2007-12-04 22:23:35 +00:00
|
|
|
//===-- SPUISelLowering.cpp - Cell SPU DAG Lowering Implementation --------===//
|
|
|
|
//
|
|
|
|
// The LLVM Compiler Infrastructure
|
|
|
|
//
|
2007-12-29 20:36:04 +00:00
|
|
|
// This file is distributed under the University of Illinois Open Source
|
|
|
|
// License. See LICENSE.TXT for details.
|
2007-12-04 22:23:35 +00:00
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
//
|
|
|
|
// This file implements the SPUTargetLowering class.
|
|
|
|
//
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#include "SPURegisterNames.h"
|
|
|
|
#include "SPUISelLowering.h"
|
|
|
|
#include "SPUTargetMachine.h"
|
2008-04-30 00:30:08 +00:00
|
|
|
#include "SPUFrameInfo.h"
|
2007-12-04 22:23:35 +00:00
|
|
|
#include "llvm/ADT/VectorExtras.h"
|
|
|
|
#include "llvm/CodeGen/CallingConvLower.h"
|
|
|
|
#include "llvm/CodeGen/MachineFrameInfo.h"
|
|
|
|
#include "llvm/CodeGen/MachineFunction.h"
|
|
|
|
#include "llvm/CodeGen/MachineInstrBuilder.h"
|
2007-12-31 04:13:23 +00:00
|
|
|
#include "llvm/CodeGen/MachineRegisterInfo.h"
|
2007-12-04 22:23:35 +00:00
|
|
|
#include "llvm/CodeGen/SelectionDAG.h"
|
|
|
|
#include "llvm/Constants.h"
|
|
|
|
#include "llvm/Function.h"
|
|
|
|
#include "llvm/Intrinsics.h"
|
|
|
|
#include "llvm/Support/Debug.h"
|
|
|
|
#include "llvm/Support/MathExtras.h"
|
|
|
|
#include "llvm/Target/TargetOptions.h"
|
|
|
|
|
|
|
|
#include <map>
|
|
|
|
|
|
|
|
using namespace llvm;
|
|
|
|
|
|
|
|
// Used in getTargetNodeName() below
|
|
|
|
namespace {
|
|
|
|
std::map<unsigned, const char *> node_names;
|
|
|
|
|
2008-06-06 12:08:01 +00:00
|
|
|
//! MVT mapping to useful data for Cell SPU
|
2007-12-04 22:23:35 +00:00
|
|
|
struct valtype_map_s {
|
2008-06-06 12:08:01 +00:00
|
|
|
const MVT valtype;
|
2008-01-30 02:55:46 +00:00
|
|
|
const int prefslot_byte;
|
2007-12-04 22:23:35 +00:00
|
|
|
};
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
const valtype_map_s valtype_map[] = {
|
|
|
|
{ MVT::i1, 3 },
|
|
|
|
{ MVT::i8, 3 },
|
|
|
|
{ MVT::i16, 2 },
|
|
|
|
{ MVT::i32, 0 },
|
|
|
|
{ MVT::f32, 0 },
|
|
|
|
{ MVT::i64, 0 },
|
|
|
|
{ MVT::f64, 0 },
|
|
|
|
{ MVT::i128, 0 }
|
|
|
|
};
|
|
|
|
|
|
|
|
const size_t n_valtype_map = sizeof(valtype_map) / sizeof(valtype_map[0]);
|
|
|
|
|
2008-06-06 12:08:01 +00:00
|
|
|
const valtype_map_s *getValueTypeMapEntry(MVT VT) {
|
2007-12-04 22:23:35 +00:00
|
|
|
const valtype_map_s *retval = 0;
|
|
|
|
|
|
|
|
for (size_t i = 0; i < n_valtype_map; ++i) {
|
|
|
|
if (valtype_map[i].valtype == VT) {
|
2008-01-30 02:55:46 +00:00
|
|
|
retval = valtype_map + i;
|
|
|
|
break;
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
#ifndef NDEBUG
|
|
|
|
if (retval == 0) {
|
|
|
|
cerr << "getValueTypeMapEntry returns NULL for "
|
2008-06-06 12:08:01 +00:00
|
|
|
<< VT.getMVTString()
|
2008-01-30 02:55:46 +00:00
|
|
|
<< "\n";
|
2007-12-04 22:23:35 +00:00
|
|
|
abort();
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
return retval;
|
|
|
|
}
|
|
|
|
|
|
|
|
//! Predicate that returns true if operand is a memory target
|
|
|
|
/*!
|
|
|
|
\arg Op Operand to test
|
|
|
|
\return true if the operand is a memory target (i.e., global
|
2008-01-11 02:53:15 +00:00
|
|
|
address, external symbol, constant pool) or an A-form
|
2007-12-04 22:23:35 +00:00
|
|
|
address.
|
|
|
|
*/
|
2008-07-27 21:46:04 +00:00
|
|
|
bool isMemoryOperand(const SDValue &Op)
|
2007-12-04 22:23:35 +00:00
|
|
|
{
|
|
|
|
const unsigned Opc = Op.getOpcode();
|
|
|
|
return (Opc == ISD::GlobalAddress
|
|
|
|
|| Opc == ISD::GlobalTLSAddress
|
|
|
|
|| Opc == ISD::JumpTable
|
|
|
|
|| Opc == ISD::ConstantPool
|
|
|
|
|| Opc == ISD::ExternalSymbol
|
|
|
|
|| Opc == ISD::TargetGlobalAddress
|
|
|
|
|| Opc == ISD::TargetGlobalTLSAddress
|
|
|
|
|| Opc == ISD::TargetJumpTable
|
|
|
|
|| Opc == ISD::TargetConstantPool
|
|
|
|
|| Opc == ISD::TargetExternalSymbol
|
2008-01-11 02:53:15 +00:00
|
|
|
|| Opc == SPUISD::AFormAddr);
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
2008-01-17 20:38:41 +00:00
|
|
|
|
|
|
|
//! Predicate that returns true if the operand is an indirect target
|
2008-07-27 21:46:04 +00:00
|
|
|
bool isIndirectOperand(const SDValue &Op)
|
2008-01-17 20:38:41 +00:00
|
|
|
{
|
|
|
|
const unsigned Opc = Op.getOpcode();
|
|
|
|
return (Opc == ISD::Register
|
2008-01-30 02:55:46 +00:00
|
|
|
|| Opc == SPUISD::LDRESULT);
|
2008-01-17 20:38:41 +00:00
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
SPUTargetLowering::SPUTargetLowering(SPUTargetMachine &TM)
|
|
|
|
: TargetLowering(TM),
|
|
|
|
SPUTM(TM)
|
|
|
|
{
|
|
|
|
// Fold away setcc operations if possible.
|
|
|
|
setPow2DivIsCheap();
|
|
|
|
|
|
|
|
// Use _setjmp/_longjmp instead of setjmp/longjmp.
|
|
|
|
setUseUnderscoreSetJmp(true);
|
|
|
|
setUseUnderscoreLongJmp(true);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Set up the SPU's register classes:
|
2007-12-17 22:32:34 +00:00
|
|
|
addRegisterClass(MVT::i8, SPU::R8CRegisterClass);
|
|
|
|
addRegisterClass(MVT::i16, SPU::R16CRegisterClass);
|
|
|
|
addRegisterClass(MVT::i32, SPU::R32CRegisterClass);
|
|
|
|
addRegisterClass(MVT::i64, SPU::R64CRegisterClass);
|
|
|
|
addRegisterClass(MVT::f32, SPU::R32FPRegisterClass);
|
|
|
|
addRegisterClass(MVT::f64, SPU::R64FPRegisterClass);
|
2007-12-04 22:23:35 +00:00
|
|
|
addRegisterClass(MVT::i128, SPU::GPRCRegisterClass);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// SPU has no sign or zero extended loads for i1, i8, i16:
|
2008-01-17 20:38:41 +00:00
|
|
|
setLoadXAction(ISD::EXTLOAD, MVT::i1, Promote);
|
2007-12-04 22:23:35 +00:00
|
|
|
setLoadXAction(ISD::SEXTLOAD, MVT::i1, Promote);
|
|
|
|
setLoadXAction(ISD::ZEXTLOAD, MVT::i1, Promote);
|
2008-01-17 19:59:44 +00:00
|
|
|
setTruncStoreAction(MVT::i8, MVT::i1, Custom);
|
|
|
|
setTruncStoreAction(MVT::i16, MVT::i1, Custom);
|
|
|
|
setTruncStoreAction(MVT::i32, MVT::i1, Custom);
|
|
|
|
setTruncStoreAction(MVT::i64, MVT::i1, Custom);
|
|
|
|
setTruncStoreAction(MVT::i128, MVT::i1, Custom);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
setLoadXAction(ISD::EXTLOAD, MVT::i8, Custom);
|
|
|
|
setLoadXAction(ISD::SEXTLOAD, MVT::i8, Custom);
|
|
|
|
setLoadXAction(ISD::ZEXTLOAD, MVT::i8, Custom);
|
2008-01-17 19:59:44 +00:00
|
|
|
setTruncStoreAction(MVT::i8 , MVT::i8, Custom);
|
|
|
|
setTruncStoreAction(MVT::i16 , MVT::i8, Custom);
|
|
|
|
setTruncStoreAction(MVT::i32 , MVT::i8, Custom);
|
|
|
|
setTruncStoreAction(MVT::i64 , MVT::i8, Custom);
|
|
|
|
setTruncStoreAction(MVT::i128, MVT::i8, Custom);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
setLoadXAction(ISD::EXTLOAD, MVT::i16, Custom);
|
|
|
|
setLoadXAction(ISD::SEXTLOAD, MVT::i16, Custom);
|
|
|
|
setLoadXAction(ISD::ZEXTLOAD, MVT::i16, Custom);
|
|
|
|
|
|
|
|
// SPU constant load actions are custom lowered:
|
|
|
|
setOperationAction(ISD::Constant, MVT::i64, Custom);
|
2008-02-14 18:43:04 +00:00
|
|
|
setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
|
2007-12-04 22:23:35 +00:00
|
|
|
setOperationAction(ISD::ConstantFP, MVT::f64, Custom);
|
|
|
|
|
|
|
|
// SPU's loads and stores have to be custom lowered:
|
|
|
|
for (unsigned sctype = (unsigned) MVT::i1; sctype < (unsigned) MVT::f128;
|
|
|
|
++sctype) {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = (MVT::SimpleValueType)sctype;
|
|
|
|
|
|
|
|
setOperationAction(ISD::LOAD, VT, Custom);
|
|
|
|
setOperationAction(ISD::STORE, VT, Custom);
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-01-17 20:38:41 +00:00
|
|
|
// Custom lower BRCOND for i1, i8 to "promote" the result to
|
|
|
|
// i32 and i16, respectively.
|
|
|
|
setOperationAction(ISD::BRCOND, MVT::Other, Custom);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
// Expand the jumptable branches
|
|
|
|
setOperationAction(ISD::BR_JT, MVT::Other, Expand);
|
|
|
|
setOperationAction(ISD::BR_CC, MVT::Other, Expand);
|
2008-07-16 17:17:29 +00:00
|
|
|
setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
// SPU has no intrinsics for these particular operations:
|
2008-02-16 14:46:26 +00:00
|
|
|
setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
|
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// PowerPC has no SREM/UREM instructions
|
|
|
|
setOperationAction(ISD::SREM, MVT::i32, Expand);
|
|
|
|
setOperationAction(ISD::UREM, MVT::i32, Expand);
|
|
|
|
setOperationAction(ISD::SREM, MVT::i64, Expand);
|
|
|
|
setOperationAction(ISD::UREM, MVT::i64, Expand);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// We don't support sin/cos/sqrt/fmod
|
|
|
|
setOperationAction(ISD::FSIN , MVT::f64, Expand);
|
|
|
|
setOperationAction(ISD::FCOS , MVT::f64, Expand);
|
|
|
|
setOperationAction(ISD::FREM , MVT::f64, Expand);
|
|
|
|
setOperationAction(ISD::FSIN , MVT::f32, Expand);
|
|
|
|
setOperationAction(ISD::FCOS , MVT::f32, Expand);
|
|
|
|
setOperationAction(ISD::FREM , MVT::f32, Expand);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// If we're enabling GP optimizations, use hardware square root
|
|
|
|
setOperationAction(ISD::FSQRT, MVT::f64, Expand);
|
|
|
|
setOperationAction(ISD::FSQRT, MVT::f32, Expand);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
|
|
|
|
setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
|
|
|
|
|
|
|
|
// SPU can do rotate right and left, so legalize it... but customize for i8
|
|
|
|
// because instructions don't exist.
|
|
|
|
setOperationAction(ISD::ROTR, MVT::i32, Legal);
|
|
|
|
setOperationAction(ISD::ROTR, MVT::i16, Legal);
|
|
|
|
setOperationAction(ISD::ROTR, MVT::i8, Custom);
|
|
|
|
setOperationAction(ISD::ROTL, MVT::i32, Legal);
|
|
|
|
setOperationAction(ISD::ROTL, MVT::i16, Legal);
|
|
|
|
setOperationAction(ISD::ROTL, MVT::i8, Custom);
|
|
|
|
// SPU has no native version of shift left/right for i8
|
|
|
|
setOperationAction(ISD::SHL, MVT::i8, Custom);
|
|
|
|
setOperationAction(ISD::SRL, MVT::i8, Custom);
|
|
|
|
setOperationAction(ISD::SRA, MVT::i8, Custom);
|
2008-02-23 18:41:37 +00:00
|
|
|
// And SPU needs custom lowering for shift left/right for i64
|
|
|
|
setOperationAction(ISD::SHL, MVT::i64, Custom);
|
|
|
|
setOperationAction(ISD::SRL, MVT::i64, Custom);
|
|
|
|
setOperationAction(ISD::SRA, MVT::i64, Custom);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-16 17:17:29 +00:00
|
|
|
// Custom lower i8, i32 and i64 multiplications
|
|
|
|
setOperationAction(ISD::MUL, MVT::i8, Custom);
|
2007-12-04 22:23:35 +00:00
|
|
|
setOperationAction(ISD::MUL, MVT::i32, Custom);
|
2008-07-16 17:17:29 +00:00
|
|
|
setOperationAction(ISD::MUL, MVT::i64, Custom);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-06-02 22:18:03 +00:00
|
|
|
// Need to custom handle (some) common i8, i64 math ops
|
|
|
|
setOperationAction(ISD::ADD, MVT::i64, Custom);
|
2007-12-04 22:23:35 +00:00
|
|
|
setOperationAction(ISD::SUB, MVT::i8, Custom);
|
2008-06-02 22:18:03 +00:00
|
|
|
setOperationAction(ISD::SUB, MVT::i64, Custom);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// SPU does not have BSWAP. It does have i32 support CTLZ.
|
|
|
|
// CTPOP has to be custom lowered.
|
|
|
|
setOperationAction(ISD::BSWAP, MVT::i32, Expand);
|
|
|
|
setOperationAction(ISD::BSWAP, MVT::i64, Expand);
|
|
|
|
|
|
|
|
setOperationAction(ISD::CTPOP, MVT::i8, Custom);
|
|
|
|
setOperationAction(ISD::CTPOP, MVT::i16, Custom);
|
|
|
|
setOperationAction(ISD::CTPOP, MVT::i32, Custom);
|
|
|
|
setOperationAction(ISD::CTPOP, MVT::i64, Custom);
|
|
|
|
|
|
|
|
setOperationAction(ISD::CTTZ , MVT::i32, Expand);
|
|
|
|
setOperationAction(ISD::CTTZ , MVT::i64, Expand);
|
|
|
|
|
|
|
|
setOperationAction(ISD::CTLZ , MVT::i32, Legal);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-06-02 22:18:03 +00:00
|
|
|
// SPU has a version of select that implements (a&~c)|(b&c), just like
|
2008-03-10 23:49:09 +00:00
|
|
|
// select ought to work:
|
2008-03-10 16:58:52 +00:00
|
|
|
setOperationAction(ISD::SELECT, MVT::i1, Promote);
|
|
|
|
setOperationAction(ISD::SELECT, MVT::i8, Legal);
|
2008-03-05 23:02:02 +00:00
|
|
|
setOperationAction(ISD::SELECT, MVT::i16, Legal);
|
|
|
|
setOperationAction(ISD::SELECT, MVT::i32, Legal);
|
2007-12-04 22:23:35 +00:00
|
|
|
setOperationAction(ISD::SELECT, MVT::i64, Expand);
|
|
|
|
|
2008-03-10 16:58:52 +00:00
|
|
|
setOperationAction(ISD::SETCC, MVT::i1, Promote);
|
|
|
|
setOperationAction(ISD::SETCC, MVT::i8, Legal);
|
|
|
|
setOperationAction(ISD::SETCC, MVT::i16, Legal);
|
|
|
|
setOperationAction(ISD::SETCC, MVT::i32, Legal);
|
|
|
|
setOperationAction(ISD::SETCC, MVT::i64, Expand);
|
2008-03-05 23:02:02 +00:00
|
|
|
|
2008-02-23 18:41:37 +00:00
|
|
|
// Zero extension and sign extension for i64 have to be
|
|
|
|
// custom legalized
|
|
|
|
setOperationAction(ISD::ZERO_EXTEND, MVT::i64, Custom);
|
|
|
|
setOperationAction(ISD::SIGN_EXTEND, MVT::i64, Custom);
|
|
|
|
setOperationAction(ISD::ANY_EXTEND, MVT::i64, Custom);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// SPU has a legal FP -> signed INT instruction
|
|
|
|
setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
|
|
|
|
setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
|
|
|
|
setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
|
|
|
|
setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
|
|
|
|
|
|
|
|
// FDIV on SPU requires custom lowering
|
|
|
|
setOperationAction(ISD::FDIV, MVT::f32, Custom);
|
|
|
|
//setOperationAction(ISD::FDIV, MVT::f64, Custom);
|
|
|
|
|
|
|
|
// SPU has [U|S]INT_TO_FP
|
|
|
|
setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
|
|
|
|
setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
|
|
|
|
setOperationAction(ISD::SINT_TO_FP, MVT::i8, Promote);
|
|
|
|
setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
|
|
|
|
setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote);
|
|
|
|
setOperationAction(ISD::UINT_TO_FP, MVT::i8, Promote);
|
|
|
|
setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
|
|
|
|
setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
|
|
|
|
|
2007-12-20 00:44:13 +00:00
|
|
|
setOperationAction(ISD::BIT_CONVERT, MVT::i32, Legal);
|
|
|
|
setOperationAction(ISD::BIT_CONVERT, MVT::f32, Legal);
|
|
|
|
setOperationAction(ISD::BIT_CONVERT, MVT::i64, Legal);
|
|
|
|
setOperationAction(ISD::BIT_CONVERT, MVT::f64, Legal);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
// We cannot sextinreg(i1). Expand to shifts.
|
|
|
|
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Support label based line numbers.
|
2008-06-30 20:59:49 +00:00
|
|
|
setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
|
2007-12-04 22:23:35 +00:00
|
|
|
setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
|
|
|
// We want to legalize GlobalAddress and ConstantPool nodes into the
|
2007-12-04 22:23:35 +00:00
|
|
|
// appropriate instructions to materialize the address.
|
2008-01-29 02:16:57 +00:00
|
|
|
for (unsigned sctype = (unsigned) MVT::i1; sctype < (unsigned) MVT::f128;
|
|
|
|
++sctype) {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = (MVT::SimpleValueType)sctype;
|
|
|
|
|
|
|
|
setOperationAction(ISD::GlobalAddress, VT, Custom);
|
|
|
|
setOperationAction(ISD::ConstantPool, VT, Custom);
|
|
|
|
setOperationAction(ISD::JumpTable, VT, Custom);
|
2008-01-29 02:16:57 +00:00
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
// RET must be custom lowered, to meet ABI requirements
|
|
|
|
setOperationAction(ISD::RET, MVT::Other, Custom);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// VASTART needs to be custom lowered to use the VarArgsFrameIndex
|
|
|
|
setOperationAction(ISD::VASTART , MVT::Other, Custom);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Use the default implementation.
|
|
|
|
setOperationAction(ISD::VAARG , MVT::Other, Expand);
|
|
|
|
setOperationAction(ISD::VACOPY , MVT::Other, Expand);
|
|
|
|
setOperationAction(ISD::VAEND , MVT::Other, Expand);
|
2008-07-16 17:17:29 +00:00
|
|
|
setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
|
2007-12-04 22:23:35 +00:00
|
|
|
setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand);
|
|
|
|
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
|
|
|
|
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Expand);
|
|
|
|
|
|
|
|
// Cell SPU has instructions for converting between i64 and fp.
|
|
|
|
setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
|
|
|
|
setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
|
|
|
|
setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
|
|
|
|
|
|
|
|
// BUILD_PAIR can't be handled natively, and should be expanded to shl/or
|
|
|
|
setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
|
|
|
|
|
|
|
|
// First set operation action for all vector types to expand. Then we
|
|
|
|
// will selectively turn on ones that can be effectively codegen'd.
|
|
|
|
addRegisterClass(MVT::v16i8, SPU::VECREGRegisterClass);
|
|
|
|
addRegisterClass(MVT::v8i16, SPU::VECREGRegisterClass);
|
|
|
|
addRegisterClass(MVT::v4i32, SPU::VECREGRegisterClass);
|
|
|
|
addRegisterClass(MVT::v2i64, SPU::VECREGRegisterClass);
|
|
|
|
addRegisterClass(MVT::v4f32, SPU::VECREGRegisterClass);
|
|
|
|
addRegisterClass(MVT::v2f64, SPU::VECREGRegisterClass);
|
|
|
|
|
2008-06-06 12:08:01 +00:00
|
|
|
for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
|
|
|
|
i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
|
|
|
|
MVT VT = (MVT::SimpleValueType)i;
|
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// add/sub are legal for all supported vector VT's.
|
2008-06-06 12:08:01 +00:00
|
|
|
setOperationAction(ISD::ADD , VT, Legal);
|
|
|
|
setOperationAction(ISD::SUB , VT, Legal);
|
2007-12-04 22:23:35 +00:00
|
|
|
// mul has to be custom lowered.
|
2008-06-06 12:08:01 +00:00
|
|
|
setOperationAction(ISD::MUL , VT, Custom);
|
|
|
|
|
|
|
|
setOperationAction(ISD::AND , VT, Legal);
|
|
|
|
setOperationAction(ISD::OR , VT, Legal);
|
|
|
|
setOperationAction(ISD::XOR , VT, Legal);
|
|
|
|
setOperationAction(ISD::LOAD , VT, Legal);
|
|
|
|
setOperationAction(ISD::SELECT, VT, Legal);
|
|
|
|
setOperationAction(ISD::STORE, VT, Legal);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// These operations need to be expanded:
|
2008-06-06 12:08:01 +00:00
|
|
|
setOperationAction(ISD::SDIV, VT, Expand);
|
|
|
|
setOperationAction(ISD::SREM, VT, Expand);
|
|
|
|
setOperationAction(ISD::UDIV, VT, Expand);
|
|
|
|
setOperationAction(ISD::UREM, VT, Expand);
|
|
|
|
setOperationAction(ISD::FDIV, VT, Custom);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
// Custom lower build_vector, constant pool spills, insert and
|
|
|
|
// extract vector elements:
|
2008-06-06 12:08:01 +00:00
|
|
|
setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
|
|
|
|
setOperationAction(ISD::ConstantPool, VT, Custom);
|
|
|
|
setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
|
|
|
|
setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
|
|
|
|
setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
|
|
|
|
setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
setOperationAction(ISD::MUL, MVT::v16i8, Custom);
|
|
|
|
setOperationAction(ISD::AND, MVT::v16i8, Custom);
|
|
|
|
setOperationAction(ISD::OR, MVT::v16i8, Custom);
|
|
|
|
setOperationAction(ISD::XOR, MVT::v16i8, Custom);
|
|
|
|
setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
|
2008-01-11 02:53:15 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
setShiftAmountType(MVT::i32);
|
|
|
|
setSetCCResultContents(ZeroOrOneSetCCResult);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
setStackPointerRegisterToSaveRestore(SPU::R1);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// We have target-specific dag combine patterns for the following nodes:
|
2008-01-29 02:16:57 +00:00
|
|
|
setTargetDAGCombine(ISD::ADD);
|
2008-02-23 18:41:37 +00:00
|
|
|
setTargetDAGCombine(ISD::ZERO_EXTEND);
|
|
|
|
setTargetDAGCombine(ISD::SIGN_EXTEND);
|
|
|
|
setTargetDAGCombine(ISD::ANY_EXTEND);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
computeRegisterProperties();
|
|
|
|
}
|
|
|
|
|
|
|
|
const char *
|
|
|
|
SPUTargetLowering::getTargetNodeName(unsigned Opcode) const
|
|
|
|
{
|
|
|
|
if (node_names.empty()) {
|
|
|
|
node_names[(unsigned) SPUISD::RET_FLAG] = "SPUISD::RET_FLAG";
|
|
|
|
node_names[(unsigned) SPUISD::Hi] = "SPUISD::Hi";
|
|
|
|
node_names[(unsigned) SPUISD::Lo] = "SPUISD::Lo";
|
|
|
|
node_names[(unsigned) SPUISD::PCRelAddr] = "SPUISD::PCRelAddr";
|
2008-01-11 02:53:15 +00:00
|
|
|
node_names[(unsigned) SPUISD::AFormAddr] = "SPUISD::AFormAddr";
|
2008-01-29 02:16:57 +00:00
|
|
|
node_names[(unsigned) SPUISD::IndirectAddr] = "SPUISD::IndirectAddr";
|
2007-12-04 22:23:35 +00:00
|
|
|
node_names[(unsigned) SPUISD::LDRESULT] = "SPUISD::LDRESULT";
|
|
|
|
node_names[(unsigned) SPUISD::CALL] = "SPUISD::CALL";
|
|
|
|
node_names[(unsigned) SPUISD::SHUFB] = "SPUISD::SHUFB";
|
|
|
|
node_names[(unsigned) SPUISD::INSERT_MASK] = "SPUISD::INSERT_MASK";
|
|
|
|
node_names[(unsigned) SPUISD::CNTB] = "SPUISD::CNTB";
|
|
|
|
node_names[(unsigned) SPUISD::PROMOTE_SCALAR] = "SPUISD::PROMOTE_SCALAR";
|
|
|
|
node_names[(unsigned) SPUISD::EXTRACT_ELT0] = "SPUISD::EXTRACT_ELT0";
|
|
|
|
node_names[(unsigned) SPUISD::EXTRACT_ELT0_CHAINED] = "SPUISD::EXTRACT_ELT0_CHAINED";
|
|
|
|
node_names[(unsigned) SPUISD::EXTRACT_I1_ZEXT] = "SPUISD::EXTRACT_I1_ZEXT";
|
|
|
|
node_names[(unsigned) SPUISD::EXTRACT_I1_SEXT] = "SPUISD::EXTRACT_I1_SEXT";
|
|
|
|
node_names[(unsigned) SPUISD::EXTRACT_I8_ZEXT] = "SPUISD::EXTRACT_I8_ZEXT";
|
|
|
|
node_names[(unsigned) SPUISD::EXTRACT_I8_SEXT] = "SPUISD::EXTRACT_I8_SEXT";
|
|
|
|
node_names[(unsigned) SPUISD::MPY] = "SPUISD::MPY";
|
|
|
|
node_names[(unsigned) SPUISD::MPYU] = "SPUISD::MPYU";
|
|
|
|
node_names[(unsigned) SPUISD::MPYH] = "SPUISD::MPYH";
|
|
|
|
node_names[(unsigned) SPUISD::MPYHH] = "SPUISD::MPYHH";
|
2008-02-23 18:41:37 +00:00
|
|
|
node_names[(unsigned) SPUISD::SHLQUAD_L_BITS] = "SPUISD::SHLQUAD_L_BITS";
|
|
|
|
node_names[(unsigned) SPUISD::SHLQUAD_L_BYTES] = "SPUISD::SHLQUAD_L_BYTES";
|
2007-12-04 22:23:35 +00:00
|
|
|
node_names[(unsigned) SPUISD::VEC_SHL] = "SPUISD::VEC_SHL";
|
|
|
|
node_names[(unsigned) SPUISD::VEC_SRL] = "SPUISD::VEC_SRL";
|
|
|
|
node_names[(unsigned) SPUISD::VEC_SRA] = "SPUISD::VEC_SRA";
|
|
|
|
node_names[(unsigned) SPUISD::VEC_ROTL] = "SPUISD::VEC_ROTL";
|
|
|
|
node_names[(unsigned) SPUISD::VEC_ROTR] = "SPUISD::VEC_ROTR";
|
2008-02-23 18:41:37 +00:00
|
|
|
node_names[(unsigned) SPUISD::ROTQUAD_RZ_BYTES] =
|
|
|
|
"SPUISD::ROTQUAD_RZ_BYTES";
|
|
|
|
node_names[(unsigned) SPUISD::ROTQUAD_RZ_BITS] =
|
|
|
|
"SPUISD::ROTQUAD_RZ_BITS";
|
2007-12-04 22:23:35 +00:00
|
|
|
node_names[(unsigned) SPUISD::ROTBYTES_RIGHT_S] =
|
|
|
|
"SPUISD::ROTBYTES_RIGHT_S";
|
|
|
|
node_names[(unsigned) SPUISD::ROTBYTES_LEFT] = "SPUISD::ROTBYTES_LEFT";
|
|
|
|
node_names[(unsigned) SPUISD::ROTBYTES_LEFT_CHAINED] =
|
|
|
|
"SPUISD::ROTBYTES_LEFT_CHAINED";
|
2008-06-02 22:18:03 +00:00
|
|
|
node_names[(unsigned) SPUISD::ROTBYTES_LEFT_BITS] =
|
|
|
|
"SPUISD::ROTBYTES_LEFT_BITS";
|
|
|
|
node_names[(unsigned) SPUISD::SELECT_MASK] = "SPUISD::SELECT_MASK";
|
2007-12-04 22:23:35 +00:00
|
|
|
node_names[(unsigned) SPUISD::SELB] = "SPUISD::SELB";
|
2008-06-02 22:18:03 +00:00
|
|
|
node_names[(unsigned) SPUISD::ADD_EXTENDED] = "SPUISD::ADD_EXTENDED";
|
|
|
|
node_names[(unsigned) SPUISD::CARRY_GENERATE] = "SPUISD::CARRY_GENERATE";
|
|
|
|
node_names[(unsigned) SPUISD::SUB_EXTENDED] = "SPUISD::SUB_EXTENDED";
|
|
|
|
node_names[(unsigned) SPUISD::BORROW_GENERATE] = "SPUISD::BORROW_GENERATE";
|
2007-12-04 22:23:35 +00:00
|
|
|
node_names[(unsigned) SPUISD::FPInterp] = "SPUISD::FPInterp";
|
|
|
|
node_names[(unsigned) SPUISD::FPRecipEst] = "SPUISD::FPRecipEst";
|
|
|
|
node_names[(unsigned) SPUISD::SEXT32TO64] = "SPUISD::SEXT32TO64";
|
|
|
|
}
|
|
|
|
|
|
|
|
std::map<unsigned, const char *>::iterator i = node_names.find(Opcode);
|
|
|
|
|
|
|
|
return ((i != node_names.end()) ? i->second : 0);
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
MVT SPUTargetLowering::getSetCCResultType(const SDValue &Op) const {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Op.getValueType();
|
|
|
|
if (VT.isInteger())
|
2008-03-10 23:49:09 +00:00
|
|
|
return VT;
|
|
|
|
else
|
|
|
|
return MVT::i32;
|
2008-03-10 16:58:52 +00:00
|
|
|
}
|
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Calling convention code:
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
#include "SPUGenCallingConv.inc"
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// LowerOperation implementation
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
/// Aligned load common code for CellSPU
|
|
|
|
/*!
|
|
|
|
\param[in] Op The SelectionDAG load or store operand
|
|
|
|
\param[in] DAG The selection DAG
|
|
|
|
\param[in] ST CellSPU subtarget information structure
|
|
|
|
\param[in,out] alignment Caller initializes this to the load or store node's
|
|
|
|
value from getAlignment(), may be updated while generating the aligned load
|
|
|
|
\param[in,out] alignOffs Aligned offset; set by AlignedLoad to the aligned
|
|
|
|
offset (divisible by 16, modulo 16 == 0)
|
|
|
|
\param[in,out] prefSlotOffs Preferred slot offset; set by AlignedLoad to the
|
|
|
|
offset of the preferred slot (modulo 16 != 0)
|
|
|
|
\param[in,out] VT Caller initializes this value type to the the load or store
|
|
|
|
node's loaded or stored value type; may be updated if an i1-extended load or
|
|
|
|
store.
|
|
|
|
\param[out] was16aligned true if the base pointer had 16-byte alignment,
|
|
|
|
otherwise false. Can help to determine if the chunk needs to be rotated.
|
|
|
|
|
|
|
|
Both load and store lowering load a block of data aligned on a 16-byte
|
|
|
|
boundary. This is the common aligned load code shared between both.
|
|
|
|
*/
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue
|
|
|
|
AlignedLoad(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST,
|
2008-01-11 02:53:15 +00:00
|
|
|
LSBaseSDNode *LSN,
|
|
|
|
unsigned &alignment, int &alignOffs, int &prefSlotOffs,
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT &VT, bool &was16aligned)
|
2008-01-11 02:53:15 +00:00
|
|
|
{
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
|
2008-01-11 02:53:15 +00:00
|
|
|
const valtype_map_s *vtm = getValueTypeMapEntry(VT);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue basePtr = LSN->getBasePtr();
|
|
|
|
SDValue chain = LSN->getChain();
|
2008-01-11 02:53:15 +00:00
|
|
|
|
|
|
|
if (basePtr.getOpcode() == ISD::ADD) {
|
2008-08-28 21:40:38 +00:00
|
|
|
SDValue Op1 = basePtr.getNode()->getOperand(1);
|
2008-01-11 02:53:15 +00:00
|
|
|
|
|
|
|
if (Op1.getOpcode() == ISD::Constant || Op1.getOpcode() == ISD::TargetConstant) {
|
2008-01-17 20:38:41 +00:00
|
|
|
const ConstantSDNode *CN = cast<ConstantSDNode>(basePtr.getOperand(1));
|
2008-01-11 02:53:15 +00:00
|
|
|
|
|
|
|
alignOffs = (int) CN->getValue();
|
|
|
|
prefSlotOffs = (int) (alignOffs & 0xf);
|
|
|
|
|
|
|
|
// Adjust the rotation amount to ensure that the final result ends up in
|
|
|
|
// the preferred slot:
|
|
|
|
prefSlotOffs -= vtm->prefslot_byte;
|
|
|
|
basePtr = basePtr.getOperand(0);
|
|
|
|
|
2008-01-17 20:38:41 +00:00
|
|
|
// Loading from memory, can we adjust alignment?
|
|
|
|
if (basePtr.getOpcode() == SPUISD::AFormAddr) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue APtr = basePtr.getOperand(0);
|
2008-01-17 20:38:41 +00:00
|
|
|
if (APtr.getOpcode() == ISD::TargetGlobalAddress) {
|
|
|
|
GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(APtr);
|
|
|
|
alignment = GSDN->getGlobal()->getAlignment();
|
|
|
|
}
|
2008-01-11 02:53:15 +00:00
|
|
|
}
|
|
|
|
} else {
|
|
|
|
alignOffs = 0;
|
|
|
|
prefSlotOffs = -vtm->prefslot_byte;
|
|
|
|
}
|
2008-04-30 00:30:08 +00:00
|
|
|
} else if (basePtr.getOpcode() == ISD::FrameIndex) {
|
|
|
|
FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(basePtr);
|
|
|
|
alignOffs = int(FIN->getIndex() * SPUFrameInfo::stackSlotSize());
|
|
|
|
prefSlotOffs = (int) (alignOffs & 0xf);
|
|
|
|
prefSlotOffs -= vtm->prefslot_byte;
|
|
|
|
basePtr = DAG.getRegister(SPU::R1, VT);
|
2008-01-11 02:53:15 +00:00
|
|
|
} else {
|
|
|
|
alignOffs = 0;
|
|
|
|
prefSlotOffs = -vtm->prefslot_byte;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (alignment == 16) {
|
|
|
|
// Realign the base pointer as a D-Form address:
|
|
|
|
if (!isMemoryOperand(basePtr) || (alignOffs & ~0xf) != 0) {
|
2008-01-17 20:38:41 +00:00
|
|
|
basePtr = DAG.getNode(ISD::ADD, PtrVT,
|
|
|
|
basePtr,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getConstant((alignOffs & ~0xf), PtrVT));
|
2008-01-11 02:53:15 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Emit the vector load:
|
|
|
|
was16aligned = true;
|
|
|
|
return DAG.getLoad(MVT::v16i8, chain, basePtr,
|
|
|
|
LSN->getSrcValue(), LSN->getSrcValueOffset(),
|
|
|
|
LSN->isVolatile(), 16);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Unaligned load or we're using the "large memory" model, which means that
|
|
|
|
// we have to be very pessimistic:
|
2008-01-17 20:38:41 +00:00
|
|
|
if (isMemoryOperand(basePtr) || isIndirectOperand(basePtr)) {
|
2008-01-29 02:16:57 +00:00
|
|
|
basePtr = DAG.getNode(SPUISD::IndirectAddr, PtrVT, basePtr, DAG.getConstant(0, PtrVT));
|
2008-01-11 02:53:15 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Add the offset
|
2008-01-29 02:16:57 +00:00
|
|
|
basePtr = DAG.getNode(ISD::ADD, PtrVT, basePtr,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getConstant((alignOffs & ~0xf), PtrVT));
|
2008-01-11 02:53:15 +00:00
|
|
|
was16aligned = false;
|
|
|
|
return DAG.getLoad(MVT::v16i8, chain, basePtr,
|
|
|
|
LSN->getSrcValue(), LSN->getSrcValueOffset(),
|
|
|
|
LSN->isVolatile(), 16);
|
|
|
|
}
|
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
/// Custom lower loads for CellSPU
|
|
|
|
/*!
|
|
|
|
All CellSPU loads and stores are aligned to 16-byte boundaries, so for elements
|
|
|
|
within a 16-byte block, we have to rotate to extract the requested element.
|
|
|
|
*/
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue
|
|
|
|
LowerLOAD(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
|
2007-12-04 22:23:35 +00:00
|
|
|
LoadSDNode *LN = cast<LoadSDNode>(Op);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue the_chain = LN->getChain();
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = LN->getMemoryVT();
|
2008-08-28 21:40:38 +00:00
|
|
|
MVT OpVT = Op.getNode()->getValueType(0);
|
2007-12-04 22:23:35 +00:00
|
|
|
ISD::LoadExtType ExtType = LN->getExtensionType();
|
|
|
|
unsigned alignment = LN->getAlignment();
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Ops[8];
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
switch (LN->getAddressingMode()) {
|
|
|
|
case ISD::UNINDEXED: {
|
2008-01-11 02:53:15 +00:00
|
|
|
int offset, rotamt;
|
|
|
|
bool was16aligned;
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue result =
|
2008-01-11 02:53:15 +00:00
|
|
|
AlignedLoad(Op, DAG, ST, LN,alignment, offset, rotamt, VT, was16aligned);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-08-28 21:40:38 +00:00
|
|
|
if (result.getNode() == 0)
|
2008-01-11 02:53:15 +00:00
|
|
|
return result;
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
the_chain = result.getValue(1);
|
|
|
|
// Rotate the chunk if necessary
|
|
|
|
if (rotamt < 0)
|
|
|
|
rotamt += 16;
|
2008-01-11 21:01:19 +00:00
|
|
|
if (rotamt != 0 || !was16aligned) {
|
2008-01-11 02:53:15 +00:00
|
|
|
SDVTList vecvts = DAG.getVTList(MVT::v16i8, MVT::Other);
|
|
|
|
|
2008-01-17 20:38:41 +00:00
|
|
|
Ops[0] = the_chain;
|
|
|
|
Ops[1] = result;
|
2008-01-11 02:53:15 +00:00
|
|
|
if (was16aligned) {
|
|
|
|
Ops[2] = DAG.getConstant(rotamt, MVT::i16);
|
|
|
|
} else {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
|
2008-01-11 02:53:15 +00:00
|
|
|
LoadSDNode *LN1 = cast<LoadSDNode>(result);
|
2008-01-11 21:01:19 +00:00
|
|
|
Ops[2] = DAG.getNode(ISD::ADD, PtrVT, LN1->getBasePtr(),
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getConstant(rotamt, PtrVT));
|
2008-01-11 02:53:15 +00:00
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
result = DAG.getNode(SPUISD::ROTBYTES_LEFT_CHAINED, vecvts, Ops, 3);
|
|
|
|
the_chain = result.getValue(1);
|
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
if (VT == OpVT || ExtType == ISD::EXTLOAD) {
|
|
|
|
SDVTList scalarvts;
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT vecVT = MVT::v16i8;
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
// Convert the loaded v16i8 vector to the appropriate vector type
|
|
|
|
// specified by the operand:
|
|
|
|
if (OpVT == VT) {
|
|
|
|
if (VT != MVT::i1)
|
2008-06-06 12:08:01 +00:00
|
|
|
vecVT = MVT::getVectorVT(VT, (128 / VT.getSizeInBits()));
|
2008-01-11 02:53:15 +00:00
|
|
|
} else
|
2008-06-06 12:08:01 +00:00
|
|
|
vecVT = MVT::getVectorVT(OpVT, (128 / OpVT.getSizeInBits()));
|
2008-01-11 02:53:15 +00:00
|
|
|
|
|
|
|
Ops[0] = the_chain;
|
|
|
|
Ops[1] = DAG.getNode(ISD::BIT_CONVERT, vecVT, result);
|
|
|
|
scalarvts = DAG.getVTList((OpVT == VT ? VT : OpVT), MVT::Other);
|
|
|
|
result = DAG.getNode(SPUISD::EXTRACT_ELT0_CHAINED, scalarvts, Ops, 2);
|
|
|
|
the_chain = result.getValue(1);
|
2007-12-04 22:23:35 +00:00
|
|
|
} else {
|
2008-01-11 02:53:15 +00:00
|
|
|
// Handle the sign and zero-extending loads for i1 and i8:
|
|
|
|
unsigned NewOpC;
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
if (ExtType == ISD::SEXTLOAD) {
|
|
|
|
NewOpC = (OpVT == MVT::i1
|
|
|
|
? SPUISD::EXTRACT_I1_SEXT
|
|
|
|
: SPUISD::EXTRACT_I8_SEXT);
|
2007-12-04 22:23:35 +00:00
|
|
|
} else {
|
2008-01-11 02:53:15 +00:00
|
|
|
assert(ExtType == ISD::ZEXTLOAD);
|
|
|
|
NewOpC = (OpVT == MVT::i1
|
|
|
|
? SPUISD::EXTRACT_I1_ZEXT
|
|
|
|
: SPUISD::EXTRACT_I8_ZEXT);
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
result = DAG.getNode(NewOpC, OpVT, result);
|
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
SDVTList retvts = DAG.getVTList(OpVT, MVT::Other);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue retops[2] = {
|
2008-01-17 20:38:41 +00:00
|
|
|
result,
|
2008-01-30 02:55:46 +00:00
|
|
|
the_chain
|
2008-01-17 20:38:41 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
result = DAG.getNode(SPUISD::LDRESULT, retvts,
|
|
|
|
retops, sizeof(retops) / sizeof(retops[0]));
|
2008-01-11 02:53:15 +00:00
|
|
|
return result;
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
case ISD::PRE_INC:
|
|
|
|
case ISD::PRE_DEC:
|
|
|
|
case ISD::POST_INC:
|
|
|
|
case ISD::POST_DEC:
|
|
|
|
case ISD::LAST_INDEXED_MODE:
|
|
|
|
cerr << "LowerLOAD: Got a LoadSDNode with an addr mode other than "
|
|
|
|
"UNINDEXED\n";
|
|
|
|
cerr << (unsigned) LN->getAddressingMode() << "\n";
|
|
|
|
abort();
|
|
|
|
/*NOTREACHED*/
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Custom lower stores for CellSPU
|
|
|
|
/*!
|
|
|
|
All CellSPU stores are aligned to 16-byte boundaries, so for elements
|
|
|
|
within a 16-byte block, we have to generate a shuffle to insert the
|
|
|
|
requested element into its place, then store the resulting block.
|
|
|
|
*/
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue
|
|
|
|
LowerSTORE(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
|
2007-12-04 22:23:35 +00:00
|
|
|
StoreSDNode *SN = cast<StoreSDNode>(Op);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Value = SN->getValue();
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Value.getValueType();
|
|
|
|
MVT StVT = (!SN->isTruncatingStore() ? VT : SN->getMemoryVT());
|
|
|
|
MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
|
2008-01-11 02:53:15 +00:00
|
|
|
unsigned alignment = SN->getAlignment();
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
switch (SN->getAddressingMode()) {
|
|
|
|
case ISD::UNINDEXED: {
|
2008-01-11 02:53:15 +00:00
|
|
|
int chunk_offset, slot_offset;
|
|
|
|
bool was16aligned;
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
// The vector type we really want to load from the 16-byte chunk, except
|
|
|
|
// in the case of MVT::i1, which has to be v16i8.
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT vecVT, stVecVT = MVT::v16i8;
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
if (StVT != MVT::i1)
|
2008-06-06 12:08:01 +00:00
|
|
|
stVecVT = MVT::getVectorVT(StVT, (128 / StVT.getSizeInBits()));
|
|
|
|
vecVT = MVT::getVectorVT(VT, (128 / VT.getSizeInBits()));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue alignLoadVec =
|
2008-01-11 02:53:15 +00:00
|
|
|
AlignedLoad(Op, DAG, ST, SN, alignment,
|
|
|
|
chunk_offset, slot_offset, VT, was16aligned);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-08-28 21:40:38 +00:00
|
|
|
if (alignLoadVec.getNode() == 0)
|
2008-01-11 02:53:15 +00:00
|
|
|
return alignLoadVec;
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
LoadSDNode *LN = cast<LoadSDNode>(alignLoadVec);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue basePtr = LN->getBasePtr();
|
|
|
|
SDValue the_chain = alignLoadVec.getValue(1);
|
|
|
|
SDValue theValue = SN->getValue();
|
|
|
|
SDValue result;
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
if (StVT != VT
|
2008-01-30 02:55:46 +00:00
|
|
|
&& (theValue.getOpcode() == ISD::AssertZext
|
|
|
|
|| theValue.getOpcode() == ISD::AssertSext)) {
|
2007-12-04 22:23:35 +00:00
|
|
|
// Drill down and get the value for zero- and sign-extended
|
|
|
|
// quantities
|
2008-07-16 17:17:29 +00:00
|
|
|
theValue = theValue.getOperand(0);
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
chunk_offset &= 0xf;
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue insertEltOffs = DAG.getConstant(chunk_offset, PtrVT);
|
|
|
|
SDValue insertEltPtr;
|
|
|
|
SDValue insertEltOp;
|
2008-01-11 02:53:15 +00:00
|
|
|
|
|
|
|
// If the base pointer is already a D-form address, then just create
|
|
|
|
// a new D-form address with a slot offset and the orignal base pointer.
|
|
|
|
// Otherwise generate a D-form address with the slot offset relative
|
|
|
|
// to the stack pointer, which is always aligned.
|
2008-01-11 21:01:19 +00:00
|
|
|
DEBUG(cerr << "CellSPU LowerSTORE: basePtr = ");
|
2008-08-28 21:40:38 +00:00
|
|
|
DEBUG(basePtr.getNode()->dump(&DAG));
|
2008-01-11 21:01:19 +00:00
|
|
|
DEBUG(cerr << "\n");
|
|
|
|
|
2008-01-29 02:16:57 +00:00
|
|
|
if (basePtr.getOpcode() == SPUISD::IndirectAddr ||
|
|
|
|
(basePtr.getOpcode() == ISD::ADD
|
|
|
|
&& basePtr.getOperand(0).getOpcode() == SPUISD::IndirectAddr)) {
|
2008-01-11 21:01:19 +00:00
|
|
|
insertEltPtr = basePtr;
|
2008-01-11 02:53:15 +00:00
|
|
|
} else {
|
2008-01-29 02:16:57 +00:00
|
|
|
insertEltPtr = DAG.getNode(ISD::ADD, PtrVT, basePtr, insertEltOffs);
|
2008-01-11 02:53:15 +00:00
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
insertEltOp = DAG.getNode(SPUISD::INSERT_MASK, stVecVT, insertEltPtr);
|
2007-12-04 22:23:35 +00:00
|
|
|
result = DAG.getNode(SPUISD::SHUFB, vecVT,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::SCALAR_TO_VECTOR, vecVT, theValue),
|
|
|
|
alignLoadVec,
|
|
|
|
DAG.getNode(ISD::BIT_CONVERT, vecVT, insertEltOp));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
result = DAG.getStore(the_chain, result, basePtr,
|
2007-12-04 22:23:35 +00:00
|
|
|
LN->getSrcValue(), LN->getSrcValueOffset(),
|
|
|
|
LN->isVolatile(), LN->getAlignment());
|
|
|
|
|
|
|
|
return result;
|
|
|
|
/*UNREACHED*/
|
|
|
|
}
|
|
|
|
case ISD::PRE_INC:
|
|
|
|
case ISD::PRE_DEC:
|
|
|
|
case ISD::POST_INC:
|
|
|
|
case ISD::POST_DEC:
|
|
|
|
case ISD::LAST_INDEXED_MODE:
|
|
|
|
cerr << "LowerLOAD: Got a LoadSDNode with an addr mode other than "
|
|
|
|
"UNINDEXED\n";
|
|
|
|
cerr << (unsigned) SN->getAddressingMode() << "\n";
|
|
|
|
abort();
|
|
|
|
/*NOTREACHED*/
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// Generate the address of a constant pool entry.
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue
|
|
|
|
LowerConstantPool(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT PtrVT = Op.getValueType();
|
2007-12-04 22:23:35 +00:00
|
|
|
ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
|
|
|
|
Constant *C = CP->getConstVal();
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
|
|
|
|
SDValue Zero = DAG.getConstant(0, PtrVT);
|
2008-01-11 02:53:15 +00:00
|
|
|
const TargetMachine &TM = DAG.getTarget();
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
if (TM.getRelocationModel() == Reloc::Static) {
|
|
|
|
if (!ST->usingLargeMem()) {
|
2008-07-27 21:46:04 +00:00
|
|
|
// Just return the SDValue with the constant pool address in it.
|
2008-01-17 20:38:41 +00:00
|
|
|
return DAG.getNode(SPUISD::AFormAddr, PtrVT, CPI, Zero);
|
2007-12-04 22:23:35 +00:00
|
|
|
} else {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Hi = DAG.getNode(SPUISD::Hi, PtrVT, CPI, Zero);
|
|
|
|
SDValue Lo = DAG.getNode(SPUISD::Lo, PtrVT, CPI, Zero);
|
2008-02-23 18:41:37 +00:00
|
|
|
return DAG.getNode(SPUISD::IndirectAddr, PtrVT, Hi, Lo);
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
assert(0 &&
|
|
|
|
"LowerConstantPool: Relocation model other than static not supported.");
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue
|
|
|
|
LowerJumpTable(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT PtrVT = Op.getValueType();
|
2007-12-04 22:23:35 +00:00
|
|
|
JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PtrVT);
|
|
|
|
SDValue Zero = DAG.getConstant(0, PtrVT);
|
2007-12-04 22:23:35 +00:00
|
|
|
const TargetMachine &TM = DAG.getTarget();
|
|
|
|
|
|
|
|
if (TM.getRelocationModel() == Reloc::Static) {
|
2008-02-23 18:41:37 +00:00
|
|
|
if (!ST->usingLargeMem()) {
|
|
|
|
return DAG.getNode(SPUISD::AFormAddr, PtrVT, JTI, Zero);
|
|
|
|
} else {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Hi = DAG.getNode(SPUISD::Hi, PtrVT, JTI, Zero);
|
|
|
|
SDValue Lo = DAG.getNode(SPUISD::Lo, PtrVT, JTI, Zero);
|
2008-02-23 18:41:37 +00:00
|
|
|
return DAG.getNode(SPUISD::IndirectAddr, PtrVT, Hi, Lo);
|
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
assert(0 &&
|
|
|
|
"LowerJumpTable: Relocation model other than static not supported.");
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue
|
|
|
|
LowerGlobalAddress(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT PtrVT = Op.getValueType();
|
2007-12-04 22:23:35 +00:00
|
|
|
GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op);
|
|
|
|
GlobalValue *GV = GSDN->getGlobal();
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue GA = DAG.getTargetGlobalAddress(GV, PtrVT, GSDN->getOffset());
|
2007-12-04 22:23:35 +00:00
|
|
|
const TargetMachine &TM = DAG.getTarget();
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Zero = DAG.getConstant(0, PtrVT);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
if (TM.getRelocationModel() == Reloc::Static) {
|
2008-01-29 02:16:57 +00:00
|
|
|
if (!ST->usingLargeMem()) {
|
|
|
|
return DAG.getNode(SPUISD::AFormAddr, PtrVT, GA, Zero);
|
|
|
|
} else {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Hi = DAG.getNode(SPUISD::Hi, PtrVT, GA, Zero);
|
|
|
|
SDValue Lo = DAG.getNode(SPUISD::Lo, PtrVT, GA, Zero);
|
2008-01-29 02:16:57 +00:00
|
|
|
return DAG.getNode(SPUISD::IndirectAddr, PtrVT, Hi, Lo);
|
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
} else {
|
|
|
|
cerr << "LowerGlobalAddress: Relocation model other than static not "
|
2008-01-30 02:55:46 +00:00
|
|
|
<< "supported.\n";
|
2007-12-04 22:23:35 +00:00
|
|
|
abort();
|
|
|
|
/*NOTREACHED*/
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
//! Custom lower i64 integer constants
|
|
|
|
/*!
|
|
|
|
This code inserts all of the necessary juggling that needs to occur to load
|
|
|
|
a 64-bit constant into a register.
|
|
|
|
*/
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue
|
|
|
|
LowerConstant(SDValue Op, SelectionDAG &DAG) {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Op.getValueType();
|
2008-08-28 21:40:38 +00:00
|
|
|
ConstantSDNode *CN = cast<ConstantSDNode>(Op.getNode());
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
if (VT == MVT::i64) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue T = DAG.getConstant(CN->getValue(), MVT::i64);
|
2007-12-04 22:23:35 +00:00
|
|
|
return DAG.getNode(SPUISD::EXTRACT_ELT0, VT,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i64, T, T));
|
2007-12-04 22:23:35 +00:00
|
|
|
} else {
|
|
|
|
cerr << "LowerConstant: unhandled constant type "
|
2008-06-06 12:08:01 +00:00
|
|
|
<< VT.getMVTString()
|
2008-01-30 02:55:46 +00:00
|
|
|
<< "\n";
|
2007-12-04 22:23:35 +00:00
|
|
|
abort();
|
|
|
|
/*NOTREACHED*/
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-02-14 18:43:04 +00:00
|
|
|
//! Custom lower double precision floating point constants
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue
|
|
|
|
LowerConstantFP(SDValue Op, SelectionDAG &DAG) {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Op.getValueType();
|
2008-08-28 21:40:38 +00:00
|
|
|
ConstantFPSDNode *FP = cast<ConstantFPSDNode>(Op.getNode());
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
assert((FP != 0) &&
|
2008-01-30 02:55:46 +00:00
|
|
|
"LowerConstantFP: Node is not ConstantFPSDNode");
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-02-14 18:43:04 +00:00
|
|
|
if (VT == MVT::f64) {
|
2007-12-19 20:15:47 +00:00
|
|
|
uint64_t dbits = DoubleToBits(FP->getValueAPF().convertToDouble());
|
2007-12-04 22:23:35 +00:00
|
|
|
return DAG.getNode(ISD::BIT_CONVERT, VT,
|
2008-01-30 02:55:46 +00:00
|
|
|
LowerConstant(DAG.getConstant(dbits, MVT::i64), DAG));
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-01-17 20:38:41 +00:00
|
|
|
//! Lower MVT::i1, MVT::i8 brcond to a promoted type (MVT::i32, MVT::i16)
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue
|
|
|
|
LowerBRCOND(SDValue Op, SelectionDAG &DAG)
|
2008-01-17 20:38:41 +00:00
|
|
|
{
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Cond = Op.getOperand(1);
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT CondVT = Cond.getValueType();
|
|
|
|
MVT CondNVT;
|
2008-01-17 20:38:41 +00:00
|
|
|
|
|
|
|
if (CondVT == MVT::i1 || CondVT == MVT::i8) {
|
|
|
|
CondNVT = (CondVT == MVT::i1 ? MVT::i32 : MVT::i16);
|
|
|
|
return DAG.getNode(ISD::BRCOND, Op.getValueType(),
|
|
|
|
Op.getOperand(0),
|
|
|
|
DAG.getNode(ISD::ZERO_EXTEND, CondNVT, Op.getOperand(1)),
|
|
|
|
Op.getOperand(2));
|
|
|
|
} else
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue(); // Unchanged
|
2008-01-17 20:38:41 +00:00
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue
|
|
|
|
LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG, int &VarArgsFrameIndex)
|
2007-12-04 22:23:35 +00:00
|
|
|
{
|
|
|
|
MachineFunction &MF = DAG.getMachineFunction();
|
|
|
|
MachineFrameInfo *MFI = MF.getFrameInfo();
|
2007-12-31 04:13:23 +00:00
|
|
|
MachineRegisterInfo &RegInfo = MF.getRegInfo();
|
2008-07-27 21:46:04 +00:00
|
|
|
SmallVector<SDValue, 8> ArgValues;
|
|
|
|
SDValue Root = Op.getOperand(0);
|
2007-12-04 22:23:35 +00:00
|
|
|
bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
|
|
|
|
|
|
|
|
const unsigned *ArgRegs = SPURegisterInfo::getArgRegs();
|
|
|
|
const unsigned NumArgRegs = SPURegisterInfo::getNumArgRegs();
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
unsigned ArgOffset = SPUFrameInfo::minStackSize();
|
|
|
|
unsigned ArgRegIdx = 0;
|
|
|
|
unsigned StackSlotSize = SPUFrameInfo::stackSlotSize();
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Add DAG nodes to load the arguments or copy them out of registers.
|
2008-08-28 21:40:38 +00:00
|
|
|
for (unsigned ArgNo = 0, e = Op.getNode()->getNumValues()-1; ArgNo != e; ++ArgNo) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ArgVal;
|
2007-12-04 22:23:35 +00:00
|
|
|
bool needsLoad = false;
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT ObjectVT = Op.getValue(ArgNo).getValueType();
|
|
|
|
unsigned ObjSize = ObjectVT.getSizeInBits()/8;
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-06-06 12:08:01 +00:00
|
|
|
switch (ObjectVT.getSimpleVT()) {
|
2007-12-04 22:23:35 +00:00
|
|
|
default: {
|
|
|
|
cerr << "LowerFORMAL_ARGUMENTS Unhandled argument type: "
|
2008-06-06 12:08:01 +00:00
|
|
|
<< ObjectVT.getMVTString()
|
2007-12-04 22:23:35 +00:00
|
|
|
<< "\n";
|
|
|
|
abort();
|
|
|
|
}
|
|
|
|
case MVT::i8:
|
|
|
|
if (!isVarArg && ArgRegIdx < NumArgRegs) {
|
2007-12-31 04:13:23 +00:00
|
|
|
unsigned VReg = RegInfo.createVirtualRegister(&SPU::R8CRegClass);
|
|
|
|
RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
|
2007-12-04 22:23:35 +00:00
|
|
|
ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i8);
|
|
|
|
++ArgRegIdx;
|
|
|
|
} else {
|
|
|
|
needsLoad = true;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case MVT::i16:
|
|
|
|
if (!isVarArg && ArgRegIdx < NumArgRegs) {
|
2007-12-31 04:13:23 +00:00
|
|
|
unsigned VReg = RegInfo.createVirtualRegister(&SPU::R16CRegClass);
|
|
|
|
RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
|
2007-12-04 22:23:35 +00:00
|
|
|
ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i16);
|
|
|
|
++ArgRegIdx;
|
|
|
|
} else {
|
|
|
|
needsLoad = true;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case MVT::i32:
|
|
|
|
if (!isVarArg && ArgRegIdx < NumArgRegs) {
|
2007-12-31 04:13:23 +00:00
|
|
|
unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
|
|
|
|
RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
|
2007-12-04 22:23:35 +00:00
|
|
|
ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i32);
|
|
|
|
++ArgRegIdx;
|
|
|
|
} else {
|
|
|
|
needsLoad = true;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case MVT::i64:
|
|
|
|
if (!isVarArg && ArgRegIdx < NumArgRegs) {
|
2007-12-31 04:13:23 +00:00
|
|
|
unsigned VReg = RegInfo.createVirtualRegister(&SPU::R64CRegClass);
|
|
|
|
RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
|
2007-12-04 22:23:35 +00:00
|
|
|
ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::i64);
|
|
|
|
++ArgRegIdx;
|
|
|
|
} else {
|
|
|
|
needsLoad = true;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case MVT::f32:
|
|
|
|
if (!isVarArg && ArgRegIdx < NumArgRegs) {
|
2007-12-31 04:13:23 +00:00
|
|
|
unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32FPRegClass);
|
|
|
|
RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
|
2007-12-04 22:23:35 +00:00
|
|
|
ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::f32);
|
|
|
|
++ArgRegIdx;
|
|
|
|
} else {
|
|
|
|
needsLoad = true;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case MVT::f64:
|
|
|
|
if (!isVarArg && ArgRegIdx < NumArgRegs) {
|
2007-12-31 04:13:23 +00:00
|
|
|
unsigned VReg = RegInfo.createVirtualRegister(&SPU::R64FPRegClass);
|
|
|
|
RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
|
2007-12-04 22:23:35 +00:00
|
|
|
ArgVal = DAG.getCopyFromReg(Root, VReg, MVT::f64);
|
|
|
|
++ArgRegIdx;
|
|
|
|
} else {
|
|
|
|
needsLoad = true;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case MVT::v2f64:
|
|
|
|
case MVT::v4f32:
|
2008-03-05 23:02:02 +00:00
|
|
|
case MVT::v2i64:
|
2007-12-04 22:23:35 +00:00
|
|
|
case MVT::v4i32:
|
|
|
|
case MVT::v8i16:
|
|
|
|
case MVT::v16i8:
|
|
|
|
if (!isVarArg && ArgRegIdx < NumArgRegs) {
|
2007-12-31 04:13:23 +00:00
|
|
|
unsigned VReg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
|
|
|
|
RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
|
2007-12-04 22:23:35 +00:00
|
|
|
ArgVal = DAG.getCopyFromReg(Root, VReg, ObjectVT);
|
|
|
|
++ArgRegIdx;
|
|
|
|
} else {
|
|
|
|
needsLoad = true;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// We need to load the argument to a virtual register if we determined above
|
|
|
|
// that we ran out of physical registers of the appropriate type
|
|
|
|
if (needsLoad) {
|
2008-02-13 07:35:30 +00:00
|
|
|
int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
|
2008-02-13 07:35:30 +00:00
|
|
|
ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
|
2007-12-04 22:23:35 +00:00
|
|
|
ArgOffset += StackSlotSize;
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
ArgValues.push_back(ArgVal);
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// If the function takes variable number of arguments, make a frame index for
|
|
|
|
// the start of the first vararg value... for expansion of llvm.va_start.
|
|
|
|
if (isVarArg) {
|
2008-06-06 12:08:01 +00:00
|
|
|
VarArgsFrameIndex = MFI->CreateFixedObject(PtrVT.getSizeInBits()/8,
|
2007-12-04 22:23:35 +00:00
|
|
|
ArgOffset);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
|
2007-12-04 22:23:35 +00:00
|
|
|
// If this function is vararg, store any remaining integer argument regs to
|
|
|
|
// their spots on the stack so that they may be loaded by deferencing the
|
|
|
|
// result of va_next.
|
2008-07-27 21:46:04 +00:00
|
|
|
SmallVector<SDValue, 8> MemOps;
|
2007-12-04 22:23:35 +00:00
|
|
|
for (; ArgRegIdx != NumArgRegs; ++ArgRegIdx) {
|
2007-12-31 04:13:23 +00:00
|
|
|
unsigned VReg = RegInfo.createVirtualRegister(&SPU::GPRCRegClass);
|
|
|
|
RegInfo.addLiveIn(ArgRegs[ArgRegIdx], VReg);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
|
|
|
|
SDValue Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
|
2007-12-04 22:23:35 +00:00
|
|
|
MemOps.push_back(Store);
|
|
|
|
// Increment the address by four for the next argument to store
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, PtrVT);
|
2007-12-04 22:23:35 +00:00
|
|
|
FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
|
|
|
|
}
|
|
|
|
if (!MemOps.empty())
|
|
|
|
Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
ArgValues.push_back(Root);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Return the new list of results.
|
2008-08-28 21:40:38 +00:00
|
|
|
return DAG.getMergeValues(Op.getNode()->getVTList(), &ArgValues[0],
|
2008-06-30 10:19:09 +00:00
|
|
|
ArgValues.size());
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// isLSAAddress - Return the immediate to use if the specified
|
|
|
|
/// value is representable as a LSA address.
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDNode *isLSAAddress(SDValue Op, SelectionDAG &DAG) {
|
2007-12-04 22:23:35 +00:00
|
|
|
ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
|
|
|
|
if (!C) return 0;
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
int Addr = C->getValue();
|
|
|
|
if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero.
|
|
|
|
(Addr << 14 >> 14) != Addr)
|
|
|
|
return 0; // Top 14 bits have to be sext of immediate.
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-08-28 21:40:38 +00:00
|
|
|
return DAG.getConstant((int)C->getValue() >> 2, MVT::i32).getNode();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue
|
|
|
|
LowerCALL(SDValue Op, SelectionDAG &DAG, const SPUSubtarget *ST) {
|
|
|
|
SDValue Chain = Op.getOperand(0);
|
2007-12-04 22:23:35 +00:00
|
|
|
#if 0
|
|
|
|
bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
|
|
|
|
bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
|
|
|
|
#endif
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Callee = Op.getOperand(4);
|
2007-12-04 22:23:35 +00:00
|
|
|
unsigned NumOps = (Op.getNumOperands() - 5) / 2;
|
|
|
|
unsigned StackSlotSize = SPUFrameInfo::stackSlotSize();
|
|
|
|
const unsigned *ArgRegs = SPURegisterInfo::getArgRegs();
|
|
|
|
const unsigned NumArgRegs = SPURegisterInfo::getNumArgRegs();
|
|
|
|
|
|
|
|
// Handy pointer type
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Accumulate how many bytes are to be pushed on the stack, including the
|
|
|
|
// linkage area, and parameter passing area. According to the SPU ABI,
|
|
|
|
// we minimally need space for [LR] and [SP]
|
|
|
|
unsigned NumStackBytes = SPUFrameInfo::minStackSize();
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Set up a copy of the stack pointer for use loading and storing any
|
|
|
|
// arguments that may not fit in the registers available for argument
|
|
|
|
// passing.
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue StackPtr = DAG.getRegister(SPU::R1, MVT::i32);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Figure out which arguments are going to go in registers, and which in
|
|
|
|
// memory.
|
|
|
|
unsigned ArgOffset = SPUFrameInfo::minStackSize(); // Just below [LR]
|
|
|
|
unsigned ArgRegIdx = 0;
|
|
|
|
|
|
|
|
// Keep track of registers passing arguments
|
2008-07-27 21:46:04 +00:00
|
|
|
std::vector<std::pair<unsigned, SDValue> > RegsToPass;
|
2007-12-04 22:23:35 +00:00
|
|
|
// And the arguments passed on the stack
|
2008-07-27 21:46:04 +00:00
|
|
|
SmallVector<SDValue, 8> MemOpChains;
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
for (unsigned i = 0; i != NumOps; ++i) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Arg = Op.getOperand(5+2*i);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// PtrOff will be used to store the current argument to the stack if a
|
|
|
|
// register cannot be found for it.
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue PtrOff = DAG.getConstant(ArgOffset, StackPtr.getValueType());
|
2007-12-04 22:23:35 +00:00
|
|
|
PtrOff = DAG.getNode(ISD::ADD, PtrVT, StackPtr, PtrOff);
|
|
|
|
|
2008-06-06 12:08:01 +00:00
|
|
|
switch (Arg.getValueType().getSimpleVT()) {
|
2007-12-04 22:23:35 +00:00
|
|
|
default: assert(0 && "Unexpected ValueType for argument!");
|
|
|
|
case MVT::i32:
|
|
|
|
case MVT::i64:
|
|
|
|
case MVT::i128:
|
|
|
|
if (ArgRegIdx != NumArgRegs) {
|
|
|
|
RegsToPass.push_back(std::make_pair(ArgRegs[ArgRegIdx++], Arg));
|
|
|
|
} else {
|
|
|
|
MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
|
2008-01-30 02:55:46 +00:00
|
|
|
ArgOffset += StackSlotSize;
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
case MVT::f32:
|
|
|
|
case MVT::f64:
|
|
|
|
if (ArgRegIdx != NumArgRegs) {
|
|
|
|
RegsToPass.push_back(std::make_pair(ArgRegs[ArgRegIdx++], Arg));
|
|
|
|
} else {
|
|
|
|
MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
|
2008-01-30 02:55:46 +00:00
|
|
|
ArgOffset += StackSlotSize;
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
case MVT::v4f32:
|
|
|
|
case MVT::v4i32:
|
|
|
|
case MVT::v8i16:
|
|
|
|
case MVT::v16i8:
|
|
|
|
if (ArgRegIdx != NumArgRegs) {
|
|
|
|
RegsToPass.push_back(std::make_pair(ArgRegs[ArgRegIdx++], Arg));
|
|
|
|
} else {
|
|
|
|
MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
|
2008-01-30 02:55:46 +00:00
|
|
|
ArgOffset += StackSlotSize;
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// Update number of stack bytes actually used, insert a call sequence start
|
|
|
|
NumStackBytes = (ArgOffset - SPUFrameInfo::minStackSize());
|
|
|
|
Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumStackBytes, PtrVT));
|
|
|
|
|
|
|
|
if (!MemOpChains.empty()) {
|
|
|
|
// Adjust the stack pointer for the stack arguments.
|
|
|
|
Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
|
|
|
|
&MemOpChains[0], MemOpChains.size());
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Build a sequence of copy-to-reg nodes chained together with token chain
|
|
|
|
// and flag operands which copy the outgoing args into the appropriate regs.
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue InFlag;
|
2007-12-04 22:23:35 +00:00
|
|
|
for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
|
|
|
|
Chain = DAG.getCopyToReg(Chain, RegsToPass[i].first, RegsToPass[i].second,
|
|
|
|
InFlag);
|
|
|
|
InFlag = Chain.getValue(1);
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SmallVector<SDValue, 8> Ops;
|
2007-12-04 22:23:35 +00:00
|
|
|
unsigned CallOpc = SPUISD::CALL;
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
|
|
|
|
// direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
|
|
|
|
// node so that legalize doesn't hack it.
|
|
|
|
if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
|
|
|
|
GlobalValue *GV = G->getGlobal();
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT CalleeVT = Callee.getValueType();
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Zero = DAG.getConstant(0, PtrVT);
|
|
|
|
SDValue GA = DAG.getTargetGlobalAddress(GV, CalleeVT);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-01-11 02:53:15 +00:00
|
|
|
if (!ST->usingLargeMem()) {
|
|
|
|
// Turn calls to targets that are defined (i.e., have bodies) into BRSL
|
|
|
|
// style calls, otherwise, external symbols are BRASL calls. This assumes
|
|
|
|
// that declared/defined symbols are in the same compilation unit and can
|
|
|
|
// be reached through PC-relative jumps.
|
|
|
|
//
|
|
|
|
// NOTE:
|
|
|
|
// This may be an unsafe assumption for JIT and really large compilation
|
|
|
|
// units.
|
|
|
|
if (GV->isDeclaration()) {
|
|
|
|
Callee = DAG.getNode(SPUISD::AFormAddr, CalleeVT, GA, Zero);
|
|
|
|
} else {
|
|
|
|
Callee = DAG.getNode(SPUISD::PCRelAddr, CalleeVT, GA, Zero);
|
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
} else {
|
2008-01-11 02:53:15 +00:00
|
|
|
// "Large memory" mode: Turn all calls into indirect calls with a X-form
|
|
|
|
// address pairs:
|
2008-01-29 02:16:57 +00:00
|
|
|
Callee = DAG.getNode(SPUISD::IndirectAddr, PtrVT, GA, Zero);
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
} else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
|
|
|
|
Callee = DAG.getExternalSymbol(S->getSymbol(), Callee.getValueType());
|
2008-01-11 02:53:15 +00:00
|
|
|
else if (SDNode *Dest = isLSAAddress(Callee, DAG)) {
|
2007-12-04 22:23:35 +00:00
|
|
|
// If this is an absolute destination address that appears to be a legal
|
|
|
|
// local store address, use the munged value.
|
2008-07-27 21:46:04 +00:00
|
|
|
Callee = SDValue(Dest, 0);
|
2008-01-11 02:53:15 +00:00
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
Ops.push_back(Chain);
|
|
|
|
Ops.push_back(Callee);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Add argument registers to the end of the list so that they are known live
|
|
|
|
// into the call.
|
|
|
|
for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
|
2008-07-16 17:17:29 +00:00
|
|
|
Ops.push_back(DAG.getRegister(RegsToPass[i].first,
|
2007-12-04 22:23:35 +00:00
|
|
|
RegsToPass[i].second.getValueType()));
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-08-28 21:40:38 +00:00
|
|
|
if (InFlag.getNode())
|
2007-12-04 22:23:35 +00:00
|
|
|
Ops.push_back(InFlag);
|
2008-07-02 17:40:58 +00:00
|
|
|
// Returns a chain and a flag for retval copy to use.
|
|
|
|
Chain = DAG.getNode(CallOpc, DAG.getVTList(MVT::Other, MVT::Flag),
|
|
|
|
&Ops[0], Ops.size());
|
2007-12-04 22:23:35 +00:00
|
|
|
InFlag = Chain.getValue(1);
|
|
|
|
|
2008-02-05 22:44:06 +00:00
|
|
|
Chain = DAG.getCALLSEQ_END(Chain,
|
|
|
|
DAG.getConstant(NumStackBytes, PtrVT),
|
|
|
|
DAG.getConstant(0, PtrVT),
|
|
|
|
InFlag);
|
2008-08-28 21:40:38 +00:00
|
|
|
if (Op.getNode()->getValueType(0) != MVT::Other)
|
2008-02-05 22:44:06 +00:00
|
|
|
InFlag = Chain.getValue(1);
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ResultVals[3];
|
2007-12-04 22:23:35 +00:00
|
|
|
unsigned NumResults = 0;
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// If the call has results, copy the values out of the ret val registers.
|
2008-08-28 21:40:38 +00:00
|
|
|
switch (Op.getNode()->getValueType(0).getSimpleVT()) {
|
2007-12-04 22:23:35 +00:00
|
|
|
default: assert(0 && "Unexpected ret value!");
|
|
|
|
case MVT::Other: break;
|
|
|
|
case MVT::i32:
|
2008-08-28 21:40:38 +00:00
|
|
|
if (Op.getNode()->getValueType(1) == MVT::i32) {
|
2007-12-04 22:23:35 +00:00
|
|
|
Chain = DAG.getCopyFromReg(Chain, SPU::R4, MVT::i32, InFlag).getValue(1);
|
|
|
|
ResultVals[0] = Chain.getValue(0);
|
|
|
|
Chain = DAG.getCopyFromReg(Chain, SPU::R3, MVT::i32,
|
|
|
|
Chain.getValue(2)).getValue(1);
|
|
|
|
ResultVals[1] = Chain.getValue(0);
|
|
|
|
NumResults = 2;
|
|
|
|
} else {
|
|
|
|
Chain = DAG.getCopyFromReg(Chain, SPU::R3, MVT::i32, InFlag).getValue(1);
|
|
|
|
ResultVals[0] = Chain.getValue(0);
|
|
|
|
NumResults = 1;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case MVT::i64:
|
|
|
|
Chain = DAG.getCopyFromReg(Chain, SPU::R3, MVT::i64, InFlag).getValue(1);
|
|
|
|
ResultVals[0] = Chain.getValue(0);
|
|
|
|
NumResults = 1;
|
|
|
|
break;
|
|
|
|
case MVT::f32:
|
|
|
|
case MVT::f64:
|
2008-08-28 21:40:38 +00:00
|
|
|
Chain = DAG.getCopyFromReg(Chain, SPU::R3, Op.getNode()->getValueType(0),
|
2007-12-04 22:23:35 +00:00
|
|
|
InFlag).getValue(1);
|
|
|
|
ResultVals[0] = Chain.getValue(0);
|
|
|
|
NumResults = 1;
|
|
|
|
break;
|
|
|
|
case MVT::v2f64:
|
|
|
|
case MVT::v4f32:
|
|
|
|
case MVT::v4i32:
|
|
|
|
case MVT::v8i16:
|
|
|
|
case MVT::v16i8:
|
2008-08-28 21:40:38 +00:00
|
|
|
Chain = DAG.getCopyFromReg(Chain, SPU::R3, Op.getNode()->getValueType(0),
|
2007-12-04 22:23:35 +00:00
|
|
|
InFlag).getValue(1);
|
|
|
|
ResultVals[0] = Chain.getValue(0);
|
|
|
|
NumResults = 1;
|
|
|
|
break;
|
|
|
|
}
|
2008-07-02 17:40:58 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// If the function returns void, just return the chain.
|
|
|
|
if (NumResults == 0)
|
|
|
|
return Chain;
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Otherwise, merge everything together with a MERGE_VALUES node.
|
|
|
|
ResultVals[NumResults++] = Chain;
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Res = DAG.getMergeValues(ResultVals, NumResults);
|
2008-08-26 22:36:50 +00:00
|
|
|
return Res.getValue(Op.getResNo());
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue
|
|
|
|
LowerRET(SDValue Op, SelectionDAG &DAG, TargetMachine &TM) {
|
2007-12-04 22:23:35 +00:00
|
|
|
SmallVector<CCValAssign, 16> RVLocs;
|
|
|
|
unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
|
|
|
|
bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
|
|
|
|
CCState CCInfo(CC, isVarArg, TM, RVLocs);
|
2008-08-28 21:40:38 +00:00
|
|
|
CCInfo.AnalyzeReturn(Op.getNode(), RetCC_SPU);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// If this is the first return lowered for this function, add the regs to the
|
|
|
|
// liveout set for the function.
|
2007-12-31 04:13:23 +00:00
|
|
|
if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
|
2007-12-04 22:23:35 +00:00
|
|
|
for (unsigned i = 0; i != RVLocs.size(); ++i)
|
2007-12-31 04:13:23 +00:00
|
|
|
DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Chain = Op.getOperand(0);
|
|
|
|
SDValue Flag;
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Copy the result values into the output registers.
|
|
|
|
for (unsigned i = 0; i != RVLocs.size(); ++i) {
|
|
|
|
CCValAssign &VA = RVLocs[i];
|
|
|
|
assert(VA.isRegLoc() && "Can only return in registers!");
|
|
|
|
Chain = DAG.getCopyToReg(Chain, VA.getLocReg(), Op.getOperand(i*2+1), Flag);
|
|
|
|
Flag = Chain.getValue(1);
|
|
|
|
}
|
|
|
|
|
2008-08-28 21:40:38 +00:00
|
|
|
if (Flag.getNode())
|
2007-12-04 22:23:35 +00:00
|
|
|
return DAG.getNode(SPUISD::RET_FLAG, MVT::Other, Chain, Flag);
|
|
|
|
else
|
|
|
|
return DAG.getNode(SPUISD::RET_FLAG, MVT::Other, Chain);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Vector related lowering:
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
static ConstantSDNode *
|
|
|
|
getVecImm(SDNode *N) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue OpVal(0, 0);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Check to see if this buildvec has a single non-undef value in its elements.
|
|
|
|
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
|
|
|
|
if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
|
2008-08-28 21:40:38 +00:00
|
|
|
if (OpVal.getNode() == 0)
|
2007-12-04 22:23:35 +00:00
|
|
|
OpVal = N->getOperand(i);
|
|
|
|
else if (OpVal != N->getOperand(i))
|
|
|
|
return 0;
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-08-28 21:40:38 +00:00
|
|
|
if (OpVal.getNode() != 0) {
|
2007-12-04 22:23:35 +00:00
|
|
|
if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
|
|
|
|
return CN;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0; // All UNDEF: use implicit def.; not Constant node
|
|
|
|
}
|
|
|
|
|
|
|
|
/// get_vec_i18imm - Test if this vector is a vector filled with the same value
|
|
|
|
/// and the value fits into an unsigned 18-bit constant, and if so, return the
|
|
|
|
/// constant
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue SPU::get_vec_u18imm(SDNode *N, SelectionDAG &DAG,
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT ValueType) {
|
2007-12-04 22:23:35 +00:00
|
|
|
if (ConstantSDNode *CN = getVecImm(N)) {
|
|
|
|
uint64_t Value = CN->getValue();
|
2008-03-06 04:02:54 +00:00
|
|
|
if (ValueType == MVT::i64) {
|
|
|
|
uint64_t UValue = CN->getValue();
|
|
|
|
uint32_t upper = uint32_t(UValue >> 32);
|
|
|
|
uint32_t lower = uint32_t(UValue);
|
|
|
|
if (upper != lower)
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2008-03-06 04:02:54 +00:00
|
|
|
Value = Value >> 32;
|
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
if (Value <= 0x3ffff)
|
|
|
|
return DAG.getConstant(Value, ValueType);
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// get_vec_i16imm - Test if this vector is a vector filled with the same value
|
|
|
|
/// and the value fits into a signed 16-bit constant, and if so, return the
|
|
|
|
/// constant
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue SPU::get_vec_i16imm(SDNode *N, SelectionDAG &DAG,
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT ValueType) {
|
2007-12-04 22:23:35 +00:00
|
|
|
if (ConstantSDNode *CN = getVecImm(N)) {
|
2008-03-05 23:02:02 +00:00
|
|
|
int64_t Value = CN->getSignExtended();
|
2008-03-06 04:02:54 +00:00
|
|
|
if (ValueType == MVT::i64) {
|
|
|
|
uint64_t UValue = CN->getValue();
|
|
|
|
uint32_t upper = uint32_t(UValue >> 32);
|
|
|
|
uint32_t lower = uint32_t(UValue);
|
|
|
|
if (upper != lower)
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2008-03-06 04:02:54 +00:00
|
|
|
Value = Value >> 32;
|
|
|
|
}
|
2008-03-05 23:02:02 +00:00
|
|
|
if (Value >= -(1 << 15) && Value <= ((1 << 15) - 1)) {
|
|
|
|
return DAG.getConstant(Value, ValueType);
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// get_vec_i10imm - Test if this vector is a vector filled with the same value
|
|
|
|
/// and the value fits into a signed 10-bit constant, and if so, return the
|
|
|
|
/// constant
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue SPU::get_vec_i10imm(SDNode *N, SelectionDAG &DAG,
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT ValueType) {
|
2007-12-04 22:23:35 +00:00
|
|
|
if (ConstantSDNode *CN = getVecImm(N)) {
|
2008-03-05 23:02:02 +00:00
|
|
|
int64_t Value = CN->getSignExtended();
|
2008-03-06 04:02:54 +00:00
|
|
|
if (ValueType == MVT::i64) {
|
|
|
|
uint64_t UValue = CN->getValue();
|
|
|
|
uint32_t upper = uint32_t(UValue >> 32);
|
|
|
|
uint32_t lower = uint32_t(UValue);
|
|
|
|
if (upper != lower)
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2008-03-06 04:02:54 +00:00
|
|
|
Value = Value >> 32;
|
|
|
|
}
|
2008-03-05 23:02:02 +00:00
|
|
|
if (isS10Constant(Value))
|
2007-12-04 22:23:35 +00:00
|
|
|
return DAG.getConstant(Value, ValueType);
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// get_vec_i8imm - Test if this vector is a vector filled with the same value
|
|
|
|
/// and the value fits into a signed 8-bit constant, and if so, return the
|
|
|
|
/// constant.
|
|
|
|
///
|
|
|
|
/// @note: The incoming vector is v16i8 because that's the only way we can load
|
|
|
|
/// constant vectors. Thus, we test to see if the upper and lower bytes are the
|
|
|
|
/// same value.
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue SPU::get_vec_i8imm(SDNode *N, SelectionDAG &DAG,
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT ValueType) {
|
2007-12-04 22:23:35 +00:00
|
|
|
if (ConstantSDNode *CN = getVecImm(N)) {
|
|
|
|
int Value = (int) CN->getValue();
|
|
|
|
if (ValueType == MVT::i16
|
2008-01-30 02:55:46 +00:00
|
|
|
&& Value <= 0xffff /* truncated from uint64_t */
|
|
|
|
&& ((short) Value >> 8) == ((short) Value & 0xff))
|
2007-12-04 22:23:35 +00:00
|
|
|
return DAG.getConstant(Value & 0xff, ValueType);
|
|
|
|
else if (ValueType == MVT::i8
|
2008-01-30 02:55:46 +00:00
|
|
|
&& (Value & 0xff) == Value)
|
2007-12-04 22:23:35 +00:00
|
|
|
return DAG.getConstant(Value, ValueType);
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// get_ILHUvec_imm - Test if this vector is a vector filled with the same value
|
|
|
|
/// and the value fits into a signed 16-bit constant, and if so, return the
|
|
|
|
/// constant
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue SPU::get_ILHUvec_imm(SDNode *N, SelectionDAG &DAG,
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT ValueType) {
|
2007-12-04 22:23:35 +00:00
|
|
|
if (ConstantSDNode *CN = getVecImm(N)) {
|
|
|
|
uint64_t Value = CN->getValue();
|
|
|
|
if ((ValueType == MVT::i32
|
2008-01-30 02:55:46 +00:00
|
|
|
&& ((unsigned) Value & 0xffff0000) == (unsigned) Value)
|
|
|
|
|| (ValueType == MVT::i64 && (Value & 0xffff0000) == Value))
|
2007-12-04 22:23:35 +00:00
|
|
|
return DAG.getConstant(Value >> 16, ValueType);
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// get_v4i32_imm - Catch-all for general 32-bit constant vectors
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue SPU::get_v4i32_imm(SDNode *N, SelectionDAG &DAG) {
|
2007-12-04 22:23:35 +00:00
|
|
|
if (ConstantSDNode *CN = getVecImm(N)) {
|
|
|
|
return DAG.getConstant((unsigned) CN->getValue(), MVT::i32);
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// get_v4i32_imm - Catch-all for general 64-bit constant vectors
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue SPU::get_v2i64_imm(SDNode *N, SelectionDAG &DAG) {
|
2007-12-04 22:23:35 +00:00
|
|
|
if (ConstantSDNode *CN = getVecImm(N)) {
|
|
|
|
return DAG.getConstant((unsigned) CN->getValue(), MVT::i64);
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// If this is a vector of constants or undefs, get the bits. A bit in
|
2008-07-16 17:17:29 +00:00
|
|
|
// UndefBits is set if the corresponding element of the vector is an
|
2007-12-04 22:23:35 +00:00
|
|
|
// ISD::UNDEF value. For undefs, the corresponding VectorBits values are
|
|
|
|
// zero. Return true if this is not an array of constants, false if it is.
|
|
|
|
//
|
|
|
|
static bool GetConstantBuildVectorBits(SDNode *BV, uint64_t VectorBits[2],
|
|
|
|
uint64_t UndefBits[2]) {
|
|
|
|
// Start with zero'd results.
|
|
|
|
VectorBits[0] = VectorBits[1] = UndefBits[0] = UndefBits[1] = 0;
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-06-06 12:08:01 +00:00
|
|
|
unsigned EltBitSize = BV->getOperand(0).getValueType().getSizeInBits();
|
2007-12-04 22:23:35 +00:00
|
|
|
for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue OpVal = BV->getOperand(i);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
unsigned PartNo = i >= e/2; // In the upper 128 bits?
|
|
|
|
unsigned SlotNo = e/2 - (i & (e/2-1))-1; // Which subpiece of the uint64_t.
|
|
|
|
|
|
|
|
uint64_t EltBits = 0;
|
|
|
|
if (OpVal.getOpcode() == ISD::UNDEF) {
|
|
|
|
uint64_t EltUndefBits = ~0ULL >> (64-EltBitSize);
|
|
|
|
UndefBits[PartNo] |= EltUndefBits << (SlotNo*EltBitSize);
|
|
|
|
continue;
|
|
|
|
} else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
|
|
|
|
EltBits = CN->getValue() & (~0ULL >> (64-EltBitSize));
|
|
|
|
} else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
|
|
|
|
const APFloat &apf = CN->getValueAPF();
|
|
|
|
EltBits = (CN->getValueType(0) == MVT::f32
|
2008-01-30 02:55:46 +00:00
|
|
|
? FloatToBits(apf.convertToFloat())
|
|
|
|
: DoubleToBits(apf.convertToDouble()));
|
2007-12-04 22:23:35 +00:00
|
|
|
} else {
|
|
|
|
// Nonconstant element.
|
|
|
|
return true;
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
VectorBits[PartNo] |= EltBits << (SlotNo*EltBitSize);
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
|
|
|
//printf("%llx %llx %llx %llx\n",
|
2007-12-04 22:23:35 +00:00
|
|
|
// VectorBits[0], VectorBits[1], UndefBits[0], UndefBits[1]);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// If this is a splat (repetition) of a value across the whole vector, return
|
|
|
|
/// the smallest size that splats it. For example, "0x01010101010101..." is a
|
2008-07-16 17:17:29 +00:00
|
|
|
/// splat of 0x01, 0x0101, and 0x01010101. We return SplatBits = 0x01 and
|
2007-12-04 22:23:35 +00:00
|
|
|
/// SplatSize = 1 byte.
|
2008-07-16 17:17:29 +00:00
|
|
|
static bool isConstantSplat(const uint64_t Bits128[2],
|
2007-12-04 22:23:35 +00:00
|
|
|
const uint64_t Undef128[2],
|
2008-01-30 02:55:46 +00:00
|
|
|
int MinSplatBits,
|
2007-12-04 22:23:35 +00:00
|
|
|
uint64_t &SplatBits, uint64_t &SplatUndef,
|
|
|
|
int &SplatSize) {
|
|
|
|
// Don't let undefs prevent splats from matching. See if the top 64-bits are
|
|
|
|
// the same as the lower 64-bits, ignoring undefs.
|
|
|
|
uint64_t Bits64 = Bits128[0] | Bits128[1];
|
|
|
|
uint64_t Undef64 = Undef128[0] & Undef128[1];
|
|
|
|
uint32_t Bits32 = uint32_t(Bits64) | uint32_t(Bits64 >> 32);
|
|
|
|
uint32_t Undef32 = uint32_t(Undef64) & uint32_t(Undef64 >> 32);
|
|
|
|
uint16_t Bits16 = uint16_t(Bits32) | uint16_t(Bits32 >> 16);
|
|
|
|
uint16_t Undef16 = uint16_t(Undef32) & uint16_t(Undef32 >> 16);
|
|
|
|
|
|
|
|
if ((Bits128[0] & ~Undef128[1]) == (Bits128[1] & ~Undef128[0])) {
|
|
|
|
if (MinSplatBits < 64) {
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Check that the top 32-bits are the same as the lower 32-bits, ignoring
|
|
|
|
// undefs.
|
|
|
|
if ((Bits64 & (~Undef64 >> 32)) == ((Bits64 >> 32) & ~Undef64)) {
|
2008-01-30 02:55:46 +00:00
|
|
|
if (MinSplatBits < 32) {
|
|
|
|
|
|
|
|
// If the top 16-bits are different than the lower 16-bits, ignoring
|
|
|
|
// undefs, we have an i32 splat.
|
|
|
|
if ((Bits32 & (~Undef32 >> 16)) == ((Bits32 >> 16) & ~Undef32)) {
|
|
|
|
if (MinSplatBits < 16) {
|
|
|
|
// If the top 8-bits are different than the lower 8-bits, ignoring
|
|
|
|
// undefs, we have an i16 splat.
|
|
|
|
if ((Bits16 & (uint16_t(~Undef16) >> 8)) == ((Bits16 >> 8) & ~Undef16)) {
|
|
|
|
// Otherwise, we have an 8-bit splat.
|
|
|
|
SplatBits = uint8_t(Bits16) | uint8_t(Bits16 >> 8);
|
|
|
|
SplatUndef = uint8_t(Undef16) & uint8_t(Undef16 >> 8);
|
|
|
|
SplatSize = 1;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
SplatBits = Bits16;
|
|
|
|
SplatUndef = Undef16;
|
|
|
|
SplatSize = 2;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
SplatBits = Bits32;
|
|
|
|
SplatUndef = Undef32;
|
|
|
|
SplatSize = 4;
|
|
|
|
return true;
|
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
} else {
|
|
|
|
SplatBits = Bits128[0];
|
|
|
|
SplatUndef = Undef128[0];
|
|
|
|
SplatSize = 8;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return false; // Can't be a splat if two pieces don't match.
|
|
|
|
}
|
|
|
|
|
|
|
|
// If this is a case we can't handle, return null and let the default
|
|
|
|
// expansion code take care of it. If we CAN select this case, and if it
|
|
|
|
// selects to a single instruction, return Op. Otherwise, if we can codegen
|
|
|
|
// this case more efficiently than a constant pool load, lower it to the
|
|
|
|
// sequence of ops that should be used.
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Op.getValueType();
|
2007-12-04 22:23:35 +00:00
|
|
|
// If this is a vector of constants or undefs, get the bits. A bit in
|
2008-07-16 17:17:29 +00:00
|
|
|
// UndefBits is set if the corresponding element of the vector is an
|
2007-12-04 22:23:35 +00:00
|
|
|
// ISD::UNDEF value. For undefs, the corresponding VectorBits values are
|
2008-07-16 17:17:29 +00:00
|
|
|
// zero.
|
2007-12-04 22:23:35 +00:00
|
|
|
uint64_t VectorBits[2];
|
|
|
|
uint64_t UndefBits[2];
|
|
|
|
uint64_t SplatBits, SplatUndef;
|
|
|
|
int SplatSize;
|
2008-08-28 21:40:38 +00:00
|
|
|
if (GetConstantBuildVectorBits(Op.getNode(), VectorBits, UndefBits)
|
2007-12-04 22:23:35 +00:00
|
|
|
|| !isConstantSplat(VectorBits, UndefBits,
|
2008-06-06 12:08:01 +00:00
|
|
|
VT.getVectorElementType().getSizeInBits(),
|
2007-12-04 22:23:35 +00:00
|
|
|
SplatBits, SplatUndef, SplatSize))
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue(); // Not a constant vector, not a splat.
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-06-06 12:08:01 +00:00
|
|
|
switch (VT.getSimpleVT()) {
|
2007-12-04 22:23:35 +00:00
|
|
|
default:
|
|
|
|
case MVT::v4f32: {
|
|
|
|
uint32_t Value32 = SplatBits;
|
|
|
|
assert(SplatSize == 4
|
2008-01-30 02:55:46 +00:00
|
|
|
&& "LowerBUILD_VECTOR: Unexpected floating point vector element.");
|
2007-12-04 22:23:35 +00:00
|
|
|
// NOTE: pretend the constant is an integer. LLVM won't load FP constants
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue T = DAG.getConstant(Value32, MVT::i32);
|
2007-12-04 22:23:35 +00:00
|
|
|
return DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, T, T, T, T));
|
2007-12-04 22:23:35 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
case MVT::v2f64: {
|
|
|
|
uint64_t f64val = SplatBits;
|
|
|
|
assert(SplatSize == 8
|
2008-01-30 02:55:46 +00:00
|
|
|
&& "LowerBUILD_VECTOR: 64-bit float vector element: unexpected size.");
|
2007-12-04 22:23:35 +00:00
|
|
|
// NOTE: pretend the constant is an integer. LLVM won't load FP constants
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue T = DAG.getConstant(f64val, MVT::i64);
|
2007-12-04 22:23:35 +00:00
|
|
|
return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f64,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::BUILD_VECTOR, MVT::v2i64, T, T));
|
2007-12-04 22:23:35 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
case MVT::v16i8: {
|
|
|
|
// 8-bit constants have to be expanded to 16-bits
|
|
|
|
unsigned short Value16 = SplatBits | (SplatBits << 8);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Ops[8];
|
2007-12-04 22:23:35 +00:00
|
|
|
for (int i = 0; i < 8; ++i)
|
|
|
|
Ops[i] = DAG.getConstant(Value16, MVT::i16);
|
|
|
|
return DAG.getNode(ISD::BIT_CONVERT, VT,
|
|
|
|
DAG.getNode(ISD::BUILD_VECTOR, MVT::v8i16, Ops, 8));
|
|
|
|
}
|
|
|
|
case MVT::v8i16: {
|
|
|
|
unsigned short Value16;
|
2008-07-16 17:17:29 +00:00
|
|
|
if (SplatSize == 2)
|
2007-12-04 22:23:35 +00:00
|
|
|
Value16 = (unsigned short) (SplatBits & 0xffff);
|
|
|
|
else
|
|
|
|
Value16 = (unsigned short) (SplatBits | (SplatBits << 8));
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue T = DAG.getConstant(Value16, VT.getVectorElementType());
|
|
|
|
SDValue Ops[8];
|
2007-12-04 22:23:35 +00:00
|
|
|
for (int i = 0; i < 8; ++i) Ops[i] = T;
|
|
|
|
return DAG.getNode(ISD::BUILD_VECTOR, VT, Ops, 8);
|
|
|
|
}
|
|
|
|
case MVT::v4i32: {
|
|
|
|
unsigned int Value = SplatBits;
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue T = DAG.getConstant(Value, VT.getVectorElementType());
|
2007-12-04 22:23:35 +00:00
|
|
|
return DAG.getNode(ISD::BUILD_VECTOR, VT, T, T, T, T);
|
|
|
|
}
|
|
|
|
case MVT::v2i64: {
|
|
|
|
uint64_t val = SplatBits;
|
|
|
|
uint32_t upper = uint32_t(val >> 32);
|
|
|
|
uint32_t lower = uint32_t(val);
|
|
|
|
|
2008-03-06 04:02:54 +00:00
|
|
|
if (upper == lower) {
|
|
|
|
// Magic constant that can be matched by IL, ILA, et. al.
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Val = DAG.getTargetConstant(val, MVT::i64);
|
2008-03-06 04:02:54 +00:00
|
|
|
return DAG.getNode(ISD::BUILD_VECTOR, VT, Val, Val);
|
2008-03-05 23:02:02 +00:00
|
|
|
} else {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue LO32;
|
|
|
|
SDValue HI32;
|
|
|
|
SmallVector<SDValue, 16> ShufBytes;
|
|
|
|
SDValue Result;
|
2007-12-04 22:23:35 +00:00
|
|
|
bool upper_special, lower_special;
|
|
|
|
|
|
|
|
// NOTE: This code creates common-case shuffle masks that can be easily
|
|
|
|
// detected as common expressions. It is not attempting to create highly
|
|
|
|
// specialized masks to replace any and all 0's, 0xff's and 0x80's.
|
|
|
|
|
|
|
|
// Detect if the upper or lower half is a special shuffle mask pattern:
|
|
|
|
upper_special = (upper == 0 || upper == 0xffffffff || upper == 0x80000000);
|
|
|
|
lower_special = (lower == 0 || lower == 0xffffffff || lower == 0x80000000);
|
|
|
|
|
|
|
|
// Create lower vector if not a special pattern
|
|
|
|
if (!lower_special) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue LO32C = DAG.getConstant(lower, MVT::i32);
|
2008-01-30 02:55:46 +00:00
|
|
|
LO32 = DAG.getNode(ISD::BIT_CONVERT, VT,
|
|
|
|
DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
|
|
|
|
LO32C, LO32C, LO32C, LO32C));
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// Create upper vector if not a special pattern
|
|
|
|
if (!upper_special) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue HI32C = DAG.getConstant(upper, MVT::i32);
|
2008-01-30 02:55:46 +00:00
|
|
|
HI32 = DAG.getNode(ISD::BIT_CONVERT, VT,
|
|
|
|
DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
|
|
|
|
HI32C, HI32C, HI32C, HI32C));
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// If either upper or lower are special, then the two input operands are
|
|
|
|
// the same (basically, one of them is a "don't care")
|
|
|
|
if (lower_special)
|
2008-01-30 02:55:46 +00:00
|
|
|
LO32 = HI32;
|
2007-12-04 22:23:35 +00:00
|
|
|
if (upper_special)
|
2008-01-30 02:55:46 +00:00
|
|
|
HI32 = LO32;
|
2007-12-04 22:23:35 +00:00
|
|
|
if (lower_special && upper_special) {
|
2008-01-30 02:55:46 +00:00
|
|
|
// Unhappy situation... both upper and lower are special, so punt with
|
|
|
|
// a target constant:
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Zero = DAG.getConstant(0, MVT::i32);
|
2008-01-30 02:55:46 +00:00
|
|
|
HI32 = LO32 = DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32, Zero, Zero,
|
2007-12-04 22:23:35 +00:00
|
|
|
Zero, Zero);
|
|
|
|
}
|
|
|
|
|
|
|
|
for (int i = 0; i < 4; ++i) {
|
2008-06-02 22:18:03 +00:00
|
|
|
uint64_t val = 0;
|
2008-01-30 02:55:46 +00:00
|
|
|
for (int j = 0; j < 4; ++j) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue V;
|
2008-01-30 02:55:46 +00:00
|
|
|
bool process_upper, process_lower;
|
2008-06-02 22:18:03 +00:00
|
|
|
val <<= 8;
|
2008-01-30 02:55:46 +00:00
|
|
|
process_upper = (upper_special && (i & 1) == 0);
|
|
|
|
process_lower = (lower_special && (i & 1) == 1);
|
|
|
|
|
|
|
|
if (process_upper || process_lower) {
|
|
|
|
if ((process_upper && upper == 0)
|
|
|
|
|| (process_lower && lower == 0))
|
2008-06-02 22:18:03 +00:00
|
|
|
val |= 0x80;
|
2008-01-30 02:55:46 +00:00
|
|
|
else if ((process_upper && upper == 0xffffffff)
|
|
|
|
|| (process_lower && lower == 0xffffffff))
|
2008-06-02 22:18:03 +00:00
|
|
|
val |= 0xc0;
|
2008-01-30 02:55:46 +00:00
|
|
|
else if ((process_upper && upper == 0x80000000)
|
|
|
|
|| (process_lower && lower == 0x80000000))
|
2008-06-02 22:18:03 +00:00
|
|
|
val |= (j == 0 ? 0xe0 : 0x80);
|
2008-01-30 02:55:46 +00:00
|
|
|
} else
|
2008-06-02 22:18:03 +00:00
|
|
|
val |= i * 4 + j + ((i & 1) * 16);
|
2008-01-30 02:55:46 +00:00
|
|
|
}
|
2008-06-02 22:18:03 +00:00
|
|
|
|
|
|
|
ShufBytes.push_back(DAG.getConstant(val, MVT::i32));
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
return DAG.getNode(SPUISD::SHUFB, VT, HI32, LO32,
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
|
2008-01-30 02:55:46 +00:00
|
|
|
&ShufBytes[0], ShufBytes.size()));
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// LowerVECTOR_SHUFFLE - Lower a vector shuffle (V1, V2, V3) to something on
|
|
|
|
/// which the Cell can operate. The code inspects V3 to ascertain whether the
|
|
|
|
/// permutation vector, V3, is monotonically increasing with one "exception"
|
|
|
|
/// element, e.g., (0, 1, _, 3). If this is the case, then generate a
|
|
|
|
/// INSERT_MASK synthetic instruction. Otherwise, spill V3 to the constant pool.
|
|
|
|
/// In either case, the net result is going to eventually invoke SHUFB to
|
|
|
|
/// permute/shuffle the bytes from V1 and V2.
|
|
|
|
/// \note
|
|
|
|
/// INSERT_MASK is eventually selected as one of the C*D instructions, generate
|
|
|
|
/// control word for byte/halfword/word insertion. This takes care of a single
|
|
|
|
/// element move from V2 into V1.
|
|
|
|
/// \note
|
|
|
|
/// SPUISD::SHUFB is eventually selected as Cell's <i>shufb</i> instructions.
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
|
|
|
|
SDValue V1 = Op.getOperand(0);
|
|
|
|
SDValue V2 = Op.getOperand(1);
|
|
|
|
SDValue PermMask = Op.getOperand(2);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// If we have a single element being moved from V1 to V2, this can be handled
|
|
|
|
// using the C*[DX] compute mask instructions, but the vector elements have
|
|
|
|
// to be monotonically increasing with one exception element.
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT EltVT = V1.getValueType().getVectorElementType();
|
2007-12-04 22:23:35 +00:00
|
|
|
unsigned EltsFromV2 = 0;
|
|
|
|
unsigned V2Elt = 0;
|
|
|
|
unsigned V2EltIdx0 = 0;
|
|
|
|
unsigned CurrElt = 0;
|
|
|
|
bool monotonic = true;
|
|
|
|
if (EltVT == MVT::i8)
|
|
|
|
V2EltIdx0 = 16;
|
|
|
|
else if (EltVT == MVT::i16)
|
|
|
|
V2EltIdx0 = 8;
|
|
|
|
else if (EltVT == MVT::i32)
|
|
|
|
V2EltIdx0 = 4;
|
|
|
|
else
|
|
|
|
assert(0 && "Unhandled vector type in LowerVECTOR_SHUFFLE");
|
|
|
|
|
|
|
|
for (unsigned i = 0, e = PermMask.getNumOperands();
|
|
|
|
EltsFromV2 <= 1 && monotonic && i != e;
|
|
|
|
++i) {
|
|
|
|
unsigned SrcElt;
|
|
|
|
if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
|
|
|
|
SrcElt = 0;
|
2008-07-16 17:17:29 +00:00
|
|
|
else
|
2007-12-04 22:23:35 +00:00
|
|
|
SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
|
|
|
|
|
|
|
|
if (SrcElt >= V2EltIdx0) {
|
|
|
|
++EltsFromV2;
|
|
|
|
V2Elt = (V2EltIdx0 - SrcElt) << 2;
|
|
|
|
} else if (CurrElt != SrcElt) {
|
|
|
|
monotonic = false;
|
|
|
|
}
|
|
|
|
|
|
|
|
++CurrElt;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (EltsFromV2 == 1 && monotonic) {
|
|
|
|
// Compute mask and shuffle
|
|
|
|
MachineFunction &MF = DAG.getMachineFunction();
|
2007-12-31 04:13:23 +00:00
|
|
|
MachineRegisterInfo &RegInfo = MF.getRegInfo();
|
|
|
|
unsigned VReg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
|
2007-12-04 22:23:35 +00:00
|
|
|
// Initialize temporary register to 0
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue InitTempReg =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getCopyToReg(DAG.getEntryNode(), VReg, DAG.getConstant(0, PtrVT));
|
|
|
|
// Copy register's contents as index in INSERT_MASK:
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ShufMaskOp =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(SPUISD::INSERT_MASK, V1.getValueType(),
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getTargetConstant(V2Elt, MVT::i32),
|
|
|
|
DAG.getCopyFromReg(InitTempReg, VReg, PtrVT));
|
2007-12-04 22:23:35 +00:00
|
|
|
// Use shuffle mask in SHUFB synthetic instruction:
|
|
|
|
return DAG.getNode(SPUISD::SHUFB, V1.getValueType(), V2, V1, ShufMaskOp);
|
|
|
|
} else {
|
|
|
|
// Convert the SHUFFLE_VECTOR mask's input element units to the actual bytes.
|
2008-06-06 12:08:01 +00:00
|
|
|
unsigned BytesPerElement = EltVT.getSizeInBits()/8;
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SmallVector<SDValue, 16> ResultMask;
|
2007-12-04 22:23:35 +00:00
|
|
|
for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
|
|
|
|
unsigned SrcElt;
|
|
|
|
if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
|
2008-01-30 02:55:46 +00:00
|
|
|
SrcElt = 0;
|
2008-07-16 17:17:29 +00:00
|
|
|
else
|
2008-01-30 02:55:46 +00:00
|
|
|
SrcElt = cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-02-23 18:41:37 +00:00
|
|
|
for (unsigned j = 0; j < BytesPerElement; ++j) {
|
2008-01-30 02:55:46 +00:00
|
|
|
ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
|
|
|
|
MVT::i8));
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
|
2008-01-30 02:55:46 +00:00
|
|
|
&ResultMask[0], ResultMask.size());
|
2007-12-04 22:23:35 +00:00
|
|
|
return DAG.getNode(SPUISD::SHUFB, V1.getValueType(), V1, V2, VPermMask);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
|
|
|
|
SDValue Op0 = Op.getOperand(0); // Op0 = the scalar
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-08-28 21:40:38 +00:00
|
|
|
if (Op0.getNode()->getOpcode() == ISD::Constant) {
|
2007-12-04 22:23:35 +00:00
|
|
|
// For a constant, build the appropriate constant vector, which will
|
|
|
|
// eventually simplify to a vector register load.
|
|
|
|
|
2008-08-28 21:40:38 +00:00
|
|
|
ConstantSDNode *CN = cast<ConstantSDNode>(Op0.getNode());
|
2008-07-27 21:46:04 +00:00
|
|
|
SmallVector<SDValue, 16> ConstVecValues;
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT;
|
2007-12-04 22:23:35 +00:00
|
|
|
size_t n_copies;
|
|
|
|
|
|
|
|
// Create a constant vector:
|
2008-06-06 12:08:01 +00:00
|
|
|
switch (Op.getValueType().getSimpleVT()) {
|
2007-12-04 22:23:35 +00:00
|
|
|
default: assert(0 && "Unexpected constant value type in "
|
2008-01-30 02:55:46 +00:00
|
|
|
"LowerSCALAR_TO_VECTOR");
|
2007-12-04 22:23:35 +00:00
|
|
|
case MVT::v16i8: n_copies = 16; VT = MVT::i8; break;
|
|
|
|
case MVT::v8i16: n_copies = 8; VT = MVT::i16; break;
|
|
|
|
case MVT::v4i32: n_copies = 4; VT = MVT::i32; break;
|
|
|
|
case MVT::v4f32: n_copies = 4; VT = MVT::f32; break;
|
|
|
|
case MVT::v2i64: n_copies = 2; VT = MVT::i64; break;
|
|
|
|
case MVT::v2f64: n_copies = 2; VT = MVT::f64; break;
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue CValue = DAG.getConstant(CN->getValue(), VT);
|
2007-12-04 22:23:35 +00:00
|
|
|
for (size_t j = 0; j < n_copies; ++j)
|
|
|
|
ConstVecValues.push_back(CValue);
|
|
|
|
|
|
|
|
return DAG.getNode(ISD::BUILD_VECTOR, Op.getValueType(),
|
2008-01-30 02:55:46 +00:00
|
|
|
&ConstVecValues[0], ConstVecValues.size());
|
2007-12-04 22:23:35 +00:00
|
|
|
} else {
|
|
|
|
// Otherwise, copy the value from one register to another:
|
2008-06-06 12:08:01 +00:00
|
|
|
switch (Op0.getValueType().getSimpleVT()) {
|
2007-12-04 22:23:35 +00:00
|
|
|
default: assert(0 && "Unexpected value type in LowerSCALAR_TO_VECTOR");
|
|
|
|
case MVT::i8:
|
|
|
|
case MVT::i16:
|
|
|
|
case MVT::i32:
|
|
|
|
case MVT::i64:
|
|
|
|
case MVT::f32:
|
|
|
|
case MVT::f64:
|
|
|
|
return DAG.getNode(SPUISD::PROMOTE_SCALAR, Op.getValueType(), Op0, Op0);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue LowerVectorMUL(SDValue Op, SelectionDAG &DAG) {
|
2008-06-06 12:08:01 +00:00
|
|
|
switch (Op.getValueType().getSimpleVT()) {
|
|
|
|
default:
|
|
|
|
cerr << "CellSPU: Unknown vector multiplication, got "
|
|
|
|
<< Op.getValueType().getMVTString()
|
|
|
|
<< "\n";
|
|
|
|
abort();
|
|
|
|
/*NOTREACHED*/
|
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
case MVT::v4i32: {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue rA = Op.getOperand(0);
|
|
|
|
SDValue rB = Op.getOperand(1);
|
|
|
|
SDValue HiProd1 = DAG.getNode(SPUISD::MPYH, MVT::v4i32, rA, rB);
|
|
|
|
SDValue HiProd2 = DAG.getNode(SPUISD::MPYH, MVT::v4i32, rB, rA);
|
|
|
|
SDValue LoProd = DAG.getNode(SPUISD::MPYU, MVT::v4i32, rA, rB);
|
|
|
|
SDValue Residual1 = DAG.getNode(ISD::ADD, MVT::v4i32, LoProd, HiProd1);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
return DAG.getNode(ISD::ADD, MVT::v4i32, Residual1, HiProd2);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Multiply two v8i16 vectors (pipeline friendly version):
|
|
|
|
// a) multiply lower halves, mask off upper 16-bit of 32-bit product
|
|
|
|
// b) multiply upper halves, rotate left by 16 bits (inserts 16 lower zeroes)
|
|
|
|
// c) Use SELB to select upper and lower halves from the intermediate results
|
|
|
|
//
|
2008-06-02 22:18:03 +00:00
|
|
|
// NOTE: We really want to move the SELECT_MASK to earlier to actually get the
|
2007-12-04 22:23:35 +00:00
|
|
|
// dual-issue. This code does manage to do this, even if it's a little on
|
|
|
|
// the wacky side
|
|
|
|
case MVT::v8i16: {
|
|
|
|
MachineFunction &MF = DAG.getMachineFunction();
|
2007-12-31 04:13:23 +00:00
|
|
|
MachineRegisterInfo &RegInfo = MF.getRegInfo();
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Chain = Op.getOperand(0);
|
|
|
|
SDValue rA = Op.getOperand(0);
|
|
|
|
SDValue rB = Op.getOperand(1);
|
2007-12-31 04:13:23 +00:00
|
|
|
unsigned FSMBIreg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
|
|
|
|
unsigned HiProdReg = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue FSMBOp =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getCopyToReg(Chain, FSMBIreg,
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::SELECT_MASK, MVT::v8i16,
|
2008-04-30 00:30:08 +00:00
|
|
|
DAG.getConstant(0xcccc, MVT::i16)));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue HHProd =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getCopyToReg(FSMBOp, HiProdReg,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(SPUISD::MPYHH, MVT::v8i16, rA, rB));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue HHProd_v4i32 =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getCopyFromReg(HHProd, HiProdReg, MVT::v4i32));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
return DAG.getNode(SPUISD::SELB, MVT::v8i16,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(SPUISD::MPY, MVT::v8i16, rA, rB),
|
|
|
|
DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
|
|
|
|
DAG.getNode(SPUISD::VEC_SHL, MVT::v4i32,
|
|
|
|
HHProd_v4i32,
|
|
|
|
DAG.getConstant(16, MVT::i16))),
|
|
|
|
DAG.getCopyFromReg(FSMBOp, FSMBIreg, MVT::v4i32));
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
// This M00sE is N@stI! (apologies to Monty Python)
|
|
|
|
//
|
|
|
|
// SPU doesn't know how to do any 8-bit multiplication, so the solution
|
|
|
|
// is to break it all apart, sign extend, and reassemble the various
|
|
|
|
// intermediate products.
|
|
|
|
case MVT::v16i8: {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue rA = Op.getOperand(0);
|
|
|
|
SDValue rB = Op.getOperand(1);
|
|
|
|
SDValue c8 = DAG.getConstant(8, MVT::i32);
|
|
|
|
SDValue c16 = DAG.getConstant(16, MVT::i32);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue LLProd =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(SPUISD::MPY, MVT::v8i16,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rA),
|
|
|
|
DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rB));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue rALH = DAG.getNode(SPUISD::VEC_SRA, MVT::v8i16, rA, c8);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue rBLH = DAG.getNode(SPUISD::VEC_SRA, MVT::v8i16, rB, c8);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue LHProd =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(SPUISD::VEC_SHL, MVT::v8i16,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(SPUISD::MPY, MVT::v8i16, rALH, rBLH), c8);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue FSMBmask = DAG.getNode(SPUISD::SELECT_MASK, MVT::v8i16,
|
2008-04-30 00:30:08 +00:00
|
|
|
DAG.getConstant(0x2222, MVT::i16));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue LoProdParts =
|
2008-02-23 18:41:37 +00:00
|
|
|
DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32,
|
|
|
|
DAG.getNode(SPUISD::SELB, MVT::v8i16,
|
|
|
|
LLProd, LHProd, FSMBmask));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue LoProdMask = DAG.getConstant(0xffff, MVT::i32);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue LoProd =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(ISD::AND, MVT::v4i32,
|
2008-02-23 18:41:37 +00:00
|
|
|
LoProdParts,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
|
|
|
|
LoProdMask, LoProdMask,
|
|
|
|
LoProdMask, LoProdMask));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue rAH =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, rA), c16);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue rBH =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, rB), c16);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue HLProd =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(SPUISD::MPY, MVT::v8i16,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rAH),
|
|
|
|
DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16, rBH));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue HHProd_1 =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(SPUISD::MPY, MVT::v8i16,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16,
|
|
|
|
DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32, rAH, c8)),
|
|
|
|
DAG.getNode(ISD::BIT_CONVERT, MVT::v8i16,
|
|
|
|
DAG.getNode(SPUISD::VEC_SRA, MVT::v4i32, rBH, c8)));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue HHProd =
|
2008-02-23 18:41:37 +00:00
|
|
|
DAG.getNode(SPUISD::SELB, MVT::v8i16,
|
|
|
|
HLProd,
|
|
|
|
DAG.getNode(SPUISD::VEC_SHL, MVT::v8i16, HHProd_1, c8),
|
|
|
|
FSMBmask);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue HiProd =
|
2008-02-23 18:41:37 +00:00
|
|
|
DAG.getNode(SPUISD::VEC_SHL, MVT::v4i32, HHProd, c16);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
return DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::OR, MVT::v4i32,
|
|
|
|
LoProd, HiProd));
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue LowerFDIVf32(SDValue Op, SelectionDAG &DAG) {
|
2007-12-04 22:23:35 +00:00
|
|
|
MachineFunction &MF = DAG.getMachineFunction();
|
2007-12-31 04:13:23 +00:00
|
|
|
MachineRegisterInfo &RegInfo = MF.getRegInfo();
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue A = Op.getOperand(0);
|
|
|
|
SDValue B = Op.getOperand(1);
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Op.getValueType();
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
unsigned VRegBR, VRegC;
|
|
|
|
|
|
|
|
if (VT == MVT::f32) {
|
2007-12-31 04:13:23 +00:00
|
|
|
VRegBR = RegInfo.createVirtualRegister(&SPU::R32FPRegClass);
|
|
|
|
VRegC = RegInfo.createVirtualRegister(&SPU::R32FPRegClass);
|
2007-12-04 22:23:35 +00:00
|
|
|
} else {
|
2007-12-31 04:13:23 +00:00
|
|
|
VRegBR = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
|
|
|
|
VRegC = RegInfo.createVirtualRegister(&SPU::VECREGRegClass);
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
// TODO: make sure we're feeding FPInterp the right arguments
|
|
|
|
// Right now: fi B, frest(B)
|
|
|
|
|
|
|
|
// Computes BRcpl =
|
|
|
|
// (Floating Interpolate (FP Reciprocal Estimate B))
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue BRcpl =
|
2008-07-16 17:17:29 +00:00
|
|
|
DAG.getCopyToReg(DAG.getEntryNode(), VRegBR,
|
|
|
|
DAG.getNode(SPUISD::FPInterp, VT, B,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(SPUISD::FPRecipEst, VT, B)));
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
// Computes A * BRcpl and stores in a temporary register
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue AxBRcpl =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getCopyToReg(BRcpl, VRegC,
|
2008-07-16 17:17:29 +00:00
|
|
|
DAG.getNode(ISD::FMUL, VT, A,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getCopyFromReg(BRcpl, VRegBR, VT)));
|
2007-12-04 22:23:35 +00:00
|
|
|
// What's the Chain variable do? It's magic!
|
|
|
|
// TODO: set Chain = Op(0).getEntryNode()
|
2008-07-16 17:17:29 +00:00
|
|
|
|
|
|
|
return DAG.getNode(ISD::FADD, VT,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getCopyFromReg(AxBRcpl, VRegC, VT),
|
2008-07-16 17:17:29 +00:00
|
|
|
DAG.getNode(ISD::FMUL, VT,
|
|
|
|
DAG.getCopyFromReg(AxBRcpl, VRegBR, VT),
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::FSUB, VT, A,
|
2008-07-16 17:17:29 +00:00
|
|
|
DAG.getNode(ISD::FMUL, VT, B,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getCopyFromReg(AxBRcpl, VRegC, VT)))));
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Op.getValueType();
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue N = Op.getOperand(0);
|
|
|
|
SDValue Elt = Op.getOperand(1);
|
|
|
|
SDValue ShufMask[16];
|
2007-12-04 22:23:35 +00:00
|
|
|
ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt);
|
|
|
|
|
|
|
|
assert(C != 0 && "LowerEXTRACT_VECTOR_ELT expecting constant SDNode");
|
|
|
|
|
|
|
|
int EltNo = (int) C->getValue();
|
|
|
|
|
|
|
|
// sanity checks:
|
|
|
|
if (VT == MVT::i8 && EltNo >= 16)
|
|
|
|
assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i8 extraction slot > 15");
|
|
|
|
else if (VT == MVT::i16 && EltNo >= 8)
|
|
|
|
assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i16 extraction slot > 7");
|
|
|
|
else if (VT == MVT::i32 && EltNo >= 4)
|
|
|
|
assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i32 extraction slot > 4");
|
|
|
|
else if (VT == MVT::i64 && EltNo >= 2)
|
|
|
|
assert(0 && "SPU LowerEXTRACT_VECTOR_ELT: i64 extraction slot > 2");
|
|
|
|
|
|
|
|
if (EltNo == 0 && (VT == MVT::i32 || VT == MVT::i64)) {
|
|
|
|
// i32 and i64: Element 0 is the preferred slot
|
|
|
|
return DAG.getNode(SPUISD::EXTRACT_ELT0, VT, N);
|
|
|
|
}
|
|
|
|
|
|
|
|
// Need to generate shuffle mask and extract:
|
2007-12-19 21:17:42 +00:00
|
|
|
int prefslot_begin = -1, prefslot_end = -1;
|
2008-06-06 12:08:01 +00:00
|
|
|
int elt_byte = EltNo * VT.getSizeInBits() / 8;
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-06-06 12:08:01 +00:00
|
|
|
switch (VT.getSimpleVT()) {
|
|
|
|
default:
|
|
|
|
assert(false && "Invalid value type!");
|
2007-12-04 22:23:35 +00:00
|
|
|
case MVT::i8: {
|
|
|
|
prefslot_begin = prefslot_end = 3;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case MVT::i16: {
|
|
|
|
prefslot_begin = 2; prefslot_end = 3;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case MVT::i32: {
|
|
|
|
prefslot_begin = 0; prefslot_end = 3;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case MVT::i64: {
|
|
|
|
prefslot_begin = 0; prefslot_end = 7;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2007-12-19 21:17:42 +00:00
|
|
|
assert(prefslot_begin != -1 && prefslot_end != -1 &&
|
2008-01-30 02:55:46 +00:00
|
|
|
"LowerEXTRACT_VECTOR_ELT: preferred slots uninitialized");
|
2007-12-19 21:17:42 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
for (int i = 0; i < 16; ++i) {
|
|
|
|
// zero fill uppper part of preferred slot, don't care about the
|
|
|
|
// other slots:
|
|
|
|
unsigned int mask_val;
|
|
|
|
|
|
|
|
if (i <= prefslot_end) {
|
|
|
|
mask_val =
|
2008-01-30 02:55:46 +00:00
|
|
|
((i < prefslot_begin)
|
|
|
|
? 0x80
|
|
|
|
: elt_byte + (i - prefslot_begin));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2007-12-19 21:17:42 +00:00
|
|
|
ShufMask[i] = DAG.getConstant(mask_val, MVT::i8);
|
2008-07-16 17:17:29 +00:00
|
|
|
} else
|
2007-12-04 22:23:35 +00:00
|
|
|
ShufMask[i] = ShufMask[i % (prefslot_end + 1)];
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ShufMaskVec =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
|
2008-01-30 02:55:46 +00:00
|
|
|
&ShufMask[0],
|
|
|
|
sizeof(ShufMask) / sizeof(ShufMask[0]));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
return DAG.getNode(SPUISD::EXTRACT_ELT0, VT,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(SPUISD::SHUFB, N.getValueType(),
|
|
|
|
N, N, ShufMaskVec));
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
|
|
|
|
SDValue VecOp = Op.getOperand(0);
|
|
|
|
SDValue ValOp = Op.getOperand(1);
|
|
|
|
SDValue IdxOp = Op.getOperand(2);
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Op.getValueType();
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
ConstantSDNode *CN = cast<ConstantSDNode>(IdxOp);
|
|
|
|
assert(CN != 0 && "LowerINSERT_VECTOR_ELT: Index is not constant!");
|
|
|
|
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
|
2007-12-04 22:23:35 +00:00
|
|
|
// Use $2 because it's always 16-byte aligned and it's available:
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue PtrBase = DAG.getRegister(SPU::R2, PtrVT);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue result =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(SPUISD::SHUFB, VT,
|
|
|
|
DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, ValOp),
|
|
|
|
VecOp,
|
|
|
|
DAG.getNode(SPUISD::INSERT_MASK, VT,
|
|
|
|
DAG.getNode(ISD::ADD, PtrVT,
|
|
|
|
PtrBase,
|
|
|
|
DAG.getConstant(CN->getValue(),
|
2008-01-30 02:55:46 +00:00
|
|
|
PtrVT))));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue LowerI8Math(SDValue Op, SelectionDAG &DAG, unsigned Opc)
|
2008-02-23 18:41:37 +00:00
|
|
|
{
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue N0 = Op.getOperand(0); // Everything has at least one operand
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
assert(Op.getValueType() == MVT::i8);
|
|
|
|
switch (Opc) {
|
|
|
|
default:
|
|
|
|
assert(0 && "Unhandled i8 math operator");
|
|
|
|
/*NOTREACHED*/
|
|
|
|
break;
|
|
|
|
case ISD::SUB: {
|
|
|
|
// 8-bit subtraction: Promote the arguments up to 16-bits and truncate
|
|
|
|
// the result:
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue N1 = Op.getOperand(1);
|
2007-12-04 22:23:35 +00:00
|
|
|
N0 = (N0.getOpcode() != ISD::Constant
|
|
|
|
? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N0)
|
|
|
|
: DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
|
|
|
|
N1 = (N1.getOpcode() != ISD::Constant
|
|
|
|
? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N1)
|
|
|
|
: DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
|
2008-07-16 17:17:29 +00:00
|
|
|
return DAG.getNode(ISD::TRUNCATE, MVT::i8,
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(Opc, MVT::i16, N0, N1));
|
2008-07-16 17:17:29 +00:00
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
case ISD::ROTR:
|
|
|
|
case ISD::ROTL: {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue N1 = Op.getOperand(1);
|
2007-12-04 22:23:35 +00:00
|
|
|
unsigned N1Opc;
|
|
|
|
N0 = (N0.getOpcode() != ISD::Constant
|
|
|
|
? DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, N0)
|
|
|
|
: DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
|
2008-06-08 20:54:56 +00:00
|
|
|
N1Opc = N1.getValueType().bitsLT(MVT::i16) ? ISD::ZERO_EXTEND : ISD::TRUNCATE;
|
2007-12-04 22:23:35 +00:00
|
|
|
N1 = (N1.getOpcode() != ISD::Constant
|
|
|
|
? DAG.getNode(N1Opc, MVT::i16, N1)
|
|
|
|
: DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ExpandArg =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(ISD::OR, MVT::i16, N0,
|
|
|
|
DAG.getNode(ISD::SHL, MVT::i16,
|
|
|
|
N0, DAG.getConstant(8, MVT::i16)));
|
2008-07-16 17:17:29 +00:00
|
|
|
return DAG.getNode(ISD::TRUNCATE, MVT::i8,
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(Opc, MVT::i16, ExpandArg, N1));
|
|
|
|
}
|
|
|
|
case ISD::SRL:
|
|
|
|
case ISD::SHL: {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue N1 = Op.getOperand(1);
|
2007-12-04 22:23:35 +00:00
|
|
|
unsigned N1Opc;
|
|
|
|
N0 = (N0.getOpcode() != ISD::Constant
|
|
|
|
? DAG.getNode(ISD::ZERO_EXTEND, MVT::i16, N0)
|
|
|
|
: DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
|
2008-06-08 20:54:56 +00:00
|
|
|
N1Opc = N1.getValueType().bitsLT(MVT::i16) ? ISD::ZERO_EXTEND : ISD::TRUNCATE;
|
2007-12-04 22:23:35 +00:00
|
|
|
N1 = (N1.getOpcode() != ISD::Constant
|
|
|
|
? DAG.getNode(N1Opc, MVT::i16, N1)
|
|
|
|
: DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
|
2008-07-16 17:17:29 +00:00
|
|
|
return DAG.getNode(ISD::TRUNCATE, MVT::i8,
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(Opc, MVT::i16, N0, N1));
|
|
|
|
}
|
|
|
|
case ISD::SRA: {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue N1 = Op.getOperand(1);
|
2007-12-04 22:23:35 +00:00
|
|
|
unsigned N1Opc;
|
|
|
|
N0 = (N0.getOpcode() != ISD::Constant
|
|
|
|
? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N0)
|
|
|
|
: DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
|
2008-06-08 20:54:56 +00:00
|
|
|
N1Opc = N1.getValueType().bitsLT(MVT::i16) ? ISD::SIGN_EXTEND : ISD::TRUNCATE;
|
2007-12-04 22:23:35 +00:00
|
|
|
N1 = (N1.getOpcode() != ISD::Constant
|
|
|
|
? DAG.getNode(N1Opc, MVT::i16, N1)
|
|
|
|
: DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
|
2008-07-16 17:17:29 +00:00
|
|
|
return DAG.getNode(ISD::TRUNCATE, MVT::i8,
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(Opc, MVT::i16, N0, N1));
|
|
|
|
}
|
|
|
|
case ISD::MUL: {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue N1 = Op.getOperand(1);
|
2007-12-04 22:23:35 +00:00
|
|
|
unsigned N1Opc;
|
|
|
|
N0 = (N0.getOpcode() != ISD::Constant
|
|
|
|
? DAG.getNode(ISD::SIGN_EXTEND, MVT::i16, N0)
|
|
|
|
: DAG.getConstant(cast<ConstantSDNode>(N0)->getValue(), MVT::i16));
|
2008-06-08 20:54:56 +00:00
|
|
|
N1Opc = N1.getValueType().bitsLT(MVT::i16) ? ISD::SIGN_EXTEND : ISD::TRUNCATE;
|
2007-12-04 22:23:35 +00:00
|
|
|
N1 = (N1.getOpcode() != ISD::Constant
|
|
|
|
? DAG.getNode(N1Opc, MVT::i16, N1)
|
|
|
|
: DAG.getConstant(cast<ConstantSDNode>(N1)->getValue(), MVT::i16));
|
2008-07-16 17:17:29 +00:00
|
|
|
return DAG.getNode(ISD::TRUNCATE, MVT::i8,
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(Opc, MVT::i16, N0, N1));
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue LowerI64Math(SDValue Op, SelectionDAG &DAG, unsigned Opc)
|
2008-02-23 18:41:37 +00:00
|
|
|
{
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Op.getValueType();
|
|
|
|
MVT VecVT = MVT::getVectorVT(VT, (128 / VT.getSizeInBits()));
|
2008-02-23 18:41:37 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op0 = Op.getOperand(0);
|
2008-02-23 18:41:37 +00:00
|
|
|
|
|
|
|
switch (Opc) {
|
|
|
|
case ISD::ZERO_EXTEND:
|
|
|
|
case ISD::SIGN_EXTEND:
|
|
|
|
case ISD::ANY_EXTEND: {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT Op0VT = Op0.getValueType();
|
|
|
|
MVT Op0VecVT = MVT::getVectorVT(Op0VT, (128 / Op0VT.getSizeInBits()));
|
2008-02-23 18:41:37 +00:00
|
|
|
|
|
|
|
assert(Op0VT == MVT::i32
|
|
|
|
&& "CellSPU: Zero/sign extending something other than i32");
|
2008-04-30 00:30:08 +00:00
|
|
|
DEBUG(cerr << "CellSPU: LowerI64Math custom lowering zero/sign/any extend\n");
|
2008-02-23 18:41:37 +00:00
|
|
|
|
|
|
|
unsigned NewOpc = (Opc == ISD::SIGN_EXTEND
|
|
|
|
? SPUISD::ROTBYTES_RIGHT_S
|
|
|
|
: SPUISD::ROTQUAD_RZ_BYTES);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue PromoteScalar =
|
2008-02-23 18:41:37 +00:00
|
|
|
DAG.getNode(SPUISD::PROMOTE_SCALAR, Op0VecVT, Op0);
|
|
|
|
|
|
|
|
return DAG.getNode(SPUISD::EXTRACT_ELT0, VT,
|
|
|
|
DAG.getNode(ISD::BIT_CONVERT, VecVT,
|
|
|
|
DAG.getNode(NewOpc, Op0VecVT,
|
|
|
|
PromoteScalar,
|
|
|
|
DAG.getConstant(4, MVT::i32))));
|
|
|
|
}
|
|
|
|
|
2008-06-02 22:18:03 +00:00
|
|
|
case ISD::ADD: {
|
|
|
|
// Turn operands into vectors to satisfy type checking (shufb works on
|
|
|
|
// vectors)
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op0 =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::PROMOTE_SCALAR, MVT::v2i64, Op.getOperand(0));
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op1 =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::PROMOTE_SCALAR, MVT::v2i64, Op.getOperand(1));
|
2008-07-27 21:46:04 +00:00
|
|
|
SmallVector<SDValue, 16> ShufBytes;
|
2008-06-02 22:18:03 +00:00
|
|
|
|
|
|
|
// Create the shuffle mask for "rotating" the borrow up one register slot
|
|
|
|
// once the borrow is generated.
|
|
|
|
ShufBytes.push_back(DAG.getConstant(0x04050607, MVT::i32));
|
|
|
|
ShufBytes.push_back(DAG.getConstant(0x80808080, MVT::i32));
|
|
|
|
ShufBytes.push_back(DAG.getConstant(0x0c0d0e0f, MVT::i32));
|
|
|
|
ShufBytes.push_back(DAG.getConstant(0x80808080, MVT::i32));
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue CarryGen =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::CARRY_GENERATE, MVT::v2i64, Op0, Op1);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ShiftedCarry =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::SHUFB, MVT::v2i64,
|
|
|
|
CarryGen, CarryGen,
|
|
|
|
DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
|
|
|
|
&ShufBytes[0], ShufBytes.size()));
|
|
|
|
|
|
|
|
return DAG.getNode(SPUISD::EXTRACT_ELT0, MVT::i64,
|
|
|
|
DAG.getNode(SPUISD::ADD_EXTENDED, MVT::v2i64,
|
|
|
|
Op0, Op1, ShiftedCarry));
|
|
|
|
}
|
|
|
|
|
|
|
|
case ISD::SUB: {
|
|
|
|
// Turn operands into vectors to satisfy type checking (shufb works on
|
|
|
|
// vectors)
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op0 =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::PROMOTE_SCALAR, MVT::v2i64, Op.getOperand(0));
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op1 =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::PROMOTE_SCALAR, MVT::v2i64, Op.getOperand(1));
|
2008-07-27 21:46:04 +00:00
|
|
|
SmallVector<SDValue, 16> ShufBytes;
|
2008-06-02 22:18:03 +00:00
|
|
|
|
|
|
|
// Create the shuffle mask for "rotating" the borrow up one register slot
|
|
|
|
// once the borrow is generated.
|
|
|
|
ShufBytes.push_back(DAG.getConstant(0x04050607, MVT::i32));
|
|
|
|
ShufBytes.push_back(DAG.getConstant(0xc0c0c0c0, MVT::i32));
|
|
|
|
ShufBytes.push_back(DAG.getConstant(0x0c0d0e0f, MVT::i32));
|
|
|
|
ShufBytes.push_back(DAG.getConstant(0xc0c0c0c0, MVT::i32));
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue BorrowGen =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::BORROW_GENERATE, MVT::v2i64, Op0, Op1);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ShiftedBorrow =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::SHUFB, MVT::v2i64,
|
|
|
|
BorrowGen, BorrowGen,
|
|
|
|
DAG.getNode(ISD::BUILD_VECTOR, MVT::v4i32,
|
|
|
|
&ShufBytes[0], ShufBytes.size()));
|
|
|
|
|
|
|
|
return DAG.getNode(SPUISD::EXTRACT_ELT0, MVT::i64,
|
|
|
|
DAG.getNode(SPUISD::SUB_EXTENDED, MVT::v2i64,
|
|
|
|
Op0, Op1, ShiftedBorrow));
|
|
|
|
}
|
|
|
|
|
2008-02-23 18:41:37 +00:00
|
|
|
case ISD::SHL: {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ShiftAmt = Op.getOperand(1);
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT ShiftAmtVT = ShiftAmt.getValueType();
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op0Vec = DAG.getNode(SPUISD::PROMOTE_SCALAR, VecVT, Op0);
|
|
|
|
SDValue MaskLower =
|
2008-02-23 18:41:37 +00:00
|
|
|
DAG.getNode(SPUISD::SELB, VecVT,
|
|
|
|
Op0Vec,
|
|
|
|
DAG.getConstant(0, VecVT),
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::SELECT_MASK, VecVT,
|
2008-02-23 18:41:37 +00:00
|
|
|
DAG.getConstant(0xff00ULL, MVT::i16)));
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ShiftAmtBytes =
|
2008-02-23 18:41:37 +00:00
|
|
|
DAG.getNode(ISD::SRL, ShiftAmtVT,
|
|
|
|
ShiftAmt,
|
|
|
|
DAG.getConstant(3, ShiftAmtVT));
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ShiftAmtBits =
|
2008-02-23 18:41:37 +00:00
|
|
|
DAG.getNode(ISD::AND, ShiftAmtVT,
|
|
|
|
ShiftAmt,
|
|
|
|
DAG.getConstant(7, ShiftAmtVT));
|
|
|
|
|
|
|
|
return DAG.getNode(SPUISD::EXTRACT_ELT0, VT,
|
|
|
|
DAG.getNode(SPUISD::SHLQUAD_L_BITS, VecVT,
|
|
|
|
DAG.getNode(SPUISD::SHLQUAD_L_BYTES, VecVT,
|
|
|
|
MaskLower, ShiftAmtBytes),
|
|
|
|
ShiftAmtBits));
|
|
|
|
}
|
|
|
|
|
|
|
|
case ISD::SRL: {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Op.getValueType();
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ShiftAmt = Op.getOperand(1);
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT ShiftAmtVT = ShiftAmt.getValueType();
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ShiftAmtBytes =
|
2008-02-23 18:41:37 +00:00
|
|
|
DAG.getNode(ISD::SRL, ShiftAmtVT,
|
|
|
|
ShiftAmt,
|
|
|
|
DAG.getConstant(3, ShiftAmtVT));
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ShiftAmtBits =
|
2008-02-23 18:41:37 +00:00
|
|
|
DAG.getNode(ISD::AND, ShiftAmtVT,
|
|
|
|
ShiftAmt,
|
|
|
|
DAG.getConstant(7, ShiftAmtVT));
|
|
|
|
|
|
|
|
return DAG.getNode(SPUISD::ROTQUAD_RZ_BITS, VT,
|
|
|
|
DAG.getNode(SPUISD::ROTQUAD_RZ_BYTES, VT,
|
|
|
|
Op0, ShiftAmtBytes),
|
|
|
|
ShiftAmtBits);
|
|
|
|
}
|
2008-06-02 22:18:03 +00:00
|
|
|
|
|
|
|
case ISD::SRA: {
|
|
|
|
// Promote Op0 to vector
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op0 =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::PROMOTE_SCALAR, MVT::v2i64, Op.getOperand(0));
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue ShiftAmt = Op.getOperand(1);
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT ShiftVT = ShiftAmt.getValueType();
|
2008-06-02 22:18:03 +00:00
|
|
|
|
|
|
|
// Negate variable shift amounts
|
|
|
|
if (!isa<ConstantSDNode>(ShiftAmt)) {
|
|
|
|
ShiftAmt = DAG.getNode(ISD::SUB, ShiftVT,
|
|
|
|
DAG.getConstant(0, ShiftVT), ShiftAmt);
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue UpperHalfSign =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::EXTRACT_ELT0, MVT::i32,
|
|
|
|
DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32,
|
|
|
|
DAG.getNode(SPUISD::VEC_SRA, MVT::v2i64,
|
|
|
|
Op0, DAG.getConstant(31, MVT::i32))));
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue UpperHalfSignMask =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::SELECT_MASK, MVT::v2i64, UpperHalfSign);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue UpperLowerMask =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::SELECT_MASK, MVT::v2i64,
|
|
|
|
DAG.getConstant(0xff00, MVT::i16));
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue UpperLowerSelect =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::SELB, MVT::v2i64,
|
|
|
|
UpperHalfSignMask, Op0, UpperLowerMask);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue RotateLeftBytes =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::ROTBYTES_LEFT_BITS, MVT::v2i64,
|
|
|
|
UpperLowerSelect, ShiftAmt);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue RotateLeftBits =
|
2008-06-02 22:18:03 +00:00
|
|
|
DAG.getNode(SPUISD::ROTBYTES_LEFT, MVT::v2i64,
|
|
|
|
RotateLeftBytes, ShiftAmt);
|
|
|
|
|
|
|
|
return DAG.getNode(SPUISD::EXTRACT_ELT0, MVT::i64,
|
|
|
|
RotateLeftBits);
|
|
|
|
}
|
2008-02-23 18:41:37 +00:00
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2008-02-23 18:41:37 +00:00
|
|
|
}
|
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
//! Lower byte immediate operations for v16i8 vectors:
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue
|
|
|
|
LowerByteImmed(SDValue Op, SelectionDAG &DAG) {
|
|
|
|
SDValue ConstVec;
|
|
|
|
SDValue Arg;
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Op.getValueType();
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
ConstVec = Op.getOperand(0);
|
|
|
|
Arg = Op.getOperand(1);
|
2008-08-28 21:40:38 +00:00
|
|
|
if (ConstVec.getNode()->getOpcode() != ISD::BUILD_VECTOR) {
|
|
|
|
if (ConstVec.getNode()->getOpcode() == ISD::BIT_CONVERT) {
|
2007-12-04 22:23:35 +00:00
|
|
|
ConstVec = ConstVec.getOperand(0);
|
|
|
|
} else {
|
|
|
|
ConstVec = Op.getOperand(1);
|
|
|
|
Arg = Op.getOperand(0);
|
2008-08-28 21:40:38 +00:00
|
|
|
if (ConstVec.getNode()->getOpcode() == ISD::BIT_CONVERT) {
|
2008-01-30 02:55:46 +00:00
|
|
|
ConstVec = ConstVec.getOperand(0);
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-08-28 21:40:38 +00:00
|
|
|
if (ConstVec.getNode()->getOpcode() == ISD::BUILD_VECTOR) {
|
2007-12-04 22:23:35 +00:00
|
|
|
uint64_t VectorBits[2];
|
|
|
|
uint64_t UndefBits[2];
|
|
|
|
uint64_t SplatBits, SplatUndef;
|
|
|
|
int SplatSize;
|
|
|
|
|
2008-08-28 21:40:38 +00:00
|
|
|
if (!GetConstantBuildVectorBits(ConstVec.getNode(), VectorBits, UndefBits)
|
2008-01-30 02:55:46 +00:00
|
|
|
&& isConstantSplat(VectorBits, UndefBits,
|
2008-06-06 12:08:01 +00:00
|
|
|
VT.getVectorElementType().getSizeInBits(),
|
2008-01-30 02:55:46 +00:00
|
|
|
SplatBits, SplatUndef, SplatSize)) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue tcVec[16];
|
|
|
|
SDValue tc = DAG.getTargetConstant(SplatBits & 0xff, MVT::i8);
|
2007-12-04 22:23:35 +00:00
|
|
|
const size_t tcVecSize = sizeof(tcVec) / sizeof(tcVec[0]);
|
|
|
|
|
|
|
|
// Turn the BUILD_VECTOR into a set of target constants:
|
|
|
|
for (size_t i = 0; i < tcVecSize; ++i)
|
2008-01-30 02:55:46 +00:00
|
|
|
tcVec[i] = tc;
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-08-28 21:40:38 +00:00
|
|
|
return DAG.getNode(Op.getNode()->getOpcode(), VT, Arg,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::BUILD_VECTOR, VT, tcVec, tcVecSize));
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
}
|
2008-07-29 19:07:27 +00:00
|
|
|
// These operations (AND, OR, XOR) are legal, they just couldn't be custom
|
|
|
|
// lowered. Return the operation, rather than a null SDValue.
|
|
|
|
return Op;
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
//! Lower i32 multiplication
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG, MVT VT,
|
2007-12-04 22:23:35 +00:00
|
|
|
unsigned Opc) {
|
2008-06-06 12:08:01 +00:00
|
|
|
switch (VT.getSimpleVT()) {
|
2007-12-04 22:23:35 +00:00
|
|
|
default:
|
|
|
|
cerr << "CellSPU: Unknown LowerMUL value type, got "
|
2008-06-06 12:08:01 +00:00
|
|
|
<< Op.getValueType().getMVTString()
|
2008-01-30 02:55:46 +00:00
|
|
|
<< "\n";
|
2007-12-04 22:23:35 +00:00
|
|
|
abort();
|
|
|
|
/*NOTREACHED*/
|
|
|
|
|
|
|
|
case MVT::i32: {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue rA = Op.getOperand(0);
|
|
|
|
SDValue rB = Op.getOperand(1);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
return DAG.getNode(ISD::ADD, MVT::i32,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::ADD, MVT::i32,
|
|
|
|
DAG.getNode(SPUISD::MPYH, MVT::i32, rA, rB),
|
|
|
|
DAG.getNode(SPUISD::MPYH, MVT::i32, rB, rA)),
|
|
|
|
DAG.getNode(SPUISD::MPYU, MVT::i32, rA, rB));
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
//! Custom lowering for CTPOP (count population)
|
|
|
|
/*!
|
|
|
|
Custom lowering code that counts the number ones in the input
|
|
|
|
operand. SPU has such an instruction, but it counts the number of
|
|
|
|
ones per byte, which then have to be accumulated.
|
|
|
|
*/
|
2008-07-27 21:46:04 +00:00
|
|
|
static SDValue LowerCTPOP(SDValue Op, SelectionDAG &DAG) {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Op.getValueType();
|
|
|
|
MVT vecVT = MVT::getVectorVT(VT, (128 / VT.getSizeInBits()));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-06-06 12:08:01 +00:00
|
|
|
switch (VT.getSimpleVT()) {
|
|
|
|
default:
|
|
|
|
assert(false && "Invalid value type!");
|
2007-12-04 22:23:35 +00:00
|
|
|
case MVT::i8: {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue N = Op.getOperand(0);
|
|
|
|
SDValue Elt0 = DAG.getConstant(0, MVT::i32);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Promote = DAG.getNode(SPUISD::PROMOTE_SCALAR, vecVT, N, N);
|
|
|
|
SDValue CNTB = DAG.getNode(SPUISD::CNTB, vecVT, Promote);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i8, CNTB, Elt0);
|
|
|
|
}
|
|
|
|
|
|
|
|
case MVT::i16: {
|
|
|
|
MachineFunction &MF = DAG.getMachineFunction();
|
2007-12-31 04:13:23 +00:00
|
|
|
MachineRegisterInfo &RegInfo = MF.getRegInfo();
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2007-12-31 04:13:23 +00:00
|
|
|
unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R16CRegClass);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue N = Op.getOperand(0);
|
|
|
|
SDValue Elt0 = DAG.getConstant(0, MVT::i16);
|
|
|
|
SDValue Mask0 = DAG.getConstant(0x0f, MVT::i16);
|
|
|
|
SDValue Shift1 = DAG.getConstant(8, MVT::i16);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Promote = DAG.getNode(SPUISD::PROMOTE_SCALAR, vecVT, N, N);
|
|
|
|
SDValue CNTB = DAG.getNode(SPUISD::CNTB, vecVT, Promote);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
// CNTB_result becomes the chain to which all of the virtual registers
|
|
|
|
// CNTB_reg, SUM1_reg become associated:
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue CNTB_result =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i16, CNTB, Elt0);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue CNTB_rescopy =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getCopyToReg(CNTB_result, CNTB_reg, CNTB_result);
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Tmp1 = DAG.getCopyFromReg(CNTB_rescopy, CNTB_reg, MVT::i16);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
return DAG.getNode(ISD::AND, MVT::i16,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getNode(ISD::ADD, MVT::i16,
|
|
|
|
DAG.getNode(ISD::SRL, MVT::i16,
|
|
|
|
Tmp1, Shift1),
|
|
|
|
Tmp1),
|
|
|
|
Mask0);
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
case MVT::i32: {
|
|
|
|
MachineFunction &MF = DAG.getMachineFunction();
|
2007-12-31 04:13:23 +00:00
|
|
|
MachineRegisterInfo &RegInfo = MF.getRegInfo();
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2007-12-31 04:13:23 +00:00
|
|
|
unsigned CNTB_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
|
|
|
|
unsigned SUM1_reg = RegInfo.createVirtualRegister(&SPU::R32CRegClass);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue N = Op.getOperand(0);
|
|
|
|
SDValue Elt0 = DAG.getConstant(0, MVT::i32);
|
|
|
|
SDValue Mask0 = DAG.getConstant(0xff, MVT::i32);
|
|
|
|
SDValue Shift1 = DAG.getConstant(16, MVT::i32);
|
|
|
|
SDValue Shift2 = DAG.getConstant(8, MVT::i32);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Promote = DAG.getNode(SPUISD::PROMOTE_SCALAR, vecVT, N, N);
|
|
|
|
SDValue CNTB = DAG.getNode(SPUISD::CNTB, vecVT, Promote);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
// CNTB_result becomes the chain to which all of the virtual registers
|
|
|
|
// CNTB_reg, SUM1_reg become associated:
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue CNTB_result =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32, CNTB, Elt0);
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue CNTB_rescopy =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getCopyToReg(CNTB_result, CNTB_reg, CNTB_result);
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Comp1 =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(ISD::SRL, MVT::i32,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getCopyFromReg(CNTB_rescopy, CNTB_reg, MVT::i32), Shift1);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Sum1 =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(ISD::ADD, MVT::i32,
|
2008-01-30 02:55:46 +00:00
|
|
|
Comp1, DAG.getCopyFromReg(CNTB_rescopy, CNTB_reg, MVT::i32));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Sum1_rescopy =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getCopyToReg(CNTB_result, SUM1_reg, Sum1);
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Comp2 =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(ISD::SRL, MVT::i32,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getCopyFromReg(Sum1_rescopy, SUM1_reg, MVT::i32),
|
|
|
|
Shift2);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Sum2 =
|
2007-12-04 22:23:35 +00:00
|
|
|
DAG.getNode(ISD::ADD, MVT::i32, Comp2,
|
2008-01-30 02:55:46 +00:00
|
|
|
DAG.getCopyFromReg(Sum1_rescopy, SUM1_reg, MVT::i32));
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
return DAG.getNode(ISD::AND, MVT::i32, Sum2, Mask0);
|
|
|
|
}
|
|
|
|
|
|
|
|
case MVT::i64:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/// LowerOperation - Provide custom lowering hooks for some operations.
|
|
|
|
///
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue
|
|
|
|
SPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
|
2007-12-04 22:23:35 +00:00
|
|
|
{
|
2008-02-23 18:41:37 +00:00
|
|
|
unsigned Opc = (unsigned) Op.getOpcode();
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT = Op.getValueType();
|
2008-02-23 18:41:37 +00:00
|
|
|
|
|
|
|
switch (Opc) {
|
2007-12-04 22:23:35 +00:00
|
|
|
default: {
|
|
|
|
cerr << "SPUTargetLowering::LowerOperation(): need to lower this!\n";
|
2008-02-23 18:41:37 +00:00
|
|
|
cerr << "Op.getOpcode() = " << Opc << "\n";
|
2008-08-28 21:40:38 +00:00
|
|
|
cerr << "*Op.getNode():\n";
|
|
|
|
Op.getNode()->dump();
|
2007-12-04 22:23:35 +00:00
|
|
|
abort();
|
|
|
|
}
|
|
|
|
case ISD::LOAD:
|
|
|
|
case ISD::SEXTLOAD:
|
|
|
|
case ISD::ZEXTLOAD:
|
|
|
|
return LowerLOAD(Op, DAG, SPUTM.getSubtargetImpl());
|
|
|
|
case ISD::STORE:
|
|
|
|
return LowerSTORE(Op, DAG, SPUTM.getSubtargetImpl());
|
|
|
|
case ISD::ConstantPool:
|
|
|
|
return LowerConstantPool(Op, DAG, SPUTM.getSubtargetImpl());
|
|
|
|
case ISD::GlobalAddress:
|
|
|
|
return LowerGlobalAddress(Op, DAG, SPUTM.getSubtargetImpl());
|
|
|
|
case ISD::JumpTable:
|
|
|
|
return LowerJumpTable(Op, DAG, SPUTM.getSubtargetImpl());
|
|
|
|
case ISD::Constant:
|
|
|
|
return LowerConstant(Op, DAG);
|
|
|
|
case ISD::ConstantFP:
|
|
|
|
return LowerConstantFP(Op, DAG);
|
2008-01-17 20:38:41 +00:00
|
|
|
case ISD::BRCOND:
|
|
|
|
return LowerBRCOND(Op, DAG);
|
2007-12-04 22:23:35 +00:00
|
|
|
case ISD::FORMAL_ARGUMENTS:
|
2008-01-17 20:38:41 +00:00
|
|
|
return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex);
|
2007-12-04 22:23:35 +00:00
|
|
|
case ISD::CALL:
|
2008-01-11 02:53:15 +00:00
|
|
|
return LowerCALL(Op, DAG, SPUTM.getSubtargetImpl());
|
2007-12-04 22:23:35 +00:00
|
|
|
case ISD::RET:
|
|
|
|
return LowerRET(Op, DAG, getTargetMachine());
|
|
|
|
|
2008-02-23 18:41:37 +00:00
|
|
|
|
|
|
|
// i8, i64 math ops:
|
|
|
|
case ISD::ZERO_EXTEND:
|
|
|
|
case ISD::SIGN_EXTEND:
|
|
|
|
case ISD::ANY_EXTEND:
|
2008-06-02 22:18:03 +00:00
|
|
|
case ISD::ADD:
|
2007-12-04 22:23:35 +00:00
|
|
|
case ISD::SUB:
|
|
|
|
case ISD::ROTR:
|
|
|
|
case ISD::ROTL:
|
|
|
|
case ISD::SRL:
|
|
|
|
case ISD::SHL:
|
2008-06-02 22:18:03 +00:00
|
|
|
case ISD::SRA: {
|
2008-02-23 18:41:37 +00:00
|
|
|
if (VT == MVT::i8)
|
|
|
|
return LowerI8Math(Op, DAG, Opc);
|
|
|
|
else if (VT == MVT::i64)
|
|
|
|
return LowerI64Math(Op, DAG, Opc);
|
|
|
|
break;
|
2008-06-02 22:18:03 +00:00
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
// Vector-related lowering.
|
|
|
|
case ISD::BUILD_VECTOR:
|
|
|
|
return LowerBUILD_VECTOR(Op, DAG);
|
|
|
|
case ISD::SCALAR_TO_VECTOR:
|
|
|
|
return LowerSCALAR_TO_VECTOR(Op, DAG);
|
|
|
|
case ISD::VECTOR_SHUFFLE:
|
|
|
|
return LowerVECTOR_SHUFFLE(Op, DAG);
|
|
|
|
case ISD::EXTRACT_VECTOR_ELT:
|
|
|
|
return LowerEXTRACT_VECTOR_ELT(Op, DAG);
|
|
|
|
case ISD::INSERT_VECTOR_ELT:
|
|
|
|
return LowerINSERT_VECTOR_ELT(Op, DAG);
|
|
|
|
|
|
|
|
// Look for ANDBI, ORBI and XORBI opportunities and lower appropriately:
|
|
|
|
case ISD::AND:
|
|
|
|
case ISD::OR:
|
|
|
|
case ISD::XOR:
|
|
|
|
return LowerByteImmed(Op, DAG);
|
|
|
|
|
|
|
|
// Vector and i8 multiply:
|
|
|
|
case ISD::MUL:
|
2008-06-06 12:08:01 +00:00
|
|
|
if (VT.isVector())
|
2007-12-04 22:23:35 +00:00
|
|
|
return LowerVectorMUL(Op, DAG);
|
2008-02-23 18:41:37 +00:00
|
|
|
else if (VT == MVT::i8)
|
|
|
|
return LowerI8Math(Op, DAG, Opc);
|
2007-12-04 22:23:35 +00:00
|
|
|
else
|
2008-02-23 18:41:37 +00:00
|
|
|
return LowerMUL(Op, DAG, VT, Opc);
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
case ISD::FDIV:
|
2008-02-23 18:41:37 +00:00
|
|
|
if (VT == MVT::f32 || VT == MVT::v4f32)
|
2007-12-04 22:23:35 +00:00
|
|
|
return LowerFDIVf32(Op, DAG);
|
|
|
|
// else if (Op.getValueType() == MVT::f64)
|
|
|
|
// return LowerFDIVf64(Op, DAG);
|
|
|
|
else
|
|
|
|
assert(0 && "Calling FDIV on unsupported MVT");
|
|
|
|
|
|
|
|
case ISD::CTPOP:
|
|
|
|
return LowerCTPOP(Op, DAG);
|
|
|
|
}
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
return SDValue();
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Target Optimization Hooks
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue
|
2007-12-04 22:23:35 +00:00
|
|
|
SPUTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const
|
|
|
|
{
|
|
|
|
#if 0
|
|
|
|
TargetMachine &TM = getTargetMachine();
|
2008-01-29 02:16:57 +00:00
|
|
|
#endif
|
|
|
|
const SPUSubtarget *ST = SPUTM.getSubtargetImpl();
|
2007-12-04 22:23:35 +00:00
|
|
|
SelectionDAG &DAG = DCI.DAG;
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op0 = N->getOperand(0); // everything has at least one operand
|
|
|
|
SDValue Result; // Initially, NULL result
|
2007-12-04 22:23:35 +00:00
|
|
|
|
|
|
|
switch (N->getOpcode()) {
|
|
|
|
default: break;
|
2008-01-29 02:16:57 +00:00
|
|
|
case ISD::ADD: {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op1 = N->getOperand(1);
|
2008-01-29 02:16:57 +00:00
|
|
|
|
2008-06-02 22:18:03 +00:00
|
|
|
if (isa<ConstantSDNode>(Op1) && Op0.getOpcode() == SPUISD::IndirectAddr) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op01 = Op0.getOperand(1);
|
2008-01-29 02:16:57 +00:00
|
|
|
if (Op01.getOpcode() == ISD::Constant
|
|
|
|
|| Op01.getOpcode() == ISD::TargetConstant) {
|
|
|
|
// (add <const>, (SPUindirect <arg>, <const>)) ->
|
|
|
|
// (SPUindirect <arg>, <const + const>)
|
|
|
|
ConstantSDNode *CN0 = cast<ConstantSDNode>(Op1);
|
|
|
|
ConstantSDNode *CN1 = cast<ConstantSDNode>(Op01);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue combinedConst =
|
2008-01-29 02:16:57 +00:00
|
|
|
DAG.getConstant(CN0->getValue() + CN1->getValue(),
|
|
|
|
Op0.getValueType());
|
|
|
|
|
|
|
|
DEBUG(cerr << "Replace: (add " << CN0->getValue() << ", "
|
|
|
|
<< "(SPUindirect <arg>, " << CN1->getValue() << "))\n");
|
|
|
|
DEBUG(cerr << "With: (SPUindirect <arg>, "
|
|
|
|
<< CN0->getValue() + CN1->getValue() << ")\n");
|
|
|
|
return DAG.getNode(SPUISD::IndirectAddr, Op0.getValueType(),
|
|
|
|
Op0.getOperand(0), combinedConst);
|
|
|
|
}
|
2008-06-02 22:18:03 +00:00
|
|
|
} else if (isa<ConstantSDNode>(Op0)
|
2008-01-29 02:16:57 +00:00
|
|
|
&& Op1.getOpcode() == SPUISD::IndirectAddr) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op11 = Op1.getOperand(1);
|
2008-01-29 02:16:57 +00:00
|
|
|
if (Op11.getOpcode() == ISD::Constant
|
|
|
|
|| Op11.getOpcode() == ISD::TargetConstant) {
|
|
|
|
// (add (SPUindirect <arg>, <const>), <const>) ->
|
|
|
|
// (SPUindirect <arg>, <const + const>)
|
|
|
|
ConstantSDNode *CN0 = cast<ConstantSDNode>(Op0);
|
|
|
|
ConstantSDNode *CN1 = cast<ConstantSDNode>(Op11);
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue combinedConst =
|
2008-01-29 02:16:57 +00:00
|
|
|
DAG.getConstant(CN0->getValue() + CN1->getValue(),
|
|
|
|
Op0.getValueType());
|
|
|
|
|
|
|
|
DEBUG(cerr << "Replace: (add " << CN0->getValue() << ", "
|
|
|
|
<< "(SPUindirect <arg>, " << CN1->getValue() << "))\n");
|
|
|
|
DEBUG(cerr << "With: (SPUindirect <arg>, "
|
|
|
|
<< CN0->getValue() + CN1->getValue() << ")\n");
|
|
|
|
|
|
|
|
return DAG.getNode(SPUISD::IndirectAddr, Op1.getValueType(),
|
|
|
|
Op1.getOperand(0), combinedConst);
|
|
|
|
}
|
|
|
|
}
|
2008-02-23 18:41:37 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
case ISD::SIGN_EXTEND:
|
|
|
|
case ISD::ZERO_EXTEND:
|
|
|
|
case ISD::ANY_EXTEND: {
|
|
|
|
if (Op0.getOpcode() == SPUISD::EXTRACT_ELT0 &&
|
|
|
|
N->getValueType(0) == Op0.getValueType()) {
|
|
|
|
// (any_extend (SPUextract_elt0 <arg>)) ->
|
|
|
|
// (SPUextract_elt0 <arg>)
|
|
|
|
// Types must match, however...
|
|
|
|
DEBUG(cerr << "Replace: ");
|
|
|
|
DEBUG(N->dump(&DAG));
|
|
|
|
DEBUG(cerr << "\nWith: ");
|
2008-08-28 21:40:38 +00:00
|
|
|
DEBUG(Op0.getNode()->dump(&DAG));
|
2008-02-23 18:41:37 +00:00
|
|
|
DEBUG(cerr << "\n");
|
|
|
|
|
|
|
|
return Op0;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case SPUISD::IndirectAddr: {
|
|
|
|
if (!ST->usingLargeMem() && Op0.getOpcode() == SPUISD::AFormAddr) {
|
|
|
|
ConstantSDNode *CN = cast<ConstantSDNode>(N->getOperand(1));
|
|
|
|
if (CN->getValue() == 0) {
|
|
|
|
// (SPUindirect (SPUaform <addr>, 0), 0) ->
|
|
|
|
// (SPUaform <addr>, 0)
|
|
|
|
|
|
|
|
DEBUG(cerr << "Replace: ");
|
|
|
|
DEBUG(N->dump(&DAG));
|
|
|
|
DEBUG(cerr << "\nWith: ");
|
2008-08-28 21:40:38 +00:00
|
|
|
DEBUG(Op0.getNode()->dump(&DAG));
|
2008-02-23 18:41:37 +00:00
|
|
|
DEBUG(cerr << "\n");
|
|
|
|
|
|
|
|
return Op0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case SPUISD::SHLQUAD_L_BITS:
|
|
|
|
case SPUISD::SHLQUAD_L_BYTES:
|
|
|
|
case SPUISD::VEC_SHL:
|
|
|
|
case SPUISD::VEC_SRL:
|
|
|
|
case SPUISD::VEC_SRA:
|
|
|
|
case SPUISD::ROTQUAD_RZ_BYTES:
|
|
|
|
case SPUISD::ROTQUAD_RZ_BITS: {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op1 = N->getOperand(1);
|
2008-02-23 18:41:37 +00:00
|
|
|
|
|
|
|
if (isa<ConstantSDNode>(Op1)) {
|
|
|
|
// Kill degenerate vector shifts:
|
|
|
|
ConstantSDNode *CN = cast<ConstantSDNode>(Op1);
|
|
|
|
|
|
|
|
if (CN->getValue() == 0) {
|
|
|
|
Result = Op0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case SPUISD::PROMOTE_SCALAR: {
|
|
|
|
switch (Op0.getOpcode()) {
|
|
|
|
default:
|
|
|
|
break;
|
|
|
|
case ISD::ANY_EXTEND:
|
|
|
|
case ISD::ZERO_EXTEND:
|
|
|
|
case ISD::SIGN_EXTEND: {
|
|
|
|
// (SPUpromote_scalar (any|sign|zero_extend (SPUextract_elt0 <arg>))) ->
|
|
|
|
// <arg>
|
|
|
|
// but only if the SPUpromote_scalar and <arg> types match.
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op00 = Op0.getOperand(0);
|
2008-02-23 18:41:37 +00:00
|
|
|
if (Op00.getOpcode() == SPUISD::EXTRACT_ELT0) {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op000 = Op00.getOperand(0);
|
2008-02-23 18:41:37 +00:00
|
|
|
if (Op000.getValueType() == N->getValueType(0)) {
|
|
|
|
Result = Op000;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case SPUISD::EXTRACT_ELT0: {
|
|
|
|
// (SPUpromote_scalar (SPUextract_elt0 <arg>)) ->
|
|
|
|
// <arg>
|
|
|
|
Result = Op0.getOperand(0);
|
|
|
|
break;
|
2008-07-16 17:17:29 +00:00
|
|
|
}
|
2008-02-23 18:41:37 +00:00
|
|
|
}
|
|
|
|
break;
|
2008-01-29 02:16:57 +00:00
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
2008-01-17 20:38:41 +00:00
|
|
|
// Otherwise, return unchanged.
|
2008-04-30 00:30:08 +00:00
|
|
|
#if 1
|
2008-08-28 21:40:38 +00:00
|
|
|
if (Result.getNode()) {
|
2008-02-23 18:41:37 +00:00
|
|
|
DEBUG(cerr << "\nReplace.SPU: ");
|
|
|
|
DEBUG(N->dump(&DAG));
|
|
|
|
DEBUG(cerr << "\nWith: ");
|
2008-08-28 21:40:38 +00:00
|
|
|
DEBUG(Result.getNode()->dump(&DAG));
|
2008-02-23 18:41:37 +00:00
|
|
|
DEBUG(cerr << "\n");
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
return Result;
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Inline Assembly Support
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
|
|
|
|
/// getConstraintType - Given a constraint letter, return the type of
|
|
|
|
/// constraint it is for this target.
|
2008-07-16 17:17:29 +00:00
|
|
|
SPUTargetLowering::ConstraintType
|
2007-12-04 22:23:35 +00:00
|
|
|
SPUTargetLowering::getConstraintType(const std::string &ConstraintLetter) const {
|
|
|
|
if (ConstraintLetter.size() == 1) {
|
|
|
|
switch (ConstraintLetter[0]) {
|
|
|
|
default: break;
|
|
|
|
case 'b':
|
|
|
|
case 'r':
|
|
|
|
case 'f':
|
|
|
|
case 'v':
|
|
|
|
case 'y':
|
|
|
|
return C_RegisterClass;
|
2008-07-16 17:17:29 +00:00
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
return TargetLowering::getConstraintType(ConstraintLetter);
|
|
|
|
}
|
|
|
|
|
2008-07-16 17:17:29 +00:00
|
|
|
std::pair<unsigned, const TargetRegisterClass*>
|
2007-12-04 22:23:35 +00:00
|
|
|
SPUTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT VT) const
|
2007-12-04 22:23:35 +00:00
|
|
|
{
|
|
|
|
if (Constraint.size() == 1) {
|
|
|
|
// GCC RS6000 Constraint Letters
|
|
|
|
switch (Constraint[0]) {
|
|
|
|
case 'b': // R1-R31
|
|
|
|
case 'r': // R0-R31
|
|
|
|
if (VT == MVT::i64)
|
|
|
|
return std::make_pair(0U, SPU::R64CRegisterClass);
|
|
|
|
return std::make_pair(0U, SPU::R32CRegisterClass);
|
|
|
|
case 'f':
|
|
|
|
if (VT == MVT::f32)
|
|
|
|
return std::make_pair(0U, SPU::R32FPRegisterClass);
|
|
|
|
else if (VT == MVT::f64)
|
|
|
|
return std::make_pair(0U, SPU::R64FPRegisterClass);
|
|
|
|
break;
|
2008-07-16 17:17:29 +00:00
|
|
|
case 'v':
|
2007-12-04 22:23:35 +00:00
|
|
|
return std::make_pair(0U, SPU::GPRCRegisterClass);
|
|
|
|
}
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
|
|
|
|
}
|
|
|
|
|
2008-02-23 18:41:37 +00:00
|
|
|
//! Compute used/known bits for a SPU operand
|
2007-12-04 22:23:35 +00:00
|
|
|
void
|
2008-07-27 21:46:04 +00:00
|
|
|
SPUTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
|
2008-02-13 22:28:48 +00:00
|
|
|
const APInt &Mask,
|
2008-07-16 17:17:29 +00:00
|
|
|
APInt &KnownZero,
|
2008-02-13 00:35:47 +00:00
|
|
|
APInt &KnownOne,
|
2008-01-30 02:55:46 +00:00
|
|
|
const SelectionDAG &DAG,
|
|
|
|
unsigned Depth ) const {
|
2008-04-30 00:30:08 +00:00
|
|
|
#if 0
|
2008-02-23 18:41:37 +00:00
|
|
|
const uint64_t uint64_sizebits = sizeof(uint64_t) * 8;
|
2008-04-30 00:30:08 +00:00
|
|
|
#endif
|
2008-02-23 18:41:37 +00:00
|
|
|
|
|
|
|
switch (Op.getOpcode()) {
|
|
|
|
default:
|
|
|
|
// KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
|
|
|
|
break;
|
|
|
|
|
|
|
|
#if 0
|
|
|
|
case CALL:
|
|
|
|
case SHUFB:
|
|
|
|
case INSERT_MASK:
|
|
|
|
case CNTB:
|
|
|
|
#endif
|
|
|
|
|
|
|
|
case SPUISD::PROMOTE_SCALAR: {
|
2008-07-27 21:46:04 +00:00
|
|
|
SDValue Op0 = Op.getOperand(0);
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT Op0VT = Op0.getValueType();
|
|
|
|
unsigned Op0VTBits = Op0VT.getSizeInBits();
|
|
|
|
uint64_t InMask = Op0VT.getIntegerVTBitMask();
|
2008-04-30 00:30:08 +00:00
|
|
|
KnownZero |= APInt(Op0VTBits, ~InMask, false);
|
|
|
|
KnownOne |= APInt(Op0VTBits, InMask, false);
|
2008-02-23 18:41:37 +00:00
|
|
|
break;
|
|
|
|
}
|
2008-07-16 17:17:29 +00:00
|
|
|
|
2008-02-23 18:41:37 +00:00
|
|
|
case SPUISD::LDRESULT:
|
|
|
|
case SPUISD::EXTRACT_ELT0:
|
|
|
|
case SPUISD::EXTRACT_ELT0_CHAINED: {
|
2008-06-06 12:08:01 +00:00
|
|
|
MVT OpVT = Op.getValueType();
|
|
|
|
unsigned OpVTBits = OpVT.getSizeInBits();
|
|
|
|
uint64_t InMask = OpVT.getIntegerVTBitMask();
|
2008-04-30 00:30:08 +00:00
|
|
|
KnownZero |= APInt(OpVTBits, ~InMask, false);
|
|
|
|
KnownOne |= APInt(OpVTBits, InMask, false);
|
2008-02-23 18:41:37 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
#if 0
|
|
|
|
case EXTRACT_I1_ZEXT:
|
|
|
|
case EXTRACT_I1_SEXT:
|
|
|
|
case EXTRACT_I8_ZEXT:
|
|
|
|
case EXTRACT_I8_SEXT:
|
|
|
|
case MPY:
|
|
|
|
case MPYU:
|
|
|
|
case MPYH:
|
|
|
|
case MPYHH:
|
2008-04-30 00:30:08 +00:00
|
|
|
case SPUISD::SHLQUAD_L_BITS:
|
|
|
|
case SPUISD::SHLQUAD_L_BYTES:
|
|
|
|
case SPUISD::VEC_SHL:
|
|
|
|
case SPUISD::VEC_SRL:
|
|
|
|
case SPUISD::VEC_SRA:
|
|
|
|
case SPUISD::VEC_ROTL:
|
|
|
|
case SPUISD::VEC_ROTR:
|
|
|
|
case SPUISD::ROTQUAD_RZ_BYTES:
|
|
|
|
case SPUISD::ROTQUAD_RZ_BITS:
|
|
|
|
case SPUISD::ROTBYTES_RIGHT_S:
|
|
|
|
case SPUISD::ROTBYTES_LEFT:
|
|
|
|
case SPUISD::ROTBYTES_LEFT_CHAINED:
|
2008-06-02 22:18:03 +00:00
|
|
|
case SPUISD::SELECT_MASK:
|
|
|
|
case SPUISD::SELB:
|
|
|
|
case SPUISD::FPInterp:
|
|
|
|
case SPUISD::FPRecipEst:
|
|
|
|
case SPUISD::SEXT32TO64:
|
2008-02-23 18:41:37 +00:00
|
|
|
#endif
|
|
|
|
}
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|
|
|
|
|
2008-04-30 00:30:08 +00:00
|
|
|
// LowerAsmOperandForConstraint
|
|
|
|
void
|
2008-07-27 21:46:04 +00:00
|
|
|
SPUTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
|
2008-04-30 00:30:08 +00:00
|
|
|
char ConstraintLetter,
|
2008-07-27 21:46:04 +00:00
|
|
|
std::vector<SDValue> &Ops,
|
2008-04-30 00:30:08 +00:00
|
|
|
SelectionDAG &DAG) const {
|
|
|
|
// Default, for the time being, to the base class handler
|
|
|
|
TargetLowering::LowerAsmOperandForConstraint(Op, ConstraintLetter, Ops, DAG);
|
|
|
|
}
|
|
|
|
|
2007-12-04 22:23:35 +00:00
|
|
|
/// isLegalAddressImmediate - Return true if the integer value can be used
|
|
|
|
/// as the offset of the target addressing mode.
|
|
|
|
bool SPUTargetLowering::isLegalAddressImmediate(int64_t V, const Type *Ty) const {
|
|
|
|
// SPU's addresses are 256K:
|
|
|
|
return (V > -(1 << 18) && V < (1 << 18) - 1);
|
|
|
|
}
|
|
|
|
|
|
|
|
bool SPUTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
|
2008-07-16 17:17:29 +00:00
|
|
|
return false;
|
2007-12-04 22:23:35 +00:00
|
|
|
}
|