Update to use the new MathExtras.h support for log2 computation.

Patch contributed by Jim Laskey!


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@22594 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Chris Lattner 2005-08-02 19:26:06 +00:00
parent 3b65576527
commit 0561b3ff9f
11 changed files with 108 additions and 188 deletions

View File

@ -16,6 +16,7 @@
#include "llvm/GlobalValue.h"
#include "llvm/Assembly/Writer.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Target/TargetLowering.h"
#include <iostream>
#include <set>
@ -45,15 +46,6 @@ static bool isAssociativeBinOp(unsigned Opcode) {
}
}
static unsigned ExactLog2(uint64_t Val) {
unsigned Count = 0;
while (Val != 1) {
Val >>= 1;
++Count;
}
return Count;
}
// isInvertibleForFree - Return true if there is no cost to emitting the logical
// inverse of this node.
static bool isInvertibleForFree(SDOperand N) {
@ -527,7 +519,7 @@ SDOperand SelectionDAG::getSetCC(ISD::CondCode Cond, MVT::ValueType VT,
// Perform the xform if the AND RHS is a single bit.
if ((AndRHS->getValue() & (AndRHS->getValue()-1)) == 0) {
return getNode(ISD::SRL, VT, N1,
getConstant(ExactLog2(AndRHS->getValue()),
getConstant(Log2_64(AndRHS->getValue()),
TLI.getShiftAmountTy()));
}
} else if (Cond == ISD::SETEQ && C2 == AndRHS->getValue()) {
@ -535,7 +527,7 @@ SDOperand SelectionDAG::getSetCC(ISD::CondCode Cond, MVT::ValueType VT,
// Perform the xform if C2 is a single bit.
if ((C2 & (C2-1)) == 0) {
return getNode(ISD::SRL, VT, N1,
getConstant(ExactLog2(C2),TLI.getShiftAmountTy()));
getConstant(Log2_64(C2),TLI.getShiftAmountTy()));
}
}
}
@ -956,7 +948,7 @@ SDOperand SelectionDAG::getNode(unsigned Opcode, MVT::ValueType VT,
// FIXME: Move this to the DAG combiner when it exists.
if ((C2 & C2-1) == 0) {
SDOperand ShAmt = getConstant(ExactLog2(C2), TLI.getShiftAmountTy());
SDOperand ShAmt = getConstant(Log2_64(C2), TLI.getShiftAmountTy());
return getNode(ISD::SHL, VT, N1, ShAmt);
}
break;
@ -974,7 +966,7 @@ SDOperand SelectionDAG::getNode(unsigned Opcode, MVT::ValueType VT,
case ISD::UDIV:
// FIXME: Move this to the DAG combiner when it exists.
if ((C2 & C2-1) == 0 && C2) {
SDOperand ShAmt = getConstant(ExactLog2(C2), TLI.getShiftAmountTy());
SDOperand ShAmt = getConstant(Log2_64(C2), TLI.getShiftAmountTy());
return getNode(ISD::SRL, VT, N1, ShAmt);
}
break;
@ -1410,7 +1402,7 @@ SDOperand SelectionDAG::getNode(unsigned Opcode, MVT::ValueType VT,
// exists.
if (ConstantSDNode *AC = dyn_cast<ConstantSDNode>(N2))
if ((AC->getValue() & (AC->getValue()-1)) == 0) {
unsigned ShCtV = ExactLog2(AC->getValue());
unsigned ShCtV = Log2_64(AC->getValue());
ShCtV = MVT::getSizeInBits(XType)-ShCtV-1;
SDOperand ShCt = getConstant(ShCtV, TLI.getShiftAmountTy());
SDOperand Shift = getNode(ISD::SRL, XType,

View File

@ -801,20 +801,6 @@ SDOperand AlphaISel::BuildUDIVSequence(SDOperand N) {
return Q;
}
//From PPC32
/// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
/// returns zero when the input is not exactly a power of two.
static unsigned ExactLog2(uint64_t Val) {
if (Val == 0 || (Val & (Val-1))) return 0;
unsigned Count = 0;
while (Val != 1) {
Val >>= 1;
++Count;
}
return Count;
}
//These describe LDAx
static const int IMM_LOW = -32768;
static const int IMM_HIGH = 32767;

View File

@ -847,29 +847,6 @@ SDOperand ISel::BuildConstmulSequence(SDOperand N) {
return finalresult;
}
/// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
/// returns zero when the input is not exactly a power of two.
static unsigned ExactLog2(uint64_t Val) {
if (Val == 0 || (Val & (Val-1))) return 0;
unsigned Count = 0;
while (Val != 1) {
Val >>= 1;
++Count;
}
return Count;
}
/// ExactLog2sub1 - This function solves for (Val == (1 << (N-1))-1)
/// and returns N. It returns 666 if Val is not 2^n -1 for some n.
static unsigned ExactLog2sub1(uint64_t Val) {
unsigned int n;
for(n=0; n<64; n++) {
if(Val==(uint64_t)((1LL<<n)-1))
return n;
}
return 666;
}
/// ponderIntegerDivisionBy - When handling integer divides, if the divide
/// is by a constant such that we can efficiently codegen it, this
/// function says what to do. Currently, it returns 0 if the division must
@ -882,7 +859,8 @@ static unsigned ponderIntegerDivisionBy(SDOperand N, bool isSigned,
int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
if ((Imm = ExactLog2(v))) { // if a division by a power of two, say so
if (isPowerOf2_64(v)) { // if a division by a power of two, say so
Imm = Log2_64(v);
return 1;
}
@ -895,7 +873,8 @@ static unsigned ponderIntegerAndWith(SDOperand N, unsigned& Imm) {
int64_t v = (int64_t)cast<ConstantSDNode>(N)->getSignExtended();
if ((Imm = ExactLog2sub1(v))!=666) { // if ANDing with ((2^n)-1) for some n
if (isMask_64(v)) { // if ANDing with ((2^n)-1) for some n
Imm = Log2_64(v);
return 1; // say so
}

View File

@ -26,11 +26,35 @@
#include "llvm/Target/TargetMachine.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/InstVisitor.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/Debug.h"
#include "llvm/ADT/Statistic.h"
#include <vector>
using namespace llvm;
// IsRunOfOnes - returns true if Val consists of one contiguous run of 1's with
// any number of 0's on either side. the 1's are allowed to wrap from LSB to
// MSB. so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
// not, since all 1's are not contiguous.
static bool IsRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
if (isShiftedMask_32(Val)) {
// look for the first non-zero bit
MB = CountLeadingZeros_32(Val);
// look for the first zero bit after the run of ones
ME = CountLeadingZeros_32((Val - 1) ^ Val);
return true;
} else if (isShiftedMask_32(Val = ~Val)) { // invert mask
// effectively look for the first zero bit
ME = CountLeadingZeros_32(Val) - 1;
// effectively look for the first one bit after the run of zeros
MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
return true;
}
// no run present
return false;
}
namespace {
/// TypeClass - Used by the PowerPC backend to group LLVM types by their basic
/// PPC Representation.
@ -2085,73 +2109,6 @@ void PPC32ISel::emitBinaryFPOperation(MachineBasicBlock *BB,
BuildMI(*BB, IP, Opcode, 2, DestReg).addReg(Op0r).addReg(Op1r);
}
// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
// returns zero when the input is not exactly a power of two.
static unsigned ExactLog2(unsigned Val) {
if (Val == 0 || (Val & (Val-1))) return 0;
unsigned Count = 0;
while (Val != 1) {
Val >>= 1;
++Count;
}
return Count;
}
// isRunOfOnes - returns true if Val consists of one contiguous run of 1's with
// any number of 0's on either side. the 1's are allowed to wrap from LSB to
// MSB. so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
// not, since all 1's are not contiguous.
static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
bool isRun = true;
MB = 0;
ME = 0;
// look for first set bit
int i = 0;
for (; i < 32; i++) {
if ((Val & (1 << (31 - i))) != 0) {
MB = i;
ME = i;
break;
}
}
// look for last set bit
for (; i < 32; i++) {
if ((Val & (1 << (31 - i))) == 0)
break;
ME = i;
}
// look for next set bit
for (; i < 32; i++) {
if ((Val & (1 << (31 - i))) != 0)
break;
}
// if we exhausted all the bits, we found a match at this point for 0*1*0*
if (i == 32)
return true;
// since we just encountered more 1's, if it doesn't wrap around to the
// most significant bit of the word, then we did not find a match to 1*0*1* so
// exit.
if (MB != 0)
return false;
// look for last set bit
for (MB = i; i < 32; i++) {
if ((Val & (1 << (31 - i))) == 0)
break;
}
// if we exhausted all the bits, then we found a match for 1*0*1*, otherwise,
// the value is not a run of ones.
if (i == 32)
return true;
return false;
}
/// isInsertAndHalf - Helper function for emitBitfieldInsert. Returns true if
/// OpUser has one use, is used by an or instruction, and is itself an and whose
/// second operand is a constant int. Optionally, set OrI to the Or instruction
@ -2281,7 +2238,7 @@ bool PPC32ISel::emitBitfieldInsert(User *OpUser, unsigned DestReg) {
// succeeded in matching one of the cases for generating rlwimi. Update the
// skip lists and users of the Instruction::Or.
unsigned MB, ME;
if (((TgtMask ^ InsMask) == 0xFFFFFFFF) && isRunOfOnes(InsMask, MB, ME)) {
if (((TgtMask ^ InsMask) == 0xFFFFFFFF) && IsRunOfOnes(InsMask, MB, ME)) {
SkipList.push_back(Op0User);
SkipList.push_back(Op1User);
SkipList.push_back(OptAndI);
@ -2320,7 +2277,7 @@ bool PPC32ISel::emitBitfieldExtract(MachineBasicBlock *MBB,
if (matched == false)
return false;
if (isRunOfOnes(Imm, MB, ME)) {
if (IsRunOfOnes(Imm, MB, ME)) {
unsigned SrcReg = getReg(Op, MBB, IP);
BuildMI(*MBB, IP, PPC::RLWINM, 4, DestReg).addReg(SrcReg).addImm(Rotate)
.addImm(MB).addImm(ME);
@ -2361,7 +2318,7 @@ void PPC32ISel::emitBinaryConstOperation(MachineBasicBlock *MBB,
if (Opcode == 2 && !Op1->isNullValue()) {
unsigned MB, ME, mask = Op1->getRawValue();
if (isRunOfOnes(mask, MB, ME)) {
if (IsRunOfOnes(mask, MB, ME)) {
BuildMI(*MBB, IP, PPC::RLWINM, 4, DestReg).addReg(Op0Reg).addImm(0)
.addImm(MB).addImm(ME);
return;
@ -2582,7 +2539,9 @@ void PPC32ISel::doMultiplyConst(MachineBasicBlock *MBB,
}
// If the element size is exactly a power of 2, use a shift to get it.
if (unsigned Shift = ExactLog2(CI->getRawValue())) {
uint64_t C = CI->getRawValue();
if (isPowerOf2_64(C)) {
unsigned Shift = Log2_64(C);
ConstantUInt *ShiftCI = ConstantUInt::get(Type::UByteTy, Shift);
emitShiftOperation(MBB, IP, Op0, ShiftCI, true, Op0->getType(), 0, DestReg);
return;
@ -2729,8 +2688,8 @@ void PPC32ISel::emitDivRemOperation(MachineBasicBlock *MBB,
return;
}
unsigned log2V = ExactLog2(V);
if (log2V != 0 && Ty->isSigned()) {
if (isPowerOf2_32(V) && Ty->isSigned()) {
unsigned log2V = Log2_32(V);
unsigned Op0Reg = getReg(Op0, MBB, IP);
unsigned TmpReg = makeAnotherReg(Op0->getType());

View File

@ -34,7 +34,7 @@
using namespace llvm;
//===----------------------------------------------------------------------===//
// PPC32TargetLowering - PPC32 Implementation of the TargetLowering interface
// PPC64TargetLowering - PPC64 Implementation of the TargetLowering interface
namespace {
class PPC64TargetLowering : public TargetLowering {
int VarArgsFrameIndex; // FrameIndex for start of varargs area.
@ -258,6 +258,7 @@ PPC64TargetLowering::LowerCallTo(SDOperand Chain,
// Just to be safe, we'll always reserve the full 48 bytes of linkage area
// plus 64 bytes of argument space in case any called code gets funky on us.
// (Required by ABI to support var arg)
if (NumBytes < 112) NumBytes = 112;
// Adjust the stack pointer for the new arguments...
@ -397,7 +398,7 @@ namespace {
Statistic<>NotLogic("ppc-codegen", "Number of inverted logical ops");
Statistic<>FusedFP("ppc-codegen", "Number of fused fp operations");
//===--------------------------------------------------------------------===//
/// ISel - PPC32 specific code to select PPC32 machine instructions for
/// ISel - PPC64 specific code to select PPC64 machine instructions for
/// SelectionDAG operations.
//===--------------------------------------------------------------------===//
class ISel : public SelectionDAGISel {
@ -447,18 +448,6 @@ public:
void SelectBranchCC(SDOperand N);
};
/// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
/// returns zero when the input is not exactly a power of two.
static unsigned ExactLog2(unsigned Val) {
if (Val == 0 || (Val & (Val-1))) return 0;
unsigned Count = 0;
while (Val != 1) {
Val >>= 1;
++Count;
}
return Count;
}
/// getImmediateForOpcode - This method returns a value indicating whether
/// the ConstantSDNode N can be used as an immediate to Opcode. The return
/// values are either 0, 1 or 2. 0 indicates that either N is not a
@ -477,25 +466,25 @@ static unsigned getImmediateForOpcode(SDOperand N, unsigned Opcode,
switch(Opcode) {
default: return 0;
case ISD::ADD:
if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
if (isInt16(v)) { Imm = v & 0xFFFF; return 1; }
if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
break;
case ISD::AND:
case ISD::XOR:
case ISD::OR:
if (v >= 0 && v <= 65535) { Imm = v & 0xFFFF; return 1; }
if (isUInt16(v)) { Imm = v & 0xFFFF; return 1; }
if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
break;
case ISD::MUL:
case ISD::SUB:
if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
if (isInt16(v)) { Imm = v & 0xFFFF; return 1; }
break;
case ISD::SETCC:
if (U && (v >= 0 && v <= 65535)) { Imm = v & 0xFFFF; return 1; }
if (!U && (v <= 32767 && v >= -32768)) { Imm = v & 0xFFFF; return 1; }
if (U && isUInt16(v)) { Imm = v & 0xFFFF; return 1; }
if (!U && isInt16(v)) { Imm = v & 0xFFFF; return 1; }
break;
case ISD::SDIV:
if ((Imm = ExactLog2(v))) { return 3; }
if (isPowerOf2_32(v)) { Imm = Log2_32(v); return 3; }
break;
}
return 0;
@ -1636,7 +1625,7 @@ void ISel::Select(SDOperand N) {
}
/// createPPC32PatternInstructionSelector - This pass converts an LLVM function
/// createPPC64PatternInstructionSelector - This pass converts an LLVM function
/// into a machine code representation using pattern matching and a machine
/// description file.
///

View File

@ -698,7 +698,7 @@ bool AIXAsmPrinter::doFinalization(Module &M) {
O << "\t.lcomm " << Name << ",16,_global.bss_c";
} else {
O << "\t.comm " << Name << "," << TD.getTypeSize(I->getType())
<< "," << log2((unsigned)TD.getTypeAlignment(I->getType()));
<< "," << Log2_32((unsigned)TD.getTypeAlignment(I->getType()));
}
O << "\t\t# ";
WriteAsOperand(O, I, true, true, &M);

View File

@ -35,6 +35,29 @@
#include <algorithm>
using namespace llvm;
// IsRunOfOnes - returns true if Val consists of one contiguous run of 1's with
// any number of 0's on either side. the 1's are allowed to wrap from LSB to
// MSB. so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
// not, since all 1's are not contiguous.
static bool IsRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
if (isShiftedMask_32(Val)) {
// look for the first non-zero bit
MB = CountLeadingZeros_32(Val);
// look for the first zero bit after the run of ones
ME = CountLeadingZeros_32((Val - 1) ^ Val);
return true;
} else if (isShiftedMask_32(Val = ~Val)) { // invert mask
// effectively look for the first zero bit
ME = CountLeadingZeros_32(Val) - 1;
// effectively look for the first one bit after the run of zeros
MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
return true;
}
// no run present
return false;
}
//===----------------------------------------------------------------------===//
// PPC32TargetLowering - PPC32 Implementation of the TargetLowering interface
namespace {
@ -321,6 +344,7 @@ PPC32TargetLowering::LowerCallTo(SDOperand Chain,
// Just to be safe, we'll always reserve the full 24 bytes of linkage area
// plus 32 bytes of argument space in case any called code gets funky on us.
// (Required by ABI to support var arg)
if (NumBytes < 56) NumBytes = 56;
// Adjust the stack pointer for the new arguments...
@ -664,36 +688,36 @@ static unsigned getImmediateForOpcode(SDOperand N, unsigned Opcode,
switch(Opcode) {
default: return 0;
case ISD::ADD:
if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
if (isInt16(v)) { Imm = v & 0xFFFF; return 1; }
if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
break;
case ISD::AND: {
unsigned MB, ME;
if (IsRunOfOnes(v, MB, ME)) { Imm = MB << 16 | ME & 0xFFFF; return 5; }
if (v >= 0 && v <= 65535) { Imm = v & 0xFFFF; return 1; }
if (isUInt16(v)) { Imm = v & 0xFFFF; return 1; }
if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
break;
}
case ISD::XOR:
case ISD::OR:
if (v >= 0 && v <= 65535) { Imm = v & 0xFFFF; return 1; }
if (isUInt16(v)) { Imm = v & 0xFFFF; return 1; }
if ((v & 0x0000FFFF) == 0) { Imm = v >> 16; return 2; }
break;
case ISD::MUL:
if (v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
if (isInt16(v)) { Imm = v & 0xFFFF; return 1; }
break;
case ISD::SUB:
// handle subtract-from separately from subtract, since subi is really addi
if (U && v <= 32767 && v >= -32768) { Imm = v & 0xFFFF; return 1; }
if (!U && v <= 32768 && v >= -32767) { Imm = (-v) & 0xFFFF; return 1; }
if (U && isInt16(v)) { Imm = v & 0xFFFF; return 1; }
if (!U && isInt16(-v)) { Imm = (-v) & 0xFFFF; return 1; }
break;
case ISD::SETCC:
if (U && (v >= 0 && v <= 65535)) { Imm = v & 0xFFFF; return 1; }
if (!U && (v <= 32767 && v >= -32768)) { Imm = v & 0xFFFF; return 1; }
if (U && isUInt16(v)) { Imm = v & 0xFFFF; return 1; }
if (!U && isInt16(v)) { Imm = v & 0xFFFF; return 1; }
break;
case ISD::SDIV:
if ((Imm = ExactLog2(v))) { return 3; }
if ((Imm = ExactLog2(-v))) { Imm = -Imm; return 3; }
if (isPowerOf2_32(v)) { Imm = Log2_32(v); return 3; }
if (isPowerOf2_32(-v)) { Imm = Log2_32(-v); return 3; }
if (v <= -2 || v >= 2) { return 4; }
break;
case ISD::UDIV:
@ -807,7 +831,7 @@ struct mu {
static struct ms magic(int d) {
int p;
unsigned int ad, anc, delta, q1, r1, q2, r2, t;
const unsigned int two31 = 2147483648U; // 2^31
const unsigned int two31 = 0x80000000U;
struct ms mag;
ad = abs(d);

View File

@ -2306,7 +2306,6 @@ CreateMulConstInstruction(const TargetMachine &target, Function* F,
constOp->getType(),
isValidConst);
if (isValidConst) {
unsigned pow;
bool needNeg = false;
if (C < 0) {
needNeg = true;
@ -2323,7 +2322,8 @@ CreateMulConstInstruction(const TargetMachine &target, Function* F,
else
M = BuildMI(V9::ADDr,3).addReg(lval).addMReg(Zero).addRegDef(destVal);
mvec.push_back(M);
} else if (isPowerOf2(C, pow)) {
} else if (isPowerOf2_64(C)) {
unsigned pow = Log2_64(C);
if(!needNeg) {
unsigned opSize = target.getTargetData().getTypeSize(resultType);
MachineOpCode opCode = (opSize <= 32)? V9::SLLr5 : V9::SLLXr6;
@ -2464,7 +2464,6 @@ static void CreateDivConstInstruction(TargetMachine &target,
const Type* resultType = instrNode->getInstruction()->getType();
if (resultType->isInteger()) {
unsigned pow;
bool isValidConst;
int64_t C = (int64_t) ConvertConstantToIntType(target, constOp,
constOp->getType(),
@ -2479,7 +2478,8 @@ static void CreateDivConstInstruction(TargetMachine &target,
if (C == 1) {
mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addMReg(ZeroReg)
.addRegDef(destVal));
} else if (isPowerOf2(C, pow)) {
} else if (isPowerOf2_64(C)) {
unsigned pow = Log2_64(C);
unsigned opCode;
Value* shiftOperand;
unsigned opSize = target.getTargetData().getTypeSize(resultType);
@ -2539,7 +2539,7 @@ static void CreateDivConstInstruction(TargetMachine &target,
.addRegDef(destVal));
}
if (needNeg && (C == 1 || isPowerOf2(C, pow))) {
if (needNeg && (C == 1 || isPowerOf2_64(C))) {
// insert <reg = SUB 0, reg> after the instr to flip the sign
mvec.push_back(CreateIntNegInstruction(target, destVal));
}

View File

@ -229,7 +229,7 @@ unsigned char TargetData::getTypeAlignment(const Type *Ty) const {
unsigned char TargetData::getTypeAlignmentShift(const Type *Ty) const {
unsigned Align = getTypeAlignment(Ty);
assert(!(Align & (Align-1)) && "Alignment is not a power of two!");
return log2(Align);
return Log2_32(Align);
}
/// getIntPtrType - Return an unsigned integer type that is the same size or

View File

@ -3023,7 +3023,7 @@ unsigned ISel::SelectExpr(SDOperand N) {
RHS = -RHS;
}
if (RHS && (RHS & (RHS-1)) == 0) { // Signed division by power of 2?
unsigned Log = log2(RHS);
unsigned Log = Log2_32(RHS);
unsigned SAROpc, SHROpc, ADDOpc, NEGOpc;
switch (N.getValueType()) {
default: assert("Unknown type to signed divide!");

View File

@ -29,6 +29,7 @@
#include "llvm/Target/TargetOptions.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/InstVisitor.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/ADT/Statistic.h"
using namespace llvm;
@ -2496,19 +2497,6 @@ void X86ISel::doMultiply(MachineBasicBlock *MBB,
}
}
// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
// returns zero when the input is not exactly a power of two.
static unsigned ExactLog2(unsigned Val) {
if (Val == 0 || (Val & (Val-1))) return 0;
unsigned Count = 0;
while (Val != 1) {
Val >>= 1;
++Count;
}
return Count+1;
}
/// doMultiplyConst - This function is specialized to efficiently codegen an 8,
/// 16, or 32-bit integer multiply by a constant.
void X86ISel::doMultiplyConst(MachineBasicBlock *MBB,
@ -2573,35 +2561,37 @@ void X86ISel::doMultiplyConst(MachineBasicBlock *MBB,
}
// If the element size is exactly a power of 2, use a shift to get it.
if (unsigned Shift = ExactLog2(ConstRHS)) {
if (isPowerOf2_32(ConstRHS)) {
unsigned Shift = Log2_32(ConstRHS);
switch (Class) {
default: assert(0 && "Unknown class for this function!");
case cByte:
BuildMI(*MBB, IP, X86::SHL8ri,2, DestReg).addReg(op0Reg).addImm(Shift-1);
BuildMI(*MBB, IP, X86::SHL8ri,2, DestReg).addReg(op0Reg).addImm(Shift);
return;
case cShort:
BuildMI(*MBB, IP, X86::SHL16ri,2, DestReg).addReg(op0Reg).addImm(Shift-1);
BuildMI(*MBB, IP, X86::SHL16ri,2, DestReg).addReg(op0Reg).addImm(Shift);
return;
case cInt:
BuildMI(*MBB, IP, X86::SHL32ri,2, DestReg).addReg(op0Reg).addImm(Shift-1);
BuildMI(*MBB, IP, X86::SHL32ri,2, DestReg).addReg(op0Reg).addImm(Shift);
return;
}
}
// If the element size is a negative power of 2, use a shift/neg to get it.
if (unsigned Shift = ExactLog2(-ConstRHS)) {
if (isPowerOf2_32(-ConstRHS)) {
unsigned Shift = Log2_32(-ConstRHS);
TmpReg = makeAnotherReg(DestTy);
BuildMI(*MBB, IP, NEGrTab[Class], 1, TmpReg).addReg(op0Reg);
switch (Class) {
default: assert(0 && "Unknown class for this function!");
case cByte:
BuildMI(*MBB, IP, X86::SHL8ri,2, DestReg).addReg(TmpReg).addImm(Shift-1);
BuildMI(*MBB, IP, X86::SHL8ri,2, DestReg).addReg(TmpReg).addImm(Shift);
return;
case cShort:
BuildMI(*MBB, IP, X86::SHL16ri,2, DestReg).addReg(TmpReg).addImm(Shift-1);
BuildMI(*MBB, IP, X86::SHL16ri,2, DestReg).addReg(TmpReg).addImm(Shift);
return;
case cInt:
BuildMI(*MBB, IP, X86::SHL32ri,2, DestReg).addReg(TmpReg).addImm(Shift-1);
BuildMI(*MBB, IP, X86::SHL32ri,2, DestReg).addReg(TmpReg).addImm(Shift);
return;
}
}
@ -2917,7 +2907,8 @@ void X86ISel::emitDivRemOperation(MachineBasicBlock *BB,
V = -V;
isNeg = true; // Maybe it's a negative power of 2.
}
if (unsigned Log = ExactLog2(V)) {
if (isPowerOf2_32(V)) {
unsigned Log = Log2_32(V);
--Log;
unsigned Op0Reg = getReg(Op0, BB, IP);
unsigned TmpReg = makeAnotherReg(Op0->getType());