regop.isa revision 5116:91881e9404de
1// Copyright (c) 2007 The Hewlett-Packard Development Company
2// All rights reserved.
3//
4// Redistribution and use of this software in source and binary forms,
5// with or without modification, are permitted provided that the
6// following conditions are met:
7//
8// The software must be used only for Non-Commercial Use which means any
9// use which is NOT directed to receiving any direct monetary
10// compensation for, or commercial advantage from such use.  Illustrative
11// examples of non-commercial use are academic research, personal study,
12// teaching, education and corporate research & development.
13// Illustrative examples of commercial use are distributing products for
14// commercial advantage and providing services using the software for
15// commercial advantage.
16//
17// If you wish to use this software or functionality therein that may be
18// covered by patents for commercial use, please contact:
19//     Director of Intellectual Property Licensing
20//     Office of Strategy and Technology
21//     Hewlett-Packard Company
22//     1501 Page Mill Road
23//     Palo Alto, California  94304
24//
25// Redistributions of source code must retain the above copyright notice,
26// this list of conditions and the following disclaimer.  Redistributions
27// in binary form must reproduce the above copyright notice, this list of
28// conditions and the following disclaimer in the documentation and/or
29// other materials provided with the distribution.  Neither the name of
30// the COPYRIGHT HOLDER(s), HEWLETT-PACKARD COMPANY, nor the names of its
31// contributors may be used to endorse or promote products derived from
32// this software without specific prior written permission.  No right of
33// sublicense is granted herewith.  Derivatives of the software and
34// output created using the software may be prepared, but only for
35// Non-Commercial Uses.  Derivatives of the software may be shared with
36// others provided: (i) the others agree to abide by the list of
37// conditions herein which includes the Non-Commercial Use restrictions;
38// and (ii) such Derivatives of the software include the above copyright
39// notice to acknowledge the contribution from this software where
40// applicable, this list of conditions and the disclaimer below.
41//
42// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
43// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
44// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
45// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
46// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
47// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
48// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
49// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
50// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
51// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
52// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
53//
54// Authors: Gabe Black
55
56//////////////////////////////////////////////////////////////////////////
57//
58// RegOp Microop templates
59//
60//////////////////////////////////////////////////////////////////////////
61
62def template MicroRegOpExecute {{
63        Fault %(class_name)s::execute(%(CPU_exec_context)s *xc,
64                Trace::InstRecord *traceData) const
65        {
66            Fault fault = NoFault;
67
68            DPRINTF(X86, "The data size is %d\n", dataSize);
69            %(op_decl)s;
70            %(op_rd)s;
71
72            if(%(cond_check)s)
73            {
74                %(code)s;
75                %(flag_code)s;
76            }
77            else
78            {
79                %(else_code)s;
80            }
81
82            //Write the resulting state to the execution context
83            if(fault == NoFault)
84            {
85                %(op_wb)s;
86            }
87            return fault;
88        }
89}};
90
91def template MicroRegOpImmExecute {{
92        Fault %(class_name)s::execute(%(CPU_exec_context)s *xc,
93                Trace::InstRecord *traceData) const
94        {
95            Fault fault = NoFault;
96
97            %(op_decl)s;
98            %(op_rd)s;
99
100            if(%(cond_check)s)
101            {
102                %(code)s;
103                %(flag_code)s;
104            }
105            else
106            {
107                %(else_code)s;
108            }
109
110            //Write the resulting state to the execution context
111            if(fault == NoFault)
112            {
113                %(op_wb)s;
114            }
115            return fault;
116        }
117}};
118
119def template MicroRegOpDeclare {{
120    class %(class_name)s : public %(base_class)s
121    {
122      protected:
123        void buildMe();
124
125      public:
126        %(class_name)s(ExtMachInst _machInst,
127                const char * instMnem,
128                bool isMicro, bool isDelayed, bool isFirst, bool isLast,
129                RegIndex _src1, RegIndex _src2, RegIndex _dest,
130                uint8_t _dataSize, uint16_t _ext);
131
132        %(class_name)s(ExtMachInst _machInst,
133                const char * instMnem,
134                RegIndex _src1, RegIndex _src2, RegIndex _dest,
135                uint8_t _dataSize, uint16_t _ext);
136
137        %(BasicExecDeclare)s
138    };
139}};
140
141def template MicroRegOpImmDeclare {{
142
143    class %(class_name)s : public %(base_class)s
144    {
145      protected:
146        void buildMe();
147
148      public:
149        %(class_name)s(ExtMachInst _machInst,
150                const char * instMnem,
151                bool isMicro, bool isDelayed, bool isFirst, bool isLast,
152                RegIndex _src1, uint16_t _imm8, RegIndex _dest,
153                uint8_t _dataSize, uint16_t _ext);
154
155        %(class_name)s(ExtMachInst _machInst,
156                const char * instMnem,
157                RegIndex _src1, uint16_t _imm8, RegIndex _dest,
158                uint8_t _dataSize, uint16_t _ext);
159
160        %(BasicExecDeclare)s
161    };
162}};
163
164def template MicroRegOpConstructor {{
165
166    inline void %(class_name)s::buildMe()
167    {
168        %(constructor)s;
169    }
170
171    inline %(class_name)s::%(class_name)s(
172            ExtMachInst machInst, const char * instMnem,
173            RegIndex _src1, RegIndex _src2, RegIndex _dest,
174            uint8_t _dataSize, uint16_t _ext) :
175        %(base_class)s(machInst, "%(mnemonic)s", instMnem,
176                false, false, false, false,
177                _src1, _src2, _dest, _dataSize, _ext,
178                %(op_class)s)
179    {
180        buildMe();
181    }
182
183    inline %(class_name)s::%(class_name)s(
184            ExtMachInst machInst, const char * instMnem,
185            bool isMicro, bool isDelayed, bool isFirst, bool isLast,
186            RegIndex _src1, RegIndex _src2, RegIndex _dest,
187            uint8_t _dataSize, uint16_t _ext) :
188        %(base_class)s(machInst, "%(mnemonic)s", instMnem,
189                isMicro, isDelayed, isFirst, isLast,
190                _src1, _src2, _dest, _dataSize, _ext,
191                %(op_class)s)
192    {
193        buildMe();
194    }
195}};
196
197def template MicroRegOpImmConstructor {{
198
199    inline void %(class_name)s::buildMe()
200    {
201        %(constructor)s;
202    }
203
204    inline %(class_name)s::%(class_name)s(
205            ExtMachInst machInst, const char * instMnem,
206            RegIndex _src1, uint16_t _imm8, RegIndex _dest,
207            uint8_t _dataSize, uint16_t _ext) :
208        %(base_class)s(machInst, "%(mnemonic)s", instMnem,
209                false, false, false, false,
210                _src1, _imm8, _dest, _dataSize, _ext,
211                %(op_class)s)
212    {
213        buildMe();
214    }
215
216    inline %(class_name)s::%(class_name)s(
217            ExtMachInst machInst, const char * instMnem,
218            bool isMicro, bool isDelayed, bool isFirst, bool isLast,
219            RegIndex _src1, uint16_t _imm8, RegIndex _dest,
220            uint8_t _dataSize, uint16_t _ext) :
221        %(base_class)s(machInst, "%(mnemonic)s", instMnem,
222                isMicro, isDelayed, isFirst, isLast,
223                _src1, _imm8, _dest, _dataSize, _ext,
224                %(op_class)s)
225    {
226        buildMe();
227    }
228}};
229
230output header {{
231    void
232    divide(uint64_t dividend, uint64_t divisor,
233            uint64_t &quotient, uint64_t &remainder);
234}};
235
236output decoder {{
237    void
238    divide(uint64_t dividend, uint64_t divisor,
239            uint64_t &quotient, uint64_t &remainder)
240    {
241        //Check for divide by zero.
242        if (divisor == 0)
243            panic("Divide by zero!\\n");
244        //If the divisor is bigger than the dividend, don't do anything.
245        if (divisor <= dividend) {
246            //Shift the divisor so it's msb lines up with the dividend.
247            int dividendMsb = findMsbSet(dividend);
248            int divisorMsb = findMsbSet(divisor);
249            int shift = dividendMsb - divisorMsb;
250            divisor <<= shift;
251            //Compute what we'll add to the quotient if the divisor isn't
252            //now larger than the dividend.
253            uint64_t quotientBit = 1;
254            quotientBit <<= shift;
255            //If we need to step back a bit (no pun intended) because the
256            //divisor got too to large, do that here. This is the "or two"
257            //part of one or two bit division.
258            if (divisor > dividend) {
259                quotientBit >>= 1;
260                divisor >>= 1;
261            }
262            //Decrement the remainder and increment the quotient.
263            quotient += quotientBit;
264            remainder -= divisor;
265        }
266    }
267}};
268
269let {{
270    # Make these empty strings so that concatenating onto
271    # them will always work.
272    header_output = ""
273    decoder_output = ""
274    exec_output = ""
275
276    immTemplates = (
277            MicroRegOpImmDeclare,
278            MicroRegOpImmConstructor,
279            MicroRegOpImmExecute)
280
281    regTemplates = (
282            MicroRegOpDeclare,
283            MicroRegOpConstructor,
284            MicroRegOpExecute)
285
286    class RegOpMeta(type):
287        def buildCppClasses(self, name, Name, suffix, \
288                code, flag_code, cond_check, else_code):
289
290            # Globals to stick the output in
291            global header_output
292            global decoder_output
293            global exec_output
294
295            # Stick all the code together so it can be searched at once
296            allCode = "|".join((code, flag_code, cond_check, else_code))
297
298            # If op2 is used anywhere, make register and immediate versions
299            # of this code.
300            matcher = re.compile("(?<!\\w)(?P<prefix>s?)op2(?P<typeQual>\\.\\w+)?")
301            match = matcher.search(allCode)
302            if match:
303                typeQual = ""
304                if match.group("typeQual"):
305                    typeQual = match.group("typeQual")
306                src2_name = "%spsrc2%s" % (match.group("prefix"), typeQual)
307                self.buildCppClasses(name, Name, suffix,
308                        matcher.sub(src2_name, code),
309                        matcher.sub(src2_name, flag_code),
310                        matcher.sub(src2_name, cond_check),
311                        matcher.sub(src2_name, else_code))
312                self.buildCppClasses(name + "i", Name, suffix + "Imm",
313                        matcher.sub("imm8", code),
314                        matcher.sub("imm8", flag_code),
315                        matcher.sub("imm8", cond_check),
316                        matcher.sub("imm8", else_code))
317                return
318
319            # If there's something optional to do with flags, generate
320            # a version without it and fix up this version to use it.
321            if flag_code is not "" or cond_check is not "true":
322                self.buildCppClasses(name, Name, suffix,
323                        code, "", "true", else_code)
324                suffix = "Flags" + suffix
325
326            # If psrc1 or psrc2 is used, we need to actually insert code to
327            # compute it.
328            matcher = re.compile("(?<!\w)psrc1(?!\w)")
329            if matcher.search(allCode):
330                code = "uint64_t psrc1 = pick(SrcReg1, 0, dataSize);" + code
331            matcher = re.compile("(?<!\w)psrc2(?!\w)")
332            if matcher.search(allCode):
333                code = "uint64_t psrc2 = pick(SrcReg2, 1, dataSize);" + code
334            # Also make available versions which do sign extension
335            matcher = re.compile("(?<!\w)spsrc1(?!\w)")
336            if matcher.search(allCode):
337                code = "int64_t spsrc1 = signedPick(SrcReg1, 0, dataSize);" + code
338            matcher = re.compile("(?<!\w)spsrc2(?!\w)")
339            if matcher.search(allCode):
340                code = "int64_t spsrc2 = signedPick(SrcReg2, 1, dataSize);" + code
341
342            base = "X86ISA::RegOp"
343
344            # If imm8 shows up in the code, use the immediate templates, if
345            # not, hopefully the register ones will be correct.
346            templates = regTemplates
347            matcher = re.compile("(?<!\w)imm8(?!\w)")
348            if matcher.search(allCode):
349                base += "Imm"
350                templates = immTemplates
351
352            # Get everything ready for the substitution
353            iop = InstObjParams(name, Name + suffix, base,
354                    {"code" : code,
355                     "flag_code" : flag_code,
356                     "cond_check" : cond_check,
357                     "else_code" : else_code})
358
359            # Generate the actual code (finally!)
360            header_output += templates[0].subst(iop)
361            decoder_output += templates[1].subst(iop)
362            exec_output += templates[2].subst(iop)
363
364
365        def __new__(mcls, Name, bases, dict):
366            abstract = False
367            name = Name.lower()
368            if "abstract" in dict:
369                abstract = dict['abstract']
370                del dict['abstract']
371
372            cls = super(RegOpMeta, mcls).__new__(mcls, Name, bases, dict)
373            if not abstract:
374                cls.className = Name
375                cls.base_mnemonic = name
376                code = cls.code
377                flag_code = cls.flag_code
378                cond_check = cls.cond_check
379                else_code = cls.else_code
380
381                # Set up the C++ classes
382                mcls.buildCppClasses(cls, name, Name, "",
383                        code, flag_code, cond_check, else_code)
384
385                # Hook into the microassembler dict
386                global microopClasses
387                microopClasses[name] = cls
388
389                allCode = "|".join((code, flag_code, cond_check, else_code))
390
391                # If op2 is used anywhere, make register and immediate versions
392                # of this code.
393                matcher = re.compile("op2(?P<typeQual>\\.\\w+)?")
394                if matcher.search(allCode):
395                    microopClasses[name + 'i'] = cls
396            return cls
397
398
399    class RegOp(X86Microop):
400        __metaclass__ = RegOpMeta
401        # This class itself doesn't act as a microop
402        abstract = True
403
404        # Default template parameter values
405        flag_code = ""
406        cond_check = "true"
407        else_code = ";"
408
409        def __init__(self, dest, src1, op2, flags = None, dataSize = "env.dataSize"):
410            self.dest = dest
411            self.src1 = src1
412            self.op2 = op2
413            self.flags = flags
414            self.dataSize = dataSize
415            if flags is None:
416                self.ext = 0
417            else:
418                if not isinstance(flags, (list, tuple)):
419                    raise Exception, "flags must be a list or tuple of flags"
420                self.ext = " | ".join(flags)
421                self.className += "Flags"
422
423        def getAllocator(self, *microFlags):
424            className = self.className
425            if self.mnemonic == self.base_mnemonic + 'i':
426                className += "Imm"
427            allocator = '''new %(class_name)s(machInst, mnemonic
428                    %(flags)s, %(src1)s, %(op2)s, %(dest)s,
429                    %(dataSize)s, %(ext)s)''' % {
430                "class_name" : className,
431                "flags" : self.microFlagsText(microFlags),
432                "src1" : self.src1, "op2" : self.op2,
433                "dest" : self.dest,
434                "dataSize" : self.dataSize,
435                "ext" : self.ext}
436            return allocator
437
438    class LogicRegOp(RegOp):
439        abstract = True
440        flag_code = '''
441            //Don't have genFlags handle the OF or CF bits
442            uint64_t mask = CFBit | ECFBit | OFBit;
443            ccFlagBits = genFlags(ccFlagBits, ext & ~mask, DestReg, psrc1, op2);
444            //If a logic microop wants to set these, it wants to set them to 0.
445            ccFlagBits &= ~(CFBit & ext);
446            ccFlagBits &= ~(ECFBit & ext);
447            ccFlagBits &= ~(OFBit & ext);
448        '''
449
450    class FlagRegOp(RegOp):
451        abstract = True
452        flag_code = \
453            "ccFlagBits = genFlags(ccFlagBits, ext, DestReg, psrc1, op2);"
454
455    class SubRegOp(RegOp):
456        abstract = True
457        flag_code = \
458            "ccFlagBits = genFlags(ccFlagBits, ext, DestReg, psrc1, ~op2, true);"
459
460    class CondRegOp(RegOp):
461        abstract = True
462        cond_check = "checkCondition(ccFlagBits, ext)"
463
464    class RdRegOp(RegOp):
465        abstract = True
466        def __init__(self, dest, src1=None, dataSize="env.dataSize"):
467            if not src1:
468                src1 = dest
469            super(RdRegOp, self).__init__(dest, src1, "NUM_INTREGS", None, dataSize)
470
471    class WrRegOp(RegOp):
472        abstract = True
473        def __init__(self, src1, src2, flags=None, dataSize="env.dataSize"):
474            super(WrRegOp, self).__init__("NUM_INTREGS", src1, src2, flags, dataSize)
475
476    class Add(FlagRegOp):
477        code = 'DestReg = merge(DestReg, psrc1 + op2, dataSize);'
478
479    class Or(LogicRegOp):
480        code = 'DestReg = merge(DestReg, psrc1 | op2, dataSize);'
481
482    class Adc(FlagRegOp):
483        code = '''
484            CCFlagBits flags = ccFlagBits;
485            DestReg = merge(DestReg, psrc1 + op2 + flags.CF, dataSize);
486            '''
487
488    class Sbb(SubRegOp):
489        code = '''
490            CCFlagBits flags = ccFlagBits;
491            DestReg = merge(DestReg, psrc1 - op2 - flags.CF, dataSize);
492            '''
493
494    class And(LogicRegOp):
495        code = 'DestReg = merge(DestReg, psrc1 & op2, dataSize)'
496
497    class Sub(SubRegOp):
498        code = 'DestReg = merge(DestReg, psrc1 - op2, dataSize)'
499
500    class Xor(LogicRegOp):
501        code = 'DestReg = merge(DestReg, psrc1 ^ op2, dataSize)'
502
503    # Neither of these is quite correct because it assumes that right shifting
504    # a signed or unsigned value does sign or zero extension respectively.
505    # The C standard says that what happens on a right shift with a 1 in the
506    # MSB position is undefined. On x86 and under likely most compilers the
507    # "right thing" happens, but this isn't a guarantee.
508    class Mul1s(WrRegOp):
509        code = '''
510            ProdLow = psrc1 * op2;
511            int halfSize = (dataSize * 8) / 2;
512            int64_t spsrc1_h = spsrc1 >> halfSize;
513            int64_t spsrc1_l = spsrc1 & mask(halfSize);
514            int64_t spsrc2_h = sop2 >> halfSize;
515            int64_t spsrc2_l = sop2 & mask(halfSize);
516            ProdHi = ((spsrc1_l * spsrc2_h + spsrc1_h * spsrc2_l +
517                      ((spsrc1_l * spsrc2_l) >> halfSize)) >> halfSize) +
518                     spsrc1_h * spsrc2_h;
519            '''
520
521    class Mul1u(WrRegOp):
522        code = '''
523            ProdLow = psrc1 * op2;
524            int halfSize = (dataSize * 8) / 2;
525            uint64_t psrc1_h = psrc1 >> halfSize;
526            uint64_t psrc1_l = psrc1 & mask(halfSize);
527            uint64_t psrc2_h = op2 >> halfSize;
528            uint64_t psrc2_l = op2 & mask(halfSize);
529            ProdHi = ((psrc1_l * psrc2_h + psrc1_h * psrc2_l +
530                      ((psrc1_l * psrc2_l) >> halfSize)) >> halfSize) +
531                     psrc1_h * psrc2_h;
532            '''
533
534    class Mulel(RdRegOp):
535        code = 'DestReg = merge(SrcReg1, ProdLow, dataSize);'
536
537    class Muleh(RdRegOp):
538        def __init__(self, dest, src1=None, flags=None, dataSize="env.dataSize"):
539            if not src1:
540                src1 = dest
541            super(RdRegOp, self).__init__(dest, src1, "NUM_INTREGS", flags, dataSize)
542        code = 'DestReg = merge(SrcReg1, ProdHi, dataSize);'
543        flag_code = '''
544            if (ProdHi)
545                ccFlagBits = ccFlagBits | (ext & (CFBit | OFBit | ECFBit));
546            else
547                ccFlagBits = ccFlagBits & ~(ext & (CFBit | OFBit | ECFBit));
548        '''
549
550    # One or two bit divide
551    class Div1(WrRegOp):
552        code = '''
553            //These are temporaries so that modifying them later won't make
554            //the ISA parser think they're also sources.
555            uint64_t quotient = 0;
556            uint64_t remainder = psrc1;
557            //Similarly, this is a temporary so changing it doesn't make it
558            //a source.
559            uint64_t divisor = op2;
560            //This is a temporary just for consistency and clarity.
561            uint64_t dividend = remainder;
562            //Do the division.
563            divide(dividend, divisor, quotient, remainder);
564            //Record the final results.
565            Remainder = remainder;
566            Quotient = quotient;
567            Divisor = divisor;
568            '''
569
570    # Step divide
571    class Div2(RegOp):
572        code = '''
573            uint64_t dividend = Remainder;
574            uint64_t divisor = Divisor;
575            uint64_t quotient = Quotient;
576            uint64_t remainder = dividend;
577            int remaining = op2;
578            //If we overshot, do nothing. This lets us unrool division loops a
579            //little.
580            if (remaining) {
581                //Shift in bits from the low order portion of the dividend
582                while(dividend < divisor && remaining) {
583                    dividend = (dividend << 1) | bits(SrcReg1, remaining - 1);
584                    quotient <<= 1;
585                    remaining--;
586                }
587                remainder = dividend;
588                //Do the division.
589                divide(dividend, divisor, quotient, remainder);
590            }
591            //Keep track of how many bits there are still to pull in.
592            DestReg = merge(DestReg, remaining, dataSize);
593            //Record the final results
594            Remainder = remainder;
595            Quotient = quotient;
596        '''
597        flag_code = '''
598            if (DestReg == 0)
599                ccFlagBits = ccFlagBits | (ext & EZFBit);
600            else
601                ccFlagBits = ccFlagBits & ~(ext & EZFBit);
602        '''
603
604    class Divq(RdRegOp):
605        code = 'DestReg = merge(SrcReg1, Quotient, dataSize);'
606
607    class Divr(RdRegOp):
608        code = 'DestReg = merge(SrcReg1, Remainder, dataSize);'
609
610    class Mov(CondRegOp):
611        code = 'DestReg = merge(SrcReg1, op2, dataSize)'
612        else_code = 'DestReg=DestReg;'
613
614    # Shift instructions
615
616    class Sll(RegOp):
617        code = '''
618            uint8_t shiftAmt = (op2 & ((dataSize == 8) ? mask(6) : mask(5)));
619            DestReg = merge(DestReg, psrc1 << shiftAmt, dataSize);
620            '''
621        flag_code = '''
622            // If the shift amount is zero, no flags should be modified.
623            if (shiftAmt) {
624                //Zero out any flags we might modify. This way we only have to
625                //worry about setting them.
626                ccFlagBits = ccFlagBits & ~(ext & (CFBit | ECFBit | OFBit));
627                int CFBits = 0;
628                //Figure out if we -would- set the CF bits if requested.
629                if (bits(SrcReg1, dataSize * 8 - shiftAmt))
630                    CFBits = 1;
631                //If some combination of the CF bits need to be set, set them.
632                if ((ext & (CFBit | ECFBit)) && CFBits)
633                    ccFlagBits = ccFlagBits | (ext & (CFBit | ECFBit));
634                //Figure out what the OF bit should be.
635                if ((ext & OFBit) && (CFBits ^ bits(DestReg, dataSize * 8 - 1)))
636                    ccFlagBits = ccFlagBits | OFBit;
637                //Use the regular mechanisms to calculate the other flags.
638                ccFlagBits = genFlags(ccFlagBits, ext & ~(CFBit | ECFBit | OFBit),
639                        DestReg, psrc1, op2);
640            }
641        '''
642
643    class Srl(RegOp):
644        code = '''
645            uint8_t shiftAmt = (op2 & ((dataSize == 8) ? mask(6) : mask(5)));
646            // Because what happens to the bits shift -in- on a right shift
647            // is not defined in the C/C++ standard, we have to mask them out
648            // to be sure they're zero.
649            uint64_t logicalMask = mask(dataSize * 8 - shiftAmt);
650            DestReg = merge(DestReg, (psrc1 >> shiftAmt) & logicalMask, dataSize);
651            '''
652        flag_code = '''
653            // If the shift amount is zero, no flags should be modified.
654            if (shiftAmt) {
655                //Zero out any flags we might modify. This way we only have to
656                //worry about setting them.
657                ccFlagBits = ccFlagBits & ~(ext & (CFBit | ECFBit | OFBit));
658                //If some combination of the CF bits need to be set, set them.
659                if ((ext & (CFBit | ECFBit)) && bits(SrcReg1, shiftAmt - 1))
660                    ccFlagBits = ccFlagBits | (ext & (CFBit | ECFBit));
661                //Figure out what the OF bit should be.
662                if ((ext & OFBit) && bits(SrcReg1, dataSize * 8 - 1))
663                    ccFlagBits = ccFlagBits | OFBit;
664                //Use the regular mechanisms to calculate the other flags.
665                ccFlagBits = genFlags(ccFlagBits, ext & ~(CFBit | ECFBit | OFBit),
666                        DestReg, psrc1, op2);
667            }
668        '''
669
670    class Sra(RegOp):
671        code = '''
672            uint8_t shiftAmt = (op2 & ((dataSize == 8) ? mask(6) : mask(5)));
673            // Because what happens to the bits shift -in- on a right shift
674            // is not defined in the C/C++ standard, we have to sign extend
675            // them manually to be sure.
676            uint64_t arithMask =
677                -bits(psrc1, dataSize * 8 - 1) << (dataSize * 8 - shiftAmt);
678            DestReg = merge(DestReg, (psrc1 >> shiftAmt) | arithMask, dataSize);
679            '''
680        flag_code = '''
681            // If the shift amount is zero, no flags should be modified.
682            if (shiftAmt) {
683                //Zero out any flags we might modify. This way we only have to
684                //worry about setting them.
685                ccFlagBits = ccFlagBits & ~(ext & (CFBit | ECFBit | OFBit));
686                //If some combination of the CF bits need to be set, set them.
687                if ((ext & (CFBit | ECFBit)) && bits(SrcReg1, shiftAmt - 1))
688                    ccFlagBits = ccFlagBits | (ext & (CFBit | ECFBit));
689                //Use the regular mechanisms to calculate the other flags.
690                ccFlagBits = genFlags(ccFlagBits, ext & ~(CFBit | ECFBit | OFBit),
691                        DestReg, psrc1, op2);
692            }
693        '''
694
695    class Ror(RegOp):
696        code = '''
697            uint8_t shiftAmt =
698                (op2 & ((dataSize == 8) ? mask(6) : mask(5)));
699            if(shiftAmt)
700            {
701                uint64_t top = psrc1 << (dataSize * 8 - shiftAmt);
702                uint64_t bottom = bits(psrc1, dataSize * 8, shiftAmt);
703                DestReg = merge(DestReg, top | bottom, dataSize);
704            }
705            else
706                DestReg = DestReg;
707            '''
708        flag_code = '''
709            // If the shift amount is zero, no flags should be modified.
710            if (shiftAmt) {
711                //Zero out any flags we might modify. This way we only have to
712                //worry about setting them.
713                ccFlagBits = ccFlagBits & ~(ext & (CFBit | ECFBit | OFBit));
714                //Find the most and second most significant bits of the result.
715                int msb = bits(DestReg, dataSize * 8 - 1);
716                int smsb = bits(DestReg, dataSize * 8 - 2);
717                //If some combination of the CF bits need to be set, set them.
718                if ((ext & (CFBit | ECFBit)) && msb)
719                    ccFlagBits = ccFlagBits | (ext & (CFBit | ECFBit));
720                //Figure out what the OF bit should be.
721                if ((ext & OFBit) && (msb ^ smsb))
722                    ccFlagBits = ccFlagBits | OFBit;
723                //Use the regular mechanisms to calculate the other flags.
724                ccFlagBits = genFlags(ccFlagBits, ext & ~(CFBit | ECFBit | OFBit),
725                        DestReg, psrc1, op2);
726            }
727        '''
728
729    class Rcr(RegOp):
730        code = '''
731            uint8_t shiftAmt =
732                (op2 & ((dataSize == 8) ? mask(6) : mask(5)));
733            if(shiftAmt)
734            {
735                CCFlagBits flags = ccFlagBits;
736                uint64_t top = flags.CF << (dataSize * 8 - shiftAmt);
737                if(shiftAmt > 1)
738                    top |= psrc1 << (dataSize * 8 - shiftAmt - 1);
739                uint64_t bottom = bits(psrc1, dataSize * 8, shiftAmt);
740                DestReg = merge(DestReg, top | bottom, dataSize);
741            }
742            else
743                DestReg = DestReg;
744            '''
745        flag_code = '''
746            // If the shift amount is zero, no flags should be modified.
747            if (shiftAmt) {
748                //Zero out any flags we might modify. This way we only have to
749                //worry about setting them.
750                ccFlagBits = ccFlagBits & ~(ext & (CFBit | ECFBit | OFBit));
751                //Figure out what the OF bit should be.
752                if ((ext & OFBit) && ((ccFlagBits & CFBit) ^
753                                      bits(SrcReg1, dataSize * 8 - 1)))
754                    ccFlagBits = ccFlagBits | OFBit;
755                //If some combination of the CF bits need to be set, set them.
756                if ((ext & (CFBit | ECFBit)) && bits(SrcReg1, shiftAmt - 1))
757                    ccFlagBits = ccFlagBits | (ext & (CFBit | ECFBit));
758                //Use the regular mechanisms to calculate the other flags.
759                ccFlagBits = genFlags(ccFlagBits, ext & ~(CFBit | ECFBit | OFBit),
760                        DestReg, psrc1, op2);
761            }
762        '''
763
764    class Rol(RegOp):
765        code = '''
766            uint8_t shiftAmt =
767                (op2 & ((dataSize == 8) ? mask(6) : mask(5)));
768            if(shiftAmt)
769            {
770                uint64_t top = psrc1 << shiftAmt;
771                uint64_t bottom =
772                    bits(psrc1, dataSize * 8 - 1, dataSize * 8 - shiftAmt);
773                DestReg = merge(DestReg, top | bottom, dataSize);
774            }
775            else
776                DestReg = DestReg;
777            '''
778        flag_code = '''
779            // If the shift amount is zero, no flags should be modified.
780            if (shiftAmt) {
781                //Zero out any flags we might modify. This way we only have to
782                //worry about setting them.
783                ccFlagBits = ccFlagBits & ~(ext & (CFBit | ECFBit | OFBit));
784                //The CF bits, if set, would be set to the lsb of the result.
785                int lsb = DestReg & 0x1;
786                int msb = bits(DestReg, dataSize * 8 - 1);
787                //If some combination of the CF bits need to be set, set them.
788                if ((ext & (CFBit | ECFBit)) && lsb)
789                    ccFlagBits = ccFlagBits | (ext & (CFBit | ECFBit));
790                //Figure out what the OF bit should be.
791                if ((ext & OFBit) && (msb ^ lsb))
792                    ccFlagBits = ccFlagBits | OFBit;
793                //Use the regular mechanisms to calculate the other flags.
794                ccFlagBits = genFlags(ccFlagBits, ext & ~(CFBit | ECFBit | OFBit),
795                        DestReg, psrc1, op2);
796            }
797        '''
798
799    class Rcl(RegOp):
800        code = '''
801            uint8_t shiftAmt =
802                (op2 & ((dataSize == 8) ? mask(6) : mask(5)));
803            if(shiftAmt)
804            {
805                CCFlagBits flags = ccFlagBits;
806                uint64_t top = psrc1 << shiftAmt;
807                uint64_t bottom = flags.CF << (shiftAmt - 1);
808                if(shiftAmt > 1)
809                    bottom |=
810                        bits(psrc1, dataSize * 8 - 1,
811                                   dataSize * 8 - shiftAmt + 1);
812                DestReg = merge(DestReg, top | bottom, dataSize);
813            }
814            else
815                DestReg = DestReg;
816            '''
817        flag_code = '''
818            // If the shift amount is zero, no flags should be modified.
819            if (shiftAmt) {
820                //Zero out any flags we might modify. This way we only have to
821                //worry about setting them.
822                ccFlagBits = ccFlagBits & ~(ext & (CFBit | ECFBit | OFBit));
823                int msb = bits(DestReg, dataSize * 8 - 1);
824                int CFBits = bits(SrcReg1, dataSize * 8 - shiftAmt);
825                //If some combination of the CF bits need to be set, set them.
826                if ((ext & (CFBit | ECFBit)) && CFBits)
827                    ccFlagBits = ccFlagBits | (ext & (CFBit | ECFBit));
828                //Figure out what the OF bit should be.
829                if ((ext & OFBit) && (msb ^ CFBits))
830                    ccFlagBits = ccFlagBits | OFBit;
831                //Use the regular mechanisms to calculate the other flags.
832                ccFlagBits = genFlags(ccFlagBits, ext & ~(CFBit | ECFBit | OFBit),
833                        DestReg, psrc1, op2);
834            }
835        '''
836
837    class Wrip(WrRegOp, CondRegOp):
838        code = 'RIP = psrc1 + op2'
839        else_code="RIP = RIP;"
840
841    class Br(WrRegOp, CondRegOp):
842        code = 'nuIP = psrc1 + op2;'
843        else_code='nuIP = nuIP;'
844
845    class Wruflags(WrRegOp):
846        code = 'ccFlagBits = psrc1 ^ op2'
847
848    class Rdip(RdRegOp):
849        code = 'DestReg = RIP'
850
851    class Ruflags(RdRegOp):
852        code = 'DestReg = ccFlagBits'
853
854    class Ruflag(RegOp):
855        code = '''
856            int flag = bits(ccFlagBits, imm8);
857            DestReg = merge(DestReg, flag, dataSize);
858            ccFlagBits = (flag == 0) ? (ccFlagBits | EZFBit) :
859                                       (ccFlagBits & ~EZFBit);
860            '''
861        def __init__(self, dest, imm, flags=None, \
862                dataSize="env.dataSize"):
863            super(Ruflag, self).__init__(dest, \
864                    "NUM_INTREGS", imm, flags, dataSize)
865
866    class Sext(RegOp):
867        code = '''
868            IntReg val = psrc1;
869            int sign_bit = bits(val, imm8-1, imm8-1);
870            uint64_t maskVal = mask(imm8);
871            val = sign_bit ? (val | ~maskVal) : (val & maskVal);
872            DestReg = merge(DestReg, val, dataSize);
873            '''
874
875    class Zext(RegOp):
876        code = 'DestReg = bits(psrc1, imm8-1, 0);'
877}};
878