
llvm-svn: 68576
2009-04-08 02:15:30 +02:00
|
|
|
|
|
|
|
/// 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 22:11:51 +01:00
|
|
|
};
|
|
|
|
|
|
|
|
/// 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
|
|
|
|
/// 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
|
|
|
|
/// DemandedMask.
|
2008-07-27 23:46:04 +02:00
|
|
|
bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedMask,
|
2008-02-27 01:25:32 +01:00
|
|
|
APInt &KnownZero, APInt &KnownOne,
|
2006-02-16 22:11:51 +01:00
|
|
|
TargetLoweringOpt &TLO, unsigned Depth = 0) const;
|
2006-02-03 23:24:05 +01:00
|
|
|
|
2006-03-01 05:52:55 +01:00
|
|
|
/// computeMaskedBitsForTargetNode - Determine which of the bits specified in
|
|
|
|
/// Mask are known to be either zero or one and return them in the
|
|
|
|
/// KnownZero/KnownOne bitsets.
|
2008-07-27 23:46:04 +02:00
|
|
|
virtual void computeMaskedBitsForTargetNode(const SDValue Op,
|
2008-02-13 23:28:48 +01:00
|
|
|
const APInt &Mask,
|
2008-02-13 01:35:47 +01:00
|
|
|
APInt &KnownZero,
|
|
|
|
APInt &KnownOne,
|
2007-06-22 16:59:07 +02:00
|
|
|
const SelectionDAG &DAG,
|
2006-03-01 05:52:55 +01:00
|
|
|
unsigned Depth = 0) const;
|
2006-03-14 00:18:16 +01:00
|
|
|
|
2006-05-06 11:26:22 +02:00
|
|
|
/// ComputeNumSignBitsForTargetNode - This method can be implemented by
|
|
|
|
/// targets that want to expose additional information about sign bits to the
|
|
|
|
/// DAG Combiner.
|
2008-07-27 23:46:04 +02:00
|
|
|
virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
|
2006-05-06 11:26:22 +02:00
|
|
|
unsigned Depth = 0) const;
|
|
|
|
|
2006-03-01 05:52:55 +01:00
|
|
|
struct DAGCombinerInfo {
|
|
|
|
void *DC; // The DAG Combiner object.
|
|
|
|
bool BeforeLegalize;
|
2007-02-08 23:13:59 +01:00
|
|
|
bool CalledByLegalizer;
|
2006-03-01 05:52:55 +01:00
|
|
|
public:
|
|
|
|
SelectionDAG &DAG;
|
|
|
|
|
2007-02-08 23:13:59 +01:00
|
|
|
DAGCombinerInfo(SelectionDAG &dag, bool bl, bool cl, void *dc)
|
|
|
|
: DC(dc), BeforeLegalize(bl), CalledByLegalizer(cl), DAG(dag) {}
|
2006-03-01 05:52:55 +01:00
|
|
|
|
|
|
|
bool isBeforeLegalize() const { return BeforeLegalize; }
|
2007-02-08 23:13:59 +01:00
|
|
|
bool isCalledByLegalizer() const { return CalledByLegalizer; }
|
2006-03-01 05:52:55 +01:00
|
|
|
|
|
|
|
void AddToWorklist(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
llvm-svn: 67917
2009-03-28 06:57:29 +01:00
|
|
|
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 02:59:02 +01:00
|
|
|
|
|
|
|
void CommitTargetLoweringOpt(const TargetLoweringOpt &TLO);
|
2006-03-01 05:52:55 +01:00
|
|
|
};
|
|
|
|
|
2007-02-08 23:13:59 +01:00
|
|
|
/// SimplifySetCC - Try to simplify a setcc built with the specified operands
|
2008-07-27 23:46:04 +02:00
|
|
|
/// and cc. If it is unable to simplify it, return a null SDValue.
|
|
|
|
SDValue SimplifySetCC(MVT VT, SDValue N0, SDValue N1,
|
2007-02-08 23:13:59 +01:00
|
|
|
ISD::CondCode Cond, bool foldBooleans,
|
2009-02-03 01:47:48 +01:00
|
|
|
DAGCombinerInfo &DCI, DebugLoc dl) const;
|
2007-02-08 23:13:59 +01:00
|
|
|
|
2008-05-12 22:08:05 +02:00
|
|
|
/// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
|
|
|
|
/// node is a GlobalAddress + offset.
|
|
|
|
virtual bool
|
|
|
|
isGAPlusOffset(SDNode *N, GlobalValue* &GA, int64_t &Offset) const;
|
|
|
|
|
Adapt the x86 build_vector dagcombine to the current state of the legalizer.
build vectors with i64 elements will only appear on 32b x86 before legalize.
Since vector widening occurs during legalize, and produces i64 build_vector
elements, the dag combiner is never run on these before legalize splits them
into 32b elements.
Teach the build_vector dag combine in x86 back end to recognize consecutive
loads producing the low part of the vector.
Convert the two uses of TLI's consecutive load recognizer to pass LoadSDNodes
since that was required implicitly.
Add a testcase for the transform.
Old:
subl $28, %esp
movl 32(%esp), %eax
movl 4(%eax), %ecx
movl %ecx, 4(%esp)
movl (%eax), %eax
movl %eax, (%esp)
movaps (%esp), %xmm0
pmovzxwd %xmm0, %xmm0
movl 36(%esp), %eax
movaps %xmm0, (%eax)
addl $28, %esp
ret
New:
movl 4(%esp), %eax
pmovzxwd (%eax), %xmm0
movl 8(%esp), %eax
movaps %xmm0, (%eax)
ret
llvm-svn: 72957
2009-06-05 23:37:30 +02:00
|
|
|
/// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a
|
|
|
|
/// location that is 'Dist' units away from the location that the 'Base' load
|
|
|
|
/// is loading from.
|
|
|
|
bool isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base, unsigned Bytes,
|
|
|
|
int Dist, const MachineFrameInfo *MFI) const;
|
2008-05-12 22:08:05 +02:00
|
|
|
|
2006-03-01 05:52:55 +01:00
|
|
|
/// 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 23:46:04 +02:00
|
|
|
/// SDValue.Val == 0 - No change was made
|
|
|
|
/// SDValue.Val == N - N was replaced, is dead, and is already handled.
|
2008-11-24 15:53:14 +01:00
|
|
|
/// otherwise - N should be replaced by the returned Operand.
|
2006-03-01 05:52:55 +01:00
|
|
|
///
|
|
|
|
/// In addition, methods provided by DAGCombinerInfo may be used to perform
|
|
|
|
/// more complex transformations.
|
|
|
|
///
|
2008-07-27 23:46:04 +02:00
|
|
|
virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
|
2006-03-01 05:52:55 +01:00
|
|
|
|
2005-01-07 08:44:22 +01:00
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// TargetLowering Configuration Methods - These methods should be invoked by
|
|
|
|
// the derived class constructor to configure this object for the target.
|
|
|
|
//
|
|
|
|
|
|
|
|
protected:
|
2006-10-07 00:46:34 +02:00
|
|
|
/// setUsesGlobalOffsetTable - Specify that this target does or doesn't use a
|
|
|
|
/// GOT for PC-relative code.
|
|
|
|
void setUsesGlobalOffsetTable(bool V) { UsesGlobalOffsetTable = V; }
|
2005-01-07 08:44:22 +01:00
|
|
|
|
2005-01-17 00:59:30 +01:00
|
|
|
/// setShiftAmountType - Describe the type that should be used for shift
|
|
|
|
/// amounts. This type defaults to the pointer type.
|
2008-06-06 14:08:01 +02:00
|
|
|
void setShiftAmountType(MVT VT) { ShiftAmountTy = VT; }
|
2005-01-17 00:59:30 +01:00
|
|
|
|
2008-11-23 16:47:28 +01:00
|
|
|
/// 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; }
|
2005-04-07 21:41:18 +02:00
|
|
|
|
2006-01-25 10:09:02 +01:00
|
|
|
/// setSchedulingPreference - Specify the target scheduling preference.
|
|
|
|
void setSchedulingPreference(SchedPreference Pref) {
|
|
|
|
SchedPreferenceInfo = Pref;
|
|
|
|
}
|
|
|
|
|
2005-01-19 04:36:03 +01:00
|
|
|
/// setShiftAmountFlavor - Describe how the target handles out of range shift
|
|
|
|
/// amounts.
|
|
|
|
void setShiftAmountFlavor(OutOfRangeShiftAmount OORSA) {
|
|
|
|
ShiftAmtHandling = OORSA;
|
|
|
|
}
|
2005-04-21 22:59:05 +02:00
|
|
|
|
2006-12-11 00:12:42 +01:00
|
|
|
/// 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;
|
2005-09-28 00:13:36 +02:00
|
|
|
}
|
2006-12-11 00:12:42 +01:00
|
|
|
|
|
|
|
/// 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;
|
|
|
|
}
|
|
|
|
|
2006-01-13 18:47:52 +01:00
|
|
|
/// 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;
|
|
|
|
}
|
|
|
|
|
2007-02-22 16:37:19 +01:00
|
|
|
/// 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;
|
|
|
|
}
|
|
|
|
|
2007-01-06 00:31:08 +01:00
|
|
|
/// SelectIsExpensive - Tells the code generator not to expand operations
|
|
|
|
/// into sequences that use the select operations if possible.
|
|
|
|
void setSelectIsExpensive() { SelectIsExpensive = true; }
|
2005-05-12 04:02:21 +02:00
|
|
|
|
2005-10-21 02:02:42 +02:00
|
|
|
/// setIntDivIsCheap - Tells the code generator that integer divide is
|
2005-10-20 04:14:14 +02:00
|
|
|
/// expensive, and if possible, should be replaced by an alternate sequence
|
|
|
|
/// of instructions not containing an integer divide.
|
2005-10-21 02:02:42 +02:00
|
|
|
void setIntDivIsCheap(bool isCheap = true) { IntDivIsCheap = isCheap; }
|
|
|
|
|
|
|
|
/// 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; }
|
2005-10-20 04:14:14 +02:00
|
|
|
|
2005-01-07 08:44:22 +01:00
|
|
|
/// 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.
|
2008-06-06 14:08:01 +02:00
|
|
|
void addRegisterClass(MVT VT, TargetRegisterClass *RC) {
|
|
|
|
assert((unsigned)VT.getSimpleVT() < array_lengthof(RegClassForVT));
|
2005-01-07 08:44:22 +01:00
|
|
|
AvailableRegClasses.push_back(std::make_pair(VT, RC));
|
2008-06-06 14:08:01 +02:00
|
|
|
RegClassForVT[VT.getSimpleVT()] = RC;
|
2005-01-07 08:44:22 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/// computeRegisterProperties - Once all of the register classes are added,
|
|
|
|
/// this allows us to compute derived properties we expose.
|
|
|
|
void computeRegisterProperties();
|
2005-04-21 22:59:05 +02:00
|
|
|
|
2005-01-16 08:27:49 +01:00
|
|
|
/// setOperationAction - Indicate that the specified operation does not work
|
|
|
|
/// with the specified type and indicate what to do about it.
|
2008-06-06 14:08:01 +02:00
|
|
|
void setOperationAction(unsigned Op, MVT VT,
|
2005-01-16 08:27:49 +01:00
|
|
|
LegalizeAction Action) {
|
2009-06-25 00:22:02 +02:00
|
|
|
assert((unsigned)VT.getSimpleVT() < sizeof(OpActions[0][0])*8 &&
|
|
|
|
Op < array_lengthof(OpActions[0]) && "Table isn't big enough!");
|
|
|
|
unsigned I = (unsigned) VT.getSimpleVT();
|
|
|
|
unsigned J = I & 31;
|
|
|
|
I = I >> 5;
|
|
|
|
OpActions[I][Op] &= ~(uint64_t(3UL) << (J*2));
|
|
|
|
OpActions[I][Op] |= (uint64_t)Action << (J*2);
|
2005-01-07 08:44:22 +01:00
|
|
|
}
|
2006-04-04 02:25:10 +02:00
|
|
|
|
2008-10-14 23:26:46 +02:00
|
|
|
/// setLoadExtAction - Indicate that the specified load with extension does
|
|
|
|
/// not work with the with specified type and indicate what to do about it.
|
|
|
|
void setLoadExtAction(unsigned ExtType, MVT VT,
|
2006-10-09 22:55:20 +02:00
|
|
|
LegalizeAction Action) {
|
2008-10-14 23:26:46 +02:00
|
|
|
assert((unsigned)VT.getSimpleVT() < sizeof(LoadExtActions[0])*4 &&
|
|
|
|
ExtType < array_lengthof(LoadExtActions) &&
|
2006-10-04 02:56:09 +02:00
|
|
|
"Table isn't big enough!");
|
2008-10-14 23:26:46 +02:00
|
|
|
LoadExtActions[ExtType] &= ~(uint64_t(3UL) << VT.getSimpleVT()*2);
|
|
|
|
LoadExtActions[ExtType] |= (uint64_t)Action << VT.getSimpleVT()*2;
|
2006-10-04 02:56:09 +02:00
|
|
|
}
|
|
|
|
|
2008-01-17 20:59:44 +01:00
|
|
|
/// setTruncStoreAction - Indicate that the specified truncating store does
|
2006-10-13 23:14:26 +02:00
|
|
|
/// not work with the with specified type and indicate what to do about it.
|
2008-06-06 14:08:01 +02:00
|
|
|
void setTruncStoreAction(MVT ValVT, MVT MemVT,
|
2008-01-17 20:59:44 +01:00
|
|
|
LegalizeAction Action) {
|
2008-06-06 14:08:01 +02:00
|
|
|
assert((unsigned)ValVT.getSimpleVT() < array_lengthof(TruncStoreActions) &&
|
|
|
|
(unsigned)MemVT.getSimpleVT() < sizeof(TruncStoreActions[0])*4 &&
|
|
|
|
"Table isn't big enough!");
|
|
|
|
TruncStoreActions[ValVT.getSimpleVT()] &= ~(uint64_t(3UL) <<
|
|
|
|
MemVT.getSimpleVT()*2);
|
|
|
|
TruncStoreActions[ValVT.getSimpleVT()] |= (uint64_t)Action <<
|
|
|
|
MemVT.getSimpleVT()*2;
|
2006-10-13 23:14:26 +02:00
|
|
|
}
|
|
|
|
|
2006-11-09 19:56:43 +01:00
|
|
|
/// setIndexedLoadAction - Indicate that the specified indexed load does or
|
|
|
|
/// does not work with the with specified type and indicate what to do abort
|
|
|
|
/// it. NOTE: All indexed mode loads are initialized to Expand in
|
|
|
|
/// TargetLowering.cpp
|
2008-06-06 14:08:01 +02:00
|
|
|
void setIndexedLoadAction(unsigned IdxMode, MVT VT,
|
2006-11-09 19:56:43 +01:00
|
|
|
LegalizeAction Action) {
|
2009-06-09 02:19:10 +02:00
|
|
|
assert((unsigned)VT.getSimpleVT() < MVT::LAST_VALUETYPE &&
|
|
|
|
IdxMode < array_lengthof(IndexedModeActions[0][0]) &&
|
2006-11-09 19:56:43 +01:00
|
|
|
"Table isn't big enough!");
|
2009-06-09 02:19:10 +02:00
|
|
|
IndexedModeActions[(unsigned)VT.getSimpleVT()][0][IdxMode] = (uint8_t)Action;
|
2006-11-09 19:56:43 +01:00
|
|
|
}
|
|
|
|
|
|
|
|
/// setIndexedStoreAction - Indicate that the specified indexed store does or
|
|
|
|
/// does not work with the with specified type and indicate what to do about
|
|
|
|
/// it. NOTE: All indexed mode stores are initialized to Expand in
|
|
|
|
/// TargetLowering.cpp
|
2008-06-06 14:08:01 +02:00
|
|
|
void setIndexedStoreAction(unsigned IdxMode, MVT VT,
|
2006-11-09 19:56:43 +01:00
|
|
|
LegalizeAction Action) {
|
2009-06-09 02:19:10 +02:00
|
|
|
assert((unsigned)VT.getSimpleVT() < MVT::LAST_VALUETYPE &&
|
|
|
|
IdxMode < array_lengthof(IndexedModeActions[0][1] ) &&
|
2006-11-09 19:56:43 +01:00
|
|
|
"Table isn't big enough!");
|
2009-06-09 02:19:10 +02:00
|
|
|
IndexedModeActions[(unsigned)VT.getSimpleVT()][1][IdxMode] = (uint8_t)Action;
|
2006-11-09 19:56:43 +01:00
|
|
|
}
|
|
|
|
|
2007-08-09 03:04:01 +02:00
|
|
|
/// setConvertAction - Indicate that the specified conversion does or does
|
|
|
|
/// not work with the with specified type and indicate what to do about it.
|
2008-06-06 14:08:01 +02:00
|
|
|
void setConvertAction(MVT FromVT, MVT ToVT,
|
2007-08-09 03:04:01 +02:00
|
|
|
LegalizeAction Action) {
|
2008-06-06 14:08:01 +02:00
|
|
|
assert((unsigned)FromVT.getSimpleVT() < array_lengthof(ConvertActions) &&
|
|
|
|
(unsigned)ToVT.getSimpleVT() < sizeof(ConvertActions[0])*4 &&
|
|
|
|
"Table isn't big enough!");
|
|
|
|
ConvertActions[FromVT.getSimpleVT()] &= ~(uint64_t(3UL) <<
|
|
|
|
ToVT.getSimpleVT()*2);
|
|
|
|
ConvertActions[FromVT.getSimpleVT()] |= (uint64_t)Action <<
|
|
|
|
ToVT.getSimpleVT()*2;
|
2007-08-09 03:04:01 +02:00
|
|
|
}
|
|
|
|
|
2008-10-15 04:05:31 +02:00
|
|
|
/// setCondCodeAction - Indicate that the specified condition code is or isn't
|
|
|
|
/// supported on the target and indicate what to do about it.
|
|
|
|
void setCondCodeAction(ISD::CondCode CC, MVT VT, LegalizeAction Action) {
|
|
|
|
assert((unsigned)VT.getSimpleVT() < sizeof(CondCodeActions[0])*4 &&
|
|
|
|
(unsigned)CC < array_lengthof(CondCodeActions) &&
|
|
|
|
"Table isn't big enough!");
|
|
|
|
CondCodeActions[(unsigned)CC] &= ~(uint64_t(3UL) << VT.getSimpleVT()*2);
|
|
|
|
CondCodeActions[(unsigned)CC] |= (uint64_t)Action << VT.getSimpleVT()*2;
|
|
|
|
}
|
|
|
|
|
2006-04-04 02:25:10 +02:00
|
|
|
/// 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.
|
2008-06-06 14:08:01 +02:00
|
|
|
void AddPromotedToType(unsigned Opc, MVT OrigVT, MVT DestVT) {
|
2008-06-08 22:54:56 +02:00
|
|
|
PromoteToType[std::make_pair(Opc, OrigVT.getSimpleVT())] =
|
|
|
|
DestVT.getSimpleVT();
|
2006-04-04 02:25:10 +02:00
|
|
|
}
|
2005-01-07 08:44:22 +01:00
|
|
|
|
|
|
|
/// addLegalFPImmediate - Indicate that this target can instruction select
|
|
|
|
/// the specified FP immediate natively.
|
2007-08-30 02:23:21 +02:00
|
|
|
void addLegalFPImmediate(const APFloat& Imm) {
|
2005-01-07 08:44:22 +01:00
|
|
|
LegalFPImmediates.push_back(Imm);
|
|
|
|
}
|
|
|
|
|
2006-03-01 05:52:55 +01:00
|
|
|
/// 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 11:20:53 +01:00
|
|
|
assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray));
|
2006-03-01 05:52:55 +01:00
|
|
|
TargetDAGCombineArray[NT >> 3] |= 1 << (NT&7);
|
|
|
|
}
|
|
|
|
|
2006-09-04 08:21:35 +02:00
|
|
|
/// 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;
|
|
|
|
}
|
2007-05-17 01:45:53 +02:00
|
|
|
|
2007-05-18 02:00:30 +02:00
|
|
|
/// setIfCvtBlockSizeLimit - Set the target's if-conversion block size
|
2007-06-01 10:25:24 +02:00
|
|
|
/// limit (in number of instructions); default is 2.
|
2007-05-17 01:45:53 +02:00
|
|
|
void setIfCvtBlockSizeLimit(unsigned Limit) {
|
|
|
|
IfCvtBlockSizeLimit = Limit;
|
|
|
|
}
|
2006-09-04 08:21:35 +02:00
|
|
|
|
2007-06-01 10:25:24 +02:00
|
|
|
/// setIfCvtDupBlockSizeLimit - Set the target's block size limit (in number
|
|
|
|
/// of instructions) to be considered for code duplication during
|
|
|
|
/// if-conversion; default is 2.
|
|
|
|
void setIfCvtDupBlockSizeLimit(unsigned Limit) {
|
|
|
|
IfCvtDupBlockSizeLimit = Limit;
|
|
|
|
}
|
2008-02-28 01:43:03 +01:00
|
|
|
|
|
|
|
/// setPrefLoopAlignment - Set the target's preferred loop alignment. Default
|
|
|
|
/// alignment is zero, it means the target does not care about loop alignment.
|
|
|
|
void setPrefLoopAlignment(unsigned Align) {
|
|
|
|
PrefLoopAlignment = Align;
|
|
|
|
}
|
2007-06-01 10:25:24 +02:00
|
|
|
|
2005-01-07 08:44:22 +01:00
|
|
|
public:
|
|
|
|
|
2007-11-06 00:12:20 +01:00
|
|
|
virtual const TargetSubtarget *getSubtarget() {
|
|
|
|
assert(0 && "Not Implemented");
|
2007-11-07 20:33:31 +01:00
|
|
|
return NULL; // this is here to silence compiler errors
|
2007-11-06 00:12:20 +01:00
|
|
|
}
|
2005-01-07 08:44:22 +01:00
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// Lowering methods - These methods must be implemented by targets so that
|
|
|
|
// the SelectionDAGLowering code knows how to lower these.
|
|
|
|
//
|
|
|
|
|
|
|
|
/// LowerArguments - This hook must be implemented to indicate how we should
|
|
|
|
/// lower the arguments for the specified function, into the specified DAG.
|
2008-06-30 22:31:15 +02:00
|
|
|
virtual void
|
|
|
|
LowerArguments(Function &F, SelectionDAG &DAG,
|
2009-01-31 00:10:59 +01:00
|
|
|
SmallVectorImpl<SDValue>& ArgValues, DebugLoc dl);
|
2005-01-07 08:44:22 +01:00
|
|
|
|
|
|
|
/// LowerCallTo - This hook lowers an abstract call to a function into an
|
2005-01-08 20:25:39 +01:00
|
|
|
/// 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.
|
2006-12-31 06:23:18 +01:00
|
|
|
struct ArgListEntry {
|
2008-07-27 23:46:04 +02:00
|
|
|
SDValue Node;
|
2006-12-31 06:23:18 +01:00
|
|
|
const Type* Ty;
|
2008-06-30 22:33:57 +02:00
|
|
|
bool isSExt : 1;
|
|
|
|
bool isZExt : 1;
|
|
|
|
bool isInReg : 1;
|
|
|
|
bool isSRet : 1;
|
|
|
|
bool isNest : 1;
|
|
|
|
bool isByVal : 1;
|
2008-02-22 18:49:45 +01:00
|
|
|
uint16_t Alignment;
|
2007-02-01 09:39:52 +01:00
|
|
|
|
2007-07-27 14:58:54 +02:00
|
|
|
ArgListEntry() : isSExt(false), isZExt(false), isInReg(false),
|
2008-02-22 18:49:45 +01:00
|
|
|
isSRet(false), isNest(false), isByVal(false), Alignment(0) { }
|
2006-12-31 06:23:18 +01:00
|
|
|
};
|
|
|
|
typedef std::vector<ArgListEntry> ArgListTy;
|
2008-07-27 23:46:04 +02:00
|
|
|
virtual std::pair<SDValue, SDValue>
|
|
|
|
LowerCallTo(SDValue Chain, const Type *RetTy, bool RetSExt, bool RetZExt,
|
2009-07-03 08:44:53 +02:00
|
|
|
bool isVarArg, bool isInreg, unsigned NumFixedArgs,
|
|
|
|
unsigned CallingConv, bool isTailCall, SDValue Callee,
|
|
|
|
ArgListTy &Args, SelectionDAG &DAG, DebugLoc dl);
|
2007-11-06 00:12:20 +01:00
|
|
|
|
Drop ISD::MEMSET, ISD::MEMMOVE, and ISD::MEMCPY, which are not Legal
on any current target and aren't optimized in DAGCombiner. Instead
of using intermediate nodes, expand the operations, choosing between
simple loads/stores, target-specific code, and library calls,
immediately.
Previously, the code to emit optimized code for these operations
was only used at initial SelectionDAG construction time; now it is
used at all times. This fixes some cases where rep;movs was being
used for small copies where simple loads/stores would be better.
This also cleans up code that checks for alignments less than 4;
let the targets make that decision instead of doing it in
target-independent code. This allows x86 to use rep;movs in
low-alignment cases.
Also, this fixes a bug that resulted in the use of rep;stos for
memsets of 0 with non-constant memory size when the alignment was
at least 4. It's better to use the library in this case, which
can be significantly faster when the size is large.
This also preserves more SourceValue information when memory
intrinsics are lowered into simple loads/stores.
llvm-svn: 49572
2008-04-12 06:36:06 +02:00
|
|
|
/// EmitTargetCodeForMemcpy - Emit target-specific code that performs a
|
|
|
|
/// memcpy. This can be used by targets to provide code sequences for cases
|
|
|
|
/// that don't fit the target's parameters for simple loads/stores and can be
|
|
|
|
/// more efficient than using a library call. This function can return a null
|
2008-07-27 23:46:04 +02:00
|
|
|
/// SDValue if the target declines to use custom code and a different
|
Drop ISD::MEMSET, ISD::MEMMOVE, and ISD::MEMCPY, which are not Legal
on any current target and aren't optimized in DAGCombiner. Instead
of using intermediate nodes, expand the operations, choosing between
simple loads/stores, target-specific code, and library calls,
immediately.
Previously, the code to emit optimized code for these operations
was only used at initial SelectionDAG construction time; now it is
used at all times. This fixes some cases where rep;movs was being
used for small copies where simple loads/stores would be better.
This also cleans up code that checks for alignments less than 4;
let the targets make that decision instead of doing it in
target-independent code. This allows x86 to use rep;movs in
low-alignment cases.
Also, this fixes a bug that resulted in the use of rep;stos for
memsets of 0 with non-constant memory size when the alignment was
at least 4. It's better to use the library in this case, which
can be significantly faster when the size is large.
This also preserves more SourceValue information when memory
intrinsics are lowered into simple loads/stores.
llvm-svn: 49572
2008-04-12 06:36:06 +02:00
|
|
|
/// lowering strategy should be used.
|
|
|
|
///
|
|
|
|
/// If AlwaysInline is true, the size is constant and the target should not
|
|
|
|
/// emit any calls and is strongly encouraged to attempt to emit inline code
|
|
|
|
/// even if it is beyond the usual threshold because this intrinsic is being
|
|
|
|
/// expanded in a place where calls are not feasible (e.g. within the prologue
|
|
|
|
/// for another call). If the target chooses to decline an AlwaysInline
|
|
|
|
/// request here, legalize will resort to using simple loads and stores.
|
2008-07-27 23:46:04 +02:00
|
|
|
virtual SDValue
|
2009-02-03 23:26:09 +01:00
|
|
|
EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
|
2008-07-27 23:46:04 +02:00
|
|
|
SDValue Chain,
|
|
|
|
SDValue Op1, SDValue Op2,
|
|
|
|
SDValue Op3, unsigned Align,
|
Drop ISD::MEMSET, ISD::MEMMOVE, and ISD::MEMCPY, which are not Legal
on any current target and aren't optimized in DAGCombiner. Instead
of using intermediate nodes, expand the operations, choosing between
simple loads/stores, target-specific code, and library calls,
immediately.
Previously, the code to emit optimized code for these operations
was only used at initial SelectionDAG construction time; now it is
used at all times. This fixes some cases where rep;movs was being
used for small copies where simple loads/stores would be better.
This also cleans up code that checks for alignments less than 4;
let the targets make that decision instead of doing it in
target-independent code. This allows x86 to use rep;movs in
low-alignment cases.
Also, this fixes a bug that resulted in the use of rep;stos for
memsets of 0 with non-constant memory size when the alignment was
at least 4. It's better to use the library in this case, which
can be significantly faster when the size is large.
This also preserves more SourceValue information when memory
intrinsics are lowered into simple loads/stores.
llvm-svn: 49572
2008-04-12 06:36:06 +02:00
|
|
|
bool AlwaysInline,
|
2008-04-14 19:55:48 +02:00
|
|
|
const Value *DstSV, uint64_t DstOff,
|
|
|
|
const Value *SrcSV, uint64_t SrcOff) {
|
2008-07-27 23:46:04 +02:00
|
|
|
return SDValue();
|
Drop ISD::MEMSET, ISD::MEMMOVE, and ISD::MEMCPY, which are not Legal
on any current target and aren't optimized in DAGCombiner. Instead
of using intermediate nodes, expand the operations, choosing between
simple loads/stores, target-specific code, and library calls,
immediately.
Previously, the code to emit optimized code for these operations
was only used at initial SelectionDAG construction time; now it is
used at all times. This fixes some cases where rep;movs was being
used for small copies where simple loads/stores would be better.
This also cleans up code that checks for alignments less than 4;
let the targets make that decision instead of doing it in
target-independent code. This allows x86 to use rep;movs in
low-alignment cases.
Also, this fixes a bug that resulted in the use of rep;stos for
memsets of 0 with non-constant memory size when the alignment was
at least 4. It's better to use the library in this case, which
can be significantly faster when the size is large.
This also preserves more SourceValue information when memory
intrinsics are lowered into simple loads/stores.
llvm-svn: 49572
2008-04-12 06:36:06 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/// EmitTargetCodeForMemmove - Emit target-specific code that performs a
|
|
|
|
/// memmove. This can be used by targets to provide code sequences for cases
|
|
|
|
/// that don't fit the target's parameters for simple loads/stores and can be
|
|
|
|
/// more efficient than using a library call. This function can return a null
|
2008-07-27 23:46:04 +02:00
|
|
|
/// SDValue if the target declines to use custom code and a different
|
2008-04-14 19:45:20 +02:00
|
|
|
/// lowering strategy should be used.
|
2008-07-27 23:46:04 +02:00
|
|
|
virtual SDValue
|
2009-02-03 23:26:09 +01:00
|
|
|
EmitTargetCodeForMemmove(SelectionDAG &DAG, DebugLoc dl,
|
2008-07-27 23:46:04 +02:00
|
|
|
SDValue Chain,
|
|
|
|
SDValue Op1, SDValue Op2,
|
|
|
|
SDValue Op3, unsigned Align,
|
2008-04-14 19:55:48 +02:00
|
|
|
const Value *DstSV, uint64_t DstOff,
|
|
|
|
const Value *SrcSV, uint64_t SrcOff) {
|
2008-07-27 23:46:04 +02:00
|
|
|
return SDValue();
|
Drop ISD::MEMSET, ISD::MEMMOVE, and ISD::MEMCPY, which are not Legal
on any current target and aren't optimized in DAGCombiner. Instead
of using intermediate nodes, expand the operations, choosing between
simple loads/stores, target-specific code, and library calls,
immediately.
Previously, the code to emit optimized code for these operations
was only used at initial SelectionDAG construction time; now it is
used at all times. This fixes some cases where rep;movs was being
used for small copies where simple loads/stores would be better.
This also cleans up code that checks for alignments less than 4;
let the targets make that decision instead of doing it in
target-independent code. This allows x86 to use rep;movs in
low-alignment cases.
Also, this fixes a bug that resulted in the use of rep;stos for
memsets of 0 with non-constant memory size when the alignment was
at least 4. It's better to use the library in this case, which
can be significantly faster when the size is large.
This also preserves more SourceValue information when memory
intrinsics are lowered into simple loads/stores.
llvm-svn: 49572
2008-04-12 06:36:06 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
/// EmitTargetCodeForMemset - Emit target-specific code that performs a
|
|
|
|
/// memset. This can be used by targets to provide code sequences for cases
|
|
|
|
/// that don't fit the target's parameters for simple stores and can be more
|
|
|
|
/// efficient than using a library call. This function can return a null
|
2008-07-27 23:46:04 +02:00
|
|
|
/// SDValue if the target declines to use custom code and a different
|
2008-04-14 19:45:20 +02:00
|
|
|
/// lowering strategy should be used.
|
2008-07-27 23:46:04 +02:00
|
|
|
virtual SDValue
|
2009-02-03 23:26:09 +01:00
|
|
|
EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
|
2008-07-27 23:46:04 +02:00
|
|
|
SDValue Chain,
|
|
|
|
SDValue Op1, SDValue Op2,
|
|
|
|
SDValue Op3, unsigned Align,
|
2008-10-01 02:59:58 +02:00
|
|
|
const Value *DstSV, uint64_t DstOff) {
|
2008-07-27 23:46:04 +02:00
|
|
|
return SDValue();
|
2007-11-06 00:12:20 +01:00
|
|
|
}
|
|
|
|
|
2009-01-21 10:00:29 +01:00
|
|
|
/// LowerOperationWrapper - This callback is invoked by the type legalizer
|
|
|
|
/// to legalize nodes with an illegal operand type but legal result types.
|
2009-01-21 05:48:39 +01:00
|
|
|
/// It replaces the LowerOperation callback in the type Legalizer.
|
2009-01-21 10:00:29 +01:00
|
|
|
/// The reason we can not do away with LowerOperation entirely is that
|
2009-01-21 05:48:39 +01:00
|
|
|
/// LegalizeDAG isn't yet ready to use this callback.
|
2009-01-21 10:00:29 +01:00
|
|
|
/// 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 05:48:39 +01:00
|
|
|
SmallVectorImpl<SDValue> &Results,
|
|
|
|
SelectionDAG &DAG);
|
|
|
|
|
2006-01-28 04:37:03 +01:00
|
|
|
/// LowerOperation - This callback is invoked for operations that are
|
|
|
|
/// unsupported by the target, which are registered to use 'custom' lowering,
|
|
|
|
/// and whose defined values are all legal.
|
2005-01-16 08:27:49 +01:00
|
|
|
/// If the target has no operations that require custom lowering, it need not
|
|
|
|
/// implement this. The default implementation of this aborts.
|
2008-07-27 23:46:04 +02:00
|
|
|
virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG);
|
2005-01-19 04:36:03 +01:00
|
|
|
|
2008-12-01 12:39:25 +01:00
|
|
|
/// 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 05:31:45 +02:00
|
|
|
///
|
|
|
|
/// If the target has no operations that require custom lowering, it need not
|
2008-07-04 13:47:58 +02:00
|
|
|
/// implement this. The default implementation aborts.
|
2008-12-01 12:39:25 +01:00
|
|
|
virtual void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
|
|
|
|
SelectionDAG &DAG) {
|
2008-07-04 13:47:58 +02:00
|
|
|
assert(0 && "ReplaceNodeResults not implemented for this target!");
|
2007-11-24 08:07:01 +01:00
|
|
|
}
|
2008-07-04 13:47:58 +02:00
|
|
|
|
2007-10-11 21:40:01 +02:00
|
|
|
/// IsEligibleForTailCallOptimization - Check whether the call is eligible for
|
2007-10-12 23:30:57 +02:00
|
|
|
/// tail call optimization. Targets which want to do tail call optimization
|
|
|
|
/// should override this function.
|
2008-09-13 03:54:27 +02:00
|
|
|
virtual bool IsEligibleForTailCallOptimization(CallSDNode *Call,
|
2008-07-27 23:46:04 +02:00
|
|
|
SDValue Ret,
|
2007-10-11 21:40:01 +02:00
|
|
|
SelectionDAG &DAG) const {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2008-04-30 11:16:33 +02:00
|
|
|
/// CheckTailCallReturnConstraints - Check whether CALL node immediatly
|
|
|
|
/// preceeds the RET node and whether the return uses the result of the node
|
|
|
|
/// or is a void return. This function can be used by the target to determine
|
|
|
|
/// eligiblity of tail call optimization.
|
2009-03-28 09:33:27 +01:00
|
|
|
static bool CheckTailCallReturnConstraints(CallSDNode *TheCall, SDValue Ret);
|
2008-04-30 11:16:33 +02:00
|
|
|
|
|
|
|
/// GetPossiblePreceedingTailCall - Get preceeding TailCallNodeOpCode node if
|
2008-12-09 03:18:52 +01:00
|
|
|
/// it exists. Skip a possible ISD::TokenFactor.
|
2008-07-27 23:46:04 +02:00
|
|
|
static SDValue GetPossiblePreceedingTailCall(SDValue Chain,
|
2008-04-30 11:16:33 +02:00
|
|
|
unsigned TailCallNodeOpCode) {
|
|
|
|
if (Chain.getOpcode() == TailCallNodeOpCode) {
|
|
|
|
return Chain;
|
|
|
|
} else if (Chain.getOpcode() == ISD::TokenFactor) {
|
|
|
|
if (Chain.getNumOperands() &&
|
|
|
|
Chain.getOperand(0).getOpcode() == TailCallNodeOpCode)
|
|
|
|
return Chain.getOperand(0);
|
|
|
|
}
|
|
|
|
return Chain;
|
|
|
|
}
|
|
|
|
|
2005-12-22 00:15:41 +01:00
|
|
|
/// getTargetNodeName() - This method returns the name of a target specific
|
|
|
|
/// DAG node.
|
2005-12-20 07:22:03 +01:00
|
|
|
virtual const char *getTargetNodeName(unsigned Opcode) const;
|
|
|
|
|
2008-08-19 23:26:29 +02:00
|
|
|
/// createFastISel - This method returns a target specific FastISel object,
|
|
|
|
/// or null if the target does not support "fast" ISel.
|
2008-09-04 01:12:08 +02:00
|
|
|
virtual FastISel *
|
|
|
|
createFastISel(MachineFunction &,
|
2009-01-13 01:35:13 +01:00
|
|
|
MachineModuleInfo *, DwarfWriter *,
|
2008-09-04 01:12:08 +02:00
|
|
|
DenseMap<const Value *, unsigned> &,
|
2008-09-10 22:11:02 +02:00
|
|
|
DenseMap<const BasicBlock *, MachineBasicBlock *> &,
|
2008-10-15 01:54:11 +02:00
|
|
|
DenseMap<const AllocaInst *, int> &
|
|
|
|
#ifndef NDEBUG
|
|
|
|
, SmallSet<Instruction*, 8> &CatchInfoLost
|
|
|
|
#endif
|
|
|
|
) {
|
2008-09-04 01:12:08 +02:00
|
|
|
return 0;
|
|
|
|
}
|
2008-08-19 23:26:29 +02:00
|
|
|
|
2006-01-26 21:27:33 +01:00
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// Inline Asm Support hooks
|
|
|
|
//
|
|
|
|
|
2006-02-04 03:12:09 +01:00
|
|
|
enum ConstraintType {
|
2008-11-13 22:52:36 +01:00
|
|
|
C_Register, // Constraint represents specific register(s).
|
|
|
|
C_RegisterClass, // Constraint represents any of register(s) in class.
|
2006-02-24 02:10:14 +01:00
|
|
|
C_Memory, // Memory constraint.
|
2006-02-04 03:12:09 +01:00
|
|
|
C_Other, // Something else.
|
|
|
|
C_Unknown // Unsupported constraint.
|
|
|
|
};
|
|
|
|
|
2008-02-26 03:33:44 +01:00
|
|
|
/// 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 18:21:11 +02:00
|
|
|
/// TargetLowering picks the 'best' code from ConstraintInfo::Codes that
|
|
|
|
/// most closely matches the operand.
|
2008-02-26 03:33:44 +01:00
|
|
|
std::string ConstraintCode;
|
|
|
|
|
|
|
|
/// ConstraintType - Information about the constraint code, e.g. Register,
|
|
|
|
/// RegisterClass, Memory, Other, Unknown.
|
|
|
|
TargetLowering::ConstraintType ConstraintType;
|
|
|
|
|
|
|
|
/// 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;
|
|
|
|
|
|
|
|
/// ConstraintVT - The ValueType for the operand value.
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT ConstraintVT;
|
2008-10-17 18:21:11 +02:00
|
|
|
|
2008-10-17 18:47:46 +02:00
|
|
|
/// isMatchingInputConstraint - Return true of this is an input operand that
|
|
|
|
/// is a matching constraint like "4".
|
|
|
|
bool isMatchingInputConstraint() const;
|
2008-10-17 18:21:11 +02:00
|
|
|
|
|
|
|
/// getMatchedOperand - If this is an input matching constraint, this method
|
|
|
|
/// returns the output operand it matches.
|
|
|
|
unsigned getMatchedOperand() const;
|
2008-02-26 03:33:44 +01:00
|
|
|
|
|
|
|
AsmOperandInfo(const InlineAsm::ConstraintInfo &info)
|
|
|
|
: InlineAsm::ConstraintInfo(info),
|
|
|
|
ConstraintType(TargetLowering::C_Unknown),
|
|
|
|
CallOperandVal(0), ConstraintVT(MVT::Other) {
|
|
|
|
}
|
|
|
|
};
|
|
|
|
|
2008-04-27 02:09:47 +02:00
|
|
|
/// ComputeConstraintToUse - Determines the constraint code and constraint
|
|
|
|
/// type to use for the specific AsmOperandInfo, setting
|
2008-04-27 02:37:18 +02:00
|
|
|
/// OpInfo.ConstraintCode and OpInfo.ConstraintType. If the actual operand
|
|
|
|
/// being passed in is available, it can be passed in as Op, otherwise an
|
2008-09-24 02:05:32 +02:00
|
|
|
/// empty SDValue can be passed. If hasMemory is true it means one of the asm
|
|
|
|
/// constraint of the inline asm instruction being processed is 'm'.
|
2008-04-27 02:37:18 +02:00
|
|
|
virtual void ComputeConstraintToUse(AsmOperandInfo &OpInfo,
|
2008-07-27 23:46:04 +02:00
|
|
|
SDValue Op,
|
2008-09-24 02:05:32 +02:00
|
|
|
bool hasMemory,
|
2008-04-27 02:37:18 +02:00
|
|
|
SelectionDAG *DAG = 0) const;
|
2008-04-27 02:09:47 +02:00
|
|
|
|
2007-03-25 04:14:49 +02:00
|
|
|
/// getConstraintType - Given a constraint, return the type of constraint it
|
|
|
|
/// is for this target.
|
|
|
|
virtual ConstraintType getConstraintType(const std::string &Constraint) const;
|
2006-02-04 03:12:09 +01:00
|
|
|
|
2006-02-22 01:56:39 +01:00
|
|
|
/// getRegClassForInlineAsmConstraint - Given a constraint letter (e.g. "r"),
|
|
|
|
/// return a list of registers that can be used to satisfy the constraint.
|
|
|
|
/// This should only be used for C_RegisterClass constraints.
|
2006-01-26 21:27:33 +01:00
|
|
|
virtual std::vector<unsigned>
|
2006-02-22 01:56:39 +01:00
|
|
|
getRegClassForInlineAsmConstraint(const std::string &Constraint,
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT VT) const;
|
2006-02-22 01:56:39 +01:00
|
|
|
|
|
|
|
/// getRegForInlineAsmConstraint - Given a physical register constraint (e.g.
|
|
|
|
/// {edx}), return the register number and the register class for the
|
2006-11-02 02:39:10 +01:00
|
|
|
/// 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..
|
2006-02-22 01:56:39 +01:00
|
|
|
virtual std::pair<unsigned, const TargetRegisterClass*>
|
|
|
|
getRegForInlineAsmConstraint(const std::string &Constraint,
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT VT) const;
|
2006-02-22 01:56:39 +01:00
|
|
|
|
2008-01-29 03:21:21 +01:00
|
|
|
/// LowerXConstraint - try to replace an X constraint, which matches anything,
|
|
|
|
/// with another that has more specific requirements based on the type of the
|
2008-04-27 01:02:14 +02:00
|
|
|
/// corresponding operand. This returns null if there is no replacement to
|
|
|
|
/// make.
|
2008-06-06 14:08:01 +02:00
|
|
|
virtual const char *LowerXConstraint(MVT ConstraintVT) const;
|
2006-01-26 21:27:33 +01:00
|
|
|
|
2007-08-25 02:47:38 +02:00
|
|
|
/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
|
2008-09-24 02:05:32 +02:00
|
|
|
/// vector. If it is invalid, don't add anything to Ops. If hasMemory is true
|
|
|
|
/// it means one of the asm constraint of the inline asm instruction being
|
|
|
|
/// processed is 'm'.
|
2008-07-27 23:46:04 +02:00
|
|
|
virtual void LowerAsmOperandForConstraint(SDValue Op, char ConstraintLetter,
|
2008-09-24 02:05:32 +02:00
|
|
|
bool hasMemory,
|
2008-07-27 23:46:04 +02:00
|
|
|
std::vector<SDValue> &Ops,
|
2008-04-27 01:02:14 +02:00
|
|
|
SelectionDAG &DAG) const;
|
2006-02-04 03:12:09 +01:00
|
|
|
|
2005-08-26 22:53:09 +02:00
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// Scheduler hooks
|
|
|
|
//
|
|
|
|
|
2008-01-30 19:18:23 +01:00
|
|
|
// EmitInstrWithCustomInserter - This method should be implemented by targets
|
|
|
|
// that mark instructions with the 'usesCustomDAGSchedInserter' flag. These
|
2005-08-26 22:53:09 +02:00
|
|
|
// instructions are special in various ways, which require special support to
|
|
|
|
// insert. The specified MachineInstr is created but not inserted into any
|
|
|
|
// basic blocks, and the scheduler passes ownership of it to this method.
|
2008-01-30 19:18:23 +01:00
|
|
|
virtual MachineBasicBlock *EmitInstrWithCustomInserter(MachineInstr *MI,
|
2009-02-07 17:15:20 +01:00
|
|
|
MachineBasicBlock *MBB) const;
|
2005-04-21 22:59:05 +02:00
|
|
|
|
2006-03-16 22:47:42 +01:00
|
|
|
//===--------------------------------------------------------------------===//
|
2007-03-31 01:14:50 +02:00
|
|
|
// Addressing mode description hooks (used by LSR etc).
|
2006-03-16 22:47:42 +01:00
|
|
|
//
|
2007-03-16 09:42:32 +01:00
|
|
|
|
2007-03-31 01:14:50 +02:00
|
|
|
/// AddrMode - This represents an addressing mode of:
|
|
|
|
/// BaseGV + BaseOffs + BaseReg + Scale*ScaleReg
|
|
|
|
/// If BaseGV is null, there is no BaseGV.
|
|
|
|
/// If BaseOffs is zero, there is no base offset.
|
|
|
|
/// If HasBaseReg is false, there is no base register.
|
|
|
|
/// If Scale is zero, there is no ScaleReg. Scale of 1 indicates a reg with
|
|
|
|
/// no scale.
|
|
|
|
///
|
|
|
|
struct AddrMode {
|
|
|
|
GlobalValue *BaseGV;
|
|
|
|
int64_t BaseOffs;
|
|
|
|
bool HasBaseReg;
|
|
|
|
int64_t Scale;
|
2007-04-09 23:18:34 +02:00
|
|
|
AddrMode() : BaseGV(0), BaseOffs(0), HasBaseReg(false), Scale(0) {}
|
2007-03-31 01:14:50 +02:00
|
|
|
};
|
|
|
|
|
|
|
|
/// 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 21:11:53 +02:00
|
|
|
/// 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.
|
2007-03-31 01:14:50 +02:00
|
|
|
/// TODO: Handle pre/postinc as well.
|
|
|
|
virtual bool isLegalAddressingMode(const AddrMode &AM, const Type *Ty) const;
|
2006-03-16 22:47:42 +01:00
|
|
|
|
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
llvm-svn: 43375
2007-10-26 03:56:11 +02:00
|
|
|
/// 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.
|
|
|
|
virtual bool isTruncateFree(const Type *Ty1, const Type *Ty2) const {
|
|
|
|
return false;
|
|
|
|
}
|
2007-10-29 20:58:20 +01:00
|
|
|
|
2008-06-06 14:08:01 +02:00
|
|
|
virtual bool isTruncateFree(MVT VT1, MVT VT2) const {
|
2007-10-29 20:58:20 +01:00
|
|
|
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.
llvm-svn: 68576
2009-04-08 02:15:30 +02:00
|
|
|
|
|
|
|
/// isZExtFree - Return true if any actual instruction that defines a
|
|
|
|
/// value of type Ty1 implicit zero-extends the value to Ty2 in the result
|
|
|
|
/// 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.
|
|
|
|
virtual bool isZExtFree(const Type *Ty1, const Type *Ty2) const {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
virtual bool isZExtFree(MVT VT1, MVT VT2) const {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2009-05-28 02:35:15 +02:00
|
|
|
/// 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.
|
|
|
|
virtual bool isNarrowingProfitable(MVT VT1, MVT VT2) const {
|
2009-05-29 07:31:40 +02:00
|
|
|
return false;
|
2009-05-28 02:35:15 +02:00
|
|
|
}
|
|
|
|
|
2006-05-16 19:42:15 +02:00
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// Div utility functions
|
|
|
|
//
|
2008-07-27 23:46:04 +02:00
|
|
|
SDValue BuildSDIV(SDNode *N, SelectionDAG &DAG,
|
2007-04-16 20:10:23 +02:00
|
|
|
std::vector<SDNode*>* Created) const;
|
2008-07-27 23:46:04 +02:00
|
|
|
SDValue BuildUDIV(SDNode *N, SelectionDAG &DAG,
|
2007-04-16 20:10:23 +02:00
|
|
|
std::vector<SDNode*>* Created) const;
|
2006-05-16 19:42:15 +02:00
|
|
|
|
|
|
|
|
2007-01-12 03:11:51 +01:00
|
|
|
//===--------------------------------------------------------------------===//
|
|
|
|
// Runtime Library hooks
|
|
|
|
//
|
|
|
|
|
|
|
|
/// setLibcallName - Rename the default libcall routine name for the specified
|
|
|
|
/// libcall.
|
2007-01-12 23:49:32 +01:00
|
|
|
void setLibcallName(RTLIB::Libcall Call, const char *Name) {
|
2007-01-12 03:11:51 +01:00
|
|
|
LibcallRoutineNames[Call] = Name;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// getLibcallName - Get the libcall routine name for the specified libcall.
|
|
|
|
///
|
|
|
|
const char *getLibcallName(RTLIB::Libcall Call) const {
|
2007-01-12 23:49:32 +01:00
|
|
|
return LibcallRoutineNames[Call];
|
2007-01-12 03:11:51 +01:00
|
|
|
}
|
|
|
|
|
2007-01-31 10:29:11 +01:00
|
|
|
/// 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];
|
|
|
|
}
|
|
|
|
|
2006-03-16 22:47:42 +01:00
|
|
|
private:
|
2005-01-19 04:36:03 +01:00
|
|
|
TargetMachine &TM;
|
2006-05-03 03:29:57 +02:00
|
|
|
const TargetData *TD;
|
2005-01-19 04:36:03 +01:00
|
|
|
|
|
|
|
/// PointerTy - The type to use for pointers, usually i32 or i64.
|
|
|
|
///
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT PointerTy;
|
2005-01-19 04:36:03 +01:00
|
|
|
|
2008-06-30 22:36:26 +02:00
|
|
|
/// IsLittleEndian - True if this is a little endian target.
|
|
|
|
///
|
|
|
|
bool IsLittleEndian;
|
|
|
|
|
2006-10-07 00:46:34 +02:00
|
|
|
/// UsesGlobalOffsetTable - True if this target uses a GOT for PIC codegen.
|
|
|
|
///
|
|
|
|
bool UsesGlobalOffsetTable;
|
|
|
|
|
2007-01-06 00:31:08 +01:00
|
|
|
/// SelectIsExpensive - Tells the code generator not to expand operations
|
|
|
|
/// into sequences that use the select operations if possible.
|
|
|
|
bool SelectIsExpensive;
|
2005-05-12 04:02:21 +02:00
|
|
|
|
2005-10-21 02:02:42 +02:00
|
|
|
/// 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;
|
|
|
|
|
|
|
|
/// 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;
|
2005-10-20 04:14:14 +02:00
|
|
|
|
2006-12-11 00:12:42 +01:00
|
|
|
/// 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;
|
|
|
|
|
2008-06-30 22:36:26 +02:00
|
|
|
/// ShiftAmountTy - The type to use for shift amounts, usually i8 or whatever
|
|
|
|
/// PointerTy is.
|
|
|
|
MVT ShiftAmountTy;
|
|
|
|
|
|
|
|
OutOfRangeShiftAmount ShiftAmtHandling;
|
|
|
|
|
2008-11-23 16:47:28 +01:00
|
|
|
/// BooleanContents - Information about the contents of the high-bits in
|
|
|
|
/// boolean values held in a type wider than i1. See getBooleanContents.
|
|
|
|
BooleanContent BooleanContents;
|
2008-06-30 22:36:26 +02:00
|
|
|
|
|
|
|
/// SchedPreferenceInfo - The target scheduling preference: shortest possible
|
|
|
|
/// total cycles or lowest register usage.
|
|
|
|
SchedPreference SchedPreferenceInfo;
|
|
|
|
|
2006-09-04 08:21:35 +02:00
|
|
|
/// JumpBufSize - The size, in bytes, of the target's jmp_buf buffers
|
|
|
|
unsigned JumpBufSize;
|
|
|
|
|
|
|
|
/// JumpBufAlignment - The alignment, in bytes, of the target's jmp_buf
|
|
|
|
/// buffers
|
|
|
|
unsigned JumpBufAlignment;
|
2007-05-17 01:45:53 +02:00
|
|
|
|
2007-06-01 10:25:24 +02:00
|
|
|
/// IfCvtBlockSizeLimit - The maximum allowed size for a block to be
|
|
|
|
/// if-converted.
|
2007-05-17 01:45:53 +02:00
|
|
|
unsigned IfCvtBlockSizeLimit;
|
2006-09-04 08:21:35 +02:00
|
|
|
|
2007-06-01 10:25:24 +02:00
|
|
|
/// IfCvtDupBlockSizeLimit - The maximum allowed size for a block to be
|
|
|
|
/// duplicated during if-conversion.
|
|
|
|
unsigned IfCvtDupBlockSizeLimit;
|
|
|
|
|
2008-02-28 01:43:03 +01:00
|
|
|
/// PrefLoopAlignment - The perferred loop alignment.
|
|
|
|
///
|
|
|
|
unsigned PrefLoopAlignment;
|
|
|
|
|
2006-01-13 18:47:52 +01:00
|
|
|
/// 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 21:41:18 +02:00
|
|
|
|
2007-02-22 16:37:19 +01:00
|
|
|
/// 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 04:36:03 +01:00
|
|
|
/// RegClassForVT - This indicates the default register class to use for
|
|
|
|
/// each ValueType the target supports natively.
|
|
|
|
TargetRegisterClass *RegClassForVT[MVT::LAST_VALUETYPE];
|
2007-06-21 16:42:22 +02:00
|
|
|
unsigned char NumRegistersForVT[MVT::LAST_VALUETYPE];
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT RegisterTypeForVT[MVT::LAST_VALUETYPE];
|
2005-01-19 04:36:03 +01:00
|
|
|
|
|
|
|
/// 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).
|
2008-06-06 14:08:01 +02:00
|
|
|
MVT TransformToType[MVT::LAST_VALUETYPE];
|
2005-01-19 04:36:03 +01:00
|
|
|
|
|
|
|
/// 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.
|
2009-06-25 00:22:02 +02:00
|
|
|
/// This array is accessed using VT.getSimpleVT(), so it is subject to
|
|
|
|
/// the MVT::MAX_ALLOWED_VALUETYPE * 2 bits.
|
|
|
|
uint64_t OpActions[MVT::MAX_ALLOWED_VALUETYPE/(sizeof(uint64_t)*4)][ISD::BUILTIN_OP_END];
|
2006-01-29 09:40:37 +01:00
|
|
|
|
2008-10-14 23:26:46 +02:00
|
|
|
/// LoadExtActions - For each load of load extension type and each value type,
|
2006-10-04 02:56:09 +02:00
|
|
|
/// keep a LegalizeAction that indicates how instruction selection should deal
|
|
|
|
/// with the load.
|
2008-10-14 23:26:46 +02:00
|
|
|
uint64_t LoadExtActions[ISD::LAST_LOADEXT_TYPE];
|
2006-10-04 02:56:09 +02:00
|
|
|
|
2008-01-17 20:59:44 +01:00
|
|
|
/// TruncStoreActions - For each truncating store, keep a LegalizeAction that
|
|
|
|
/// indicates how instruction selection should deal with the store.
|
|
|
|
uint64_t TruncStoreActions[MVT::LAST_VALUETYPE];
|
2006-11-09 18:55:04 +01:00
|
|
|
|
2009-06-09 02:19:10 +02:00
|
|
|
/// IndexedModeActions - For each indexed mode and each value type,
|
|
|
|
/// keep a pair of LegalizeAction that indicates how instruction
|
|
|
|
/// selection should deal with the load / store. The first
|
|
|
|
/// dimension is now the value_type for the reference. The second
|
|
|
|
/// dimension is the load [0] vs. store[1]. The third dimension
|
|
|
|
/// represents the various modes for load store.
|
|
|
|
uint8_t IndexedModeActions[MVT::LAST_VALUETYPE][2][ISD::LAST_INDEXED_MODE];
|
2006-11-09 19:56:43 +01:00
|
|
|
|
2007-08-09 03:04:01 +02:00
|
|
|
/// ConvertActions - For each conversion from source type to destination type,
|
|
|
|
/// keep a LegalizeAction that indicates how instruction selection should
|
|
|
|
/// deal with the conversion.
|
|
|
|
/// Currently, this is used only for floating->floating conversions
|
|
|
|
/// (FP_EXTEND and FP_ROUND).
|
|
|
|
uint64_t ConvertActions[MVT::LAST_VALUETYPE];
|
|
|
|
|
2008-10-15 04:05:31 +02:00
|
|
|
/// CondCodeActions - For each condition code (ISD::CondCode) keep a
|
|
|
|
/// LegalizeAction that indicates how instruction selection should
|
|
|
|
/// deal with the condition code.
|
|
|
|
uint64_t CondCodeActions[ISD::SETCC_INVALID];
|
|
|
|
|
2006-01-29 09:40:37 +01:00
|
|
|
ValueTypeActionImpl ValueTypeActions;
|
2005-04-21 22:59:05 +02:00
|
|
|
|
2007-08-30 02:23:21 +02:00
|
|
|
std::vector<APFloat> LegalFPImmediates;
|
2005-04-21 22:59:05 +02:00
|
|
|
|
2008-06-06 14:08:01 +02:00
|
|
|
std::vector<std::pair<MVT, TargetRegisterClass*> > AvailableRegClasses;
|
2005-07-19 06:52:44 +02:00
|
|
|
|
2006-03-01 05:52:55 +01:00
|
|
|
/// 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 23:50:23 +02:00
|
|
|
unsigned char
|
2009-01-05 20:40:39 +01:00
|
|
|
TargetDAGCombineArray[(ISD::BUILTIN_OP_END+CHAR_BIT-1)/CHAR_BIT];
|
2006-03-01 05:52:55 +01:00
|
|
|
|
2006-04-04 02:25:10 +02:00
|
|
|
/// 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(..).
|
2008-06-08 22:54:56 +02:00
|
|
|
std::map<std::pair<unsigned, MVT::SimpleValueType>, MVT::SimpleValueType>
|
|
|
|
PromoteToType;
|
2007-01-12 03:11:51 +01:00
|
|
|
|
|
|
|
/// LibcallRoutineNames - Stores the name each libcall.
|
|
|
|
///
|
2007-01-12 23:49:32 +01:00
|
|
|
const char *LibcallRoutineNames[RTLIB::UNKNOWN_LIBCALL];
|
2007-01-12 03:11:51 +01:00
|
|
|
|
2007-01-31 10:29:11 +01:00
|
|
|
/// 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];
|
|
|
|
|
2005-07-19 06:52:44 +02:00
|
|
|
protected:
|
2009-03-03 03:55:14 +01:00
|
|
|
/// When lowering \@llvm.memset this field specifies the maximum number of
|
2005-07-19 06:52:44 +02:00
|
|
|
/// 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 07:53:44 +02:00
|
|
|
/// with 16-bit alignment would result in four 2-byte stores and one 1-byte
|
2005-07-19 06:52:44 +02:00
|
|
|
/// 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 09:38:30 +01:00
|
|
|
unsigned maxStoresPerMemset;
|
2005-07-19 06:52:44 +02:00
|
|
|
|
2009-03-03 03:55:14 +01:00
|
|
|
/// When lowering \@llvm.memcpy this field specifies the maximum number of
|
2005-07-19 06:52:44 +02:00
|
|
|
/// 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 09:38:30 +01:00
|
|
|
unsigned maxStoresPerMemcpy;
|
2005-07-19 06:52:44 +02:00
|
|
|
|
2009-03-03 03:55:14 +01:00
|
|
|
/// When lowering \@llvm.memmove this field specifies the maximum number of
|
2005-07-19 06:52:44 +02:00
|
|
|
/// 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 07:53:44 +02:00
|
|
|
/// 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 06:52:44 +02:00
|
|
|
/// applies to copying a constant array of constant size.
|
|
|
|
/// @brief Specify maximum bytes of store instructions per memmove call.
|
2006-02-14 09:38:30 +01:00
|
|
|
unsigned maxStoresPerMemmove;
|
2005-07-19 06:52:44 +02:00
|
|
|
|
2005-08-27 21:09:02 +02:00
|
|
|
/// This field specifies whether the target machine permits unaligned memory
|
|
|
|
/// accesses. This is used, for example, to determine the size of store
|
|
|
|
/// operations when copying small arrays and other similar tasks.
|
|
|
|
/// @brief Indicate whether the target permits unaligned memory accesses.
|
|
|
|
bool allowUnalignedMemoryAccesses;
|
2009-05-13 23:42:09 +02:00
|
|
|
|
|
|
|
/// This field specifies whether the target can benefit from code placement
|
|
|
|
/// optimization.
|
|
|
|
bool benefitFromCodePlacementOpt;
|
2005-01-07 08:44:22 +01:00
|
|
|
};
|
|
|
|
} // end llvm namespace
|
|
|
|
|
|
|
|
#endif
|