File r38/lisp/csl/cslbase/instructions.c artifact 4d7de42779 part of check-in 9992369dd3


/*
instructions.c

Copyright (C) 2003-2006 Gil Dabah, http://ragestorm.net/distorm/
This library is licensed under the BSD license. See the file COPYING.
*/


#include "instructions.h"

#include "insts.h"
#include "prefix.h"
#include "textdefs.h"
#include "x86defs.h"
#include "wstring.h"

// I use the trie data structure as I found it most fitting to a disassembler mechanism.
// When you read a byte and have to decide if it's enough or you should read more bytes, 'till you get to the instruction information.
// It's really fast because you POP the instruction info in top 3 iterates on the DB, because an instruction can be formed from two bytes + 3 bits reg from the ModR/M byte.
// For a simple explanation, check this out:
// http://www.csse.monash.edu.au/~lloyd/tildeAlgDS/Tree/Trie/
// Futher reading: http://en.wikipedia.org/wiki/Trie

/*

The first GATE (array you read from in a trie data structure), as I call them, is statically allocated by the compiler.
The second and third gates if used are being allocated dynamically by the instructions-insertion functionality.

How would such a thing look in memory, say we support 4 instructions with 3 bytes top (means 2 dynamically allocated gates).

->
|-------|                                0,
|0|     -------------------------------> |-------|
|1|RET  |      1,                        |0|AND  |
|2|     -----> |-------|                 |1|XOR  |
|3|INT3 |      |0|PUSH |                 |2|OR   |         0,3,
|-------|      |1|POP  |                 |3|     --------->|-------|
               |2|PUSHF|                 |-------|         |0|ROR  |
               |3|POPF |                                   |1|ROL  |
               |-------|                                   |2|SHR  |
                                                           |3|SHL  |
                                                           |-------|

Of course, this is NOT how Intel instructions set looks!!!
but I just wanted to give a small demonstration.
Now the instructions you get from such a trie DB goes like this:

0, 0 - AND
0, 1 - XOR
0, 2 - OR
0, 3, 0, ROR
0, 3, 1, ROL
0, 3, 2, SHR
0, 3, 3, SHL
1 - RET
2, 0 - PUSH
2, 1 - POP
2, 2 - PUSHF
2, 3 - POPF
3 - INT3

I guess it's clear by now.
So now, if you read 0, you know that you have to enter the second gate(list) with the second byte specifying the index.
But if you read 1, you know that you go to an instruction (in this case, a RET).
That's why there's an Instruction-Node structure, it tells you whether you got to an instruction or another list
so you should keep on reading byte).

In Intel, you could go through 3 gates at top, because there're instructions which are built from 2 bytes and another smaller list
for the REG part.
Therefore, Intel's first gate is 256 long, and other gates are 256 or 8 long, yes, it costs pretty much alot of memory
for non-used defined instructions, but I think that it still rocks.
*/

// This function is reponsible to return the instruction information of the first found in code.
// It returns the _InstInfo of the found instruction, otherwise NULL.
// code should point to the ModR/M byte upon exit (if used), or after the instruction binary code itself.
// This function is NOT decoding-type dependant, it is up to the caller to see whether the instruction is valid.
// Get the instruction info, using a Trie data structure.
// I call it "basic", because it simply locates an instruction, it doesn't care what bytes it's using, such as prefixes.
//
_InstInfo* locate_basic_inst(const unsigned char** code0, long* codeLen0, _OffsetType* codeOffset0, _WString* instructionHex, int isERXPrefixValid, _DecodeType dt)
{
	const unsigned char* code = *code0;
	long codeLen = *codeLen0;
	_OffsetType codeOffset = *codeOffset0;

	unsigned char tmpIndex0 = *code, tmpIndex1 = 0, tmpIndex2 = 0;
	_InstNode *in = NULL;
	_InstInfo *ii = NULL;

	// Single byte instruction (OCST_1BYTE).
	if (Instructions[tmpIndex0].type == INT_INFO) {
		str_hex_b(instructionHex, tmpIndex0);

		codeLen -= 1;
		if (codeLen < 0) return NULL;
		code += 1;
		codeOffset += 1;
		*code0 = code;
		*codeLen0 = codeLen;
		*codeOffset0 = codeOffset;

		// ARPL/MOVSXD share the same instruction number, and both have differenct operands and mnemonics, of course.
		// Practically, I couldn't come up with a comfortable way to merge the operands' types of ARPL/MOVSXD.
		// And since the DB can't be patched dynamically, because the DB has to be multi-threaded compliant,
		// I have no choice but to check for ARPL/MOVSXD right here - "right about now, the funk soul brother, check it out now, the funk soul brother...", fatboy slim
		if (tmpIndex0 == INST_ARPL_INDEX) return dt == Decode64Bits ? &II_movsxd : &II_arpl;

		return Instructions[tmpIndex0].ii;
	}

	// Single byte instruction + reg bits (OCST_13BYTES).
	if (Instructions[tmpIndex0].type == INT_LIST_GROUP) {
		str_hex_b(instructionHex, tmpIndex0);

		codeLen -= 1;
		if (codeLen <= 0) return NULL;
		code += 1;
		codeOffset += 1;
		*code0 = code;
		*codeLen0 = codeLen;
		*codeOffset0 = codeOffset;
		return Instructions[tmpIndex0].list[(*code >> 3) & 7].ii;
	}

	// Single byte instruction + reg byte OR one whole byte (OCST_1dBYTES).
	if (Instructions[tmpIndex0].type == INT_LIST_DIVIDED) {
		str_hex_b(instructionHex, tmpIndex0);

		codeLen -= 1;
		if (codeLen <= 0) return NULL;
		code += 1;
		codeOffset += 1;

		tmpIndex1 = *code;

		if (tmpIndex1 < INST_DIVIDED_MODRM) {
			// An instruction which requires a ModR/M byte. Thus it's 1.3 bytes long instruction.
			tmpIndex1 = (tmpIndex1 >> 3) & 7; // Isolate the 3 REG/OPCODE bits.
		} else { // Normal 2 bytes instruction.
			str_hex_b(instructionHex, tmpIndex1);

			codeLen -= 1;
			if (codeLen < 0) return NULL;
			code += 1;
			codeOffset += 1;

			// Divided instructions can't be in the range of 0x8-0xc0.
			// That's because 0-8 are used for 3 bits group.
			// And 0xc0-0xff are used for not-divided instruction.
			// So the inbetween range is omitted, thus saving some more place in the tables.
			tmpIndex1 -= INST_DIVIDED_MODRM - 8;
		}

		*code0 = code;
		*codeLen0 = codeLen;
		*codeOffset0 = codeOffset;
		return Instructions[tmpIndex0].list[tmpIndex1].ii;
	}

	// At least 2 bytes long instruction (might be 2.3 or 3 or 3.3 bytes long).
	if (Instructions[tmpIndex0].type == INT_LIST_FULL) {
		str_hex_b(instructionHex, tmpIndex0);

		if (isERXPrefixValid) {
			// Skip REX prefix byte.
			codeLen -= 1;
			if (codeLen <= 0) return NULL;
			code += 1;
			codeOffset += 1;
			str_hex_sp_b(instructionHex, *code);
			chrcat_WS(instructionHex, SP_CHR);
		}

		codeLen -= 1;
		if (codeLen <= 0) return NULL;
		code += 1;
		codeOffset += 1;

		tmpIndex1 = *code;
		in = &Instructions[tmpIndex0].list[tmpIndex1];

		// This is where we check if we just read two escape bytes in a row, which means it is a 3DNow! instruction.
		if ((tmpIndex0 == _3DNOW_ESCAPE_BYTE) && (tmpIndex1 == _3DNOW_ESCAPE_BYTE)) {
			str_hex_b(instructionHex, tmpIndex1);

			codeLen -= 1;
			if (codeLen < 0) return NULL;
			code += 1;
			codeOffset += 1;

			*code0 = code;
			*codeLen0 = codeLen;
			*codeOffset0 = codeOffset;
			return &II_3dnow;
		}

		// 2 bytes instruction (OCST_2BYTES).
		if (in->type == INT_INFO) {
			str_hex_b(instructionHex, tmpIndex1);

			codeLen -= 1;
			if (codeLen < 0) return NULL;
			code += 1;
			codeOffset += 1;

			*code0 = code;
			*codeLen0 = codeLen;
			*codeOffset0 = codeOffset;
			return in->ii;
		}

		// 2 bytes + reg instruction (OCST_23BYTES).
		if (in->type == INT_LIST_GROUP) {
			str_hex_b(instructionHex, tmpIndex1);

			codeLen -= 1;
			if (codeLen <= 0) return NULL;
			code += 1;
			codeOffset += 1;

			*code0 = code;
			*codeLen0 = codeLen;
			*codeOffset0 = codeOffset;
			return in->list[(*code >> 3) & 7].ii;
		}

		// 2 bytes + divided range (OCST_2dBYTES).
		if (in->type == INT_LIST_DIVIDED) {
			str_hex_b(instructionHex, tmpIndex1);

			codeLen -= 1;
			if (codeLen <= 0) return NULL;
			code += 1;
			codeOffset += 1;

			tmpIndex2 = *code;
			ii = in->list[(tmpIndex2 >> 3) & 7].ii;

			// If the instruction wasn't divided (but still it must be a 2.3) or it was an official 2.3 (because its index was less than 0xc0)
			// Then it means the instruction should be used the REG bits, otherwise give a chance to range 0xc0-0xff.
			if ((ii != NULL) && ((ii->flags & INST_NOT_DIVIDED) || (tmpIndex2 < INST_DIVIDED_MODRM))) tmpIndex2 = (tmpIndex2 >> 3) & 7;
			else if (tmpIndex2 >= INST_DIVIDED_MODRM) tmpIndex2 -= INST_DIVIDED_MODRM - 8; // Divided tables are smaller, range 0x8-0xc0 is omitted.
			ii = in->list[tmpIndex2].ii;
			if ((ii != NULL) && ((ii->flags & INST_INCLUDE_MODRM) == 0)) { // Read 3 whole bytes.
				str_hex_b(instructionHex, tmpIndex2);

				codeLen -= 1;
				if (codeLen < 0) return NULL;
				code += 1;
				codeOffset += 1;
			}

			*code0 = code;
			*codeLen0 = codeLen;
			*codeOffset0 = codeOffset;
			return ii;
		}

		// At least 3 bytes (OCST_3BYTES).
		if (in->type == INT_LIST_FULL) {
			str_hex_b(instructionHex, tmpIndex1);

			codeLen -= 1;
			if (codeLen < 0) return NULL;
			code += 1;
			codeOffset += 1;

			tmpIndex2 = *code;

			// 3.3 bytes (OCST_33BYTES).
			if (in->list[tmpIndex2].type == INT_LIST_GROUP) {
				str_hex_b(instructionHex, tmpIndex2);

				codeLen -= 1;
				if (codeLen <= 0) return NULL;
				code += 1;
				codeOffset += 1;

				*code0 = code;
				*codeLen0 = codeLen;
				*codeOffset0 = codeOffset;
				return in->list[tmpIndex2].list[(*code >> 3) & 7].ii;
			} else { // OCST_3BYTES.
				str_hex_b(instructionHex, tmpIndex2);

				codeLen -= 1;
				if (codeLen < 0) return NULL;
				code += 1;
				codeOffset += 1;

				*code0 = code;
				*codeLen0 = codeLen;
				*codeOffset0 = codeOffset;
				return in->list[tmpIndex2].ii;
			}
		}
	}

	// Kahtchinggg, damn.
	return NULL;
}

// Locate an instruction, give chance to SSE/2/3 instructions which need the prefix as a nomral byte instruction!
_InstInfo* locate_inst(const unsigned char** code,  long* codeLen, _OffsetType* codeOffset, _WString* instructionHex, _PrefixState* ps, _DecodeType dt)
{
	_InstInfo* ii = NULL;

	unsigned long lastCodeLen = *codeLen;
	_OffsetType lastCodeOffset = *codeOffset;
	const unsigned char* lastCode = *code;

	unsigned char lastBC = 0; // byte code.

	// Keep last byte code, raed from ps->last only if we skipped prefixes.
	if (ps->start < ps->last) {
		if (ps->isREXPrefixValid) lastBC = *(ps->last - 2); // Skip REX prefix byte.
		else lastBC = *(ps->last-1);
	}
	/*
	Sometimes normal prefixes become mandatory prefixes, which means they are now part of the instruction opcode bytes.

	This is a bit tricky now,
	if the first byte is a REP (F3) prefix, we will have to give a chance to an SSE instruction.
	If an instruction doesn't exist, we will make it as a prefix and re-locateinst.
	A case sach that a REP prefix is being changed into an instruction byte and also an SSE instruction will not be found can't happen,
	simply because there are no collisions between string instruction and SSE instructions (they are escaped).

	As for SSE2/3, check for F2 and 66 as well.

	In 64 bits, we have to make sure that we will skip the REX prefix, if it exists.
	There's a specific case, where a 66 is mandatory but it was dropped because REG.W was used,
	but it doesn't behave as an operand size prefix but as a mandatory, so we will have to take it into account.

	For example (64 bits decoding mode):
	66 98 CBW
	48 98 CDQE
	66 48 98: db 0x66; CDQE
	Shows that operand size is dropped.

	Now, it's a mandatory prefix and NOT an operand size one.
	66480f2dc0 db 0x48; CVTPD2PI XMM0, XMM0
	Although this instruction doesn't require a REX.W, it just shows, that even if it did - it doesn't matter.
	REX.W is dropped because it's not requried, but the decode function disabled the operand size even so.
	*/
	if ((lastBC == PREFIX_REP) || (lastBC == PREFIX_REPNZ) || (lastBC == PREFIX_OP_SIZE)) {
		// Take the prefix into account, making it an instruction byte.
		if (ps->isREXPrefixValid) {
			(*code) -= 2;
			(*codeLen) += 2;
			(*codeOffset) -= 2;
		} else {
			(*code)--;
			(*codeLen)++;
			(*codeOffset)--;
		}
		ii = locate_basic_inst(code, codeLen, codeOffset, instructionHex, ps->isREXPrefixValid, dt);
		if (ii) {
			// An SSE/2/3 instruction was found,
			// so remove the last prefix. It is now a mandatory prefix and skip REX as usual.

			// Let the decode function know that we turned a prefix into mandatory one.
			// So if some problem occurs later, we will know to handle it specially because of this end case.
			if (ps->isREXPrefixValid) {
				ps->last -= 2;
				ps->specialPrefixesSize = 2; // REX + SSE.
			}
			else {
				ps->last--;
				ps->specialPrefixesSize = 1; // SSE only.
			}

			// Remove that flag specifically.
			switch (lastBC)
			{
				case PREFIX_REP: ps->totalPrefixes &= ~INST_PRE_LOKREP_MASK; ps->lokrepPos = NULL; break;
				case PREFIX_REPNZ: ps->totalPrefixes &= ~INST_PRE_LOKREP_MASK; ps->lokrepPos = NULL; break;
				case PREFIX_OP_SIZE: ps->totalPrefixes &= ~INST_PRE_OP_SIZE; ps->opsizePos = NULL; break;
			}
		} else {
			// Undo:
			strclear_WS(instructionHex); // Remove outputed prefix.
			*code = lastCode;
			*codeLen = lastCodeLen;
			*codeOffset = lastCodeOffset;
		}
	}
	// No instruction was found before, so it's a "simple" one:
	if (ii == NULL) ii = locate_basic_inst(code, codeLen, codeOffset, instructionHex, 0, dt);
	return ii;
}

/*
3DNow! instruction handling:

This is used when we encounter a 3DNow! instruction.
We can't really locate a 3DNow! instruction before we see two escaped bytes,
0x0f, 0x0f. Then we have to extract operands which are, dest=mmx register, src=mmx register or quadword indirection.
When we are finished with the extraction of operands we can resume to locate the instruction by reading another byte
which tells us which 3DNow instruction we really tracked down...
So in order to tell the extract operands function which operands the 3DNow! instruction require, we need to set up some
generic instruction info for 3DNow! instructions.

In the locate_inst itself, when we read an OCST_3BYTES which the two first bytes are 0x0f and 0x0f.
we will return this special generic II for the specific operands we are interested in (MM, MM64).
Then after extracting the operand, we'll call a completion routine for locating the instruction
which will be called only for 3DNow! instructions, distinguished by a flag, and it will read the last byte of the 3 bytes.
*/
_InstInfo II_3dnow = {OT_MM, OT_MM64, "_3DNow! II", INST_32BITS | INST_INCLUDE_MODRM | INST_3DNOW_FETCH | INST_3DNOW};

_InstInfo* locate_3dnow_inst(_CodeInfo* ci, _WString* instructionHex)
{
	unsigned char tmpIndex2 = *ci->code;
	// Start off from the two escape bytes gates...
	_InstNode* in = &Instructions[_3DNOW_ESCAPE_BYTE].list[_3DNOW_ESCAPE_BYTE];

	if (in->list[tmpIndex2].type == INT_INFO) {
		str_hex_sp_b(instructionHex, tmpIndex2);

		ci->codeLen -= 1;
		if (ci->codeLen < 0) return NULL;
		ci->code += 1;
		ci->codeOffset += 1;
	}
	return in->list[tmpIndex2].ii;
}

/* Formats the string according to the operand type given.
I.E: ADD [BX], 5, how do you know the size of the destination in memory, if it is a BYTE destination or a WORD destination?
So the output for instance is: ADD *BYTE* [BX], 5

We use the DT for cases that prefixes affect the decoding.
If you decode in 16 bits then a pointer(indirection) will be to a WORD size.
If you decode in 32 bits then a pointer will be to a DWORD size.
The catch is that an OPERAND size prefix switches the default of these statements, I.E:
83 17 05 in 32 bits, decoded to: ADC DWORD [EDI], +0x05.
So if you add a 67(ADDR SIZE Prefix) it will affect the pointer itself (EDI) to be decoded as if it were decoded in 16 bits,
in short term, affecting the INDIRECTION size.
But how can you POINT to 16 bits if you decode in 32 bits? You use the operand size, I.E:
66 83 17 05 in 32 bits, decoded to: ADC WORD [EDI], +0x05.

It's a bit confusing, 2 guidelines to follow:
Operand Size prefix affects the size of the value whether it's an immediate or the value in memory which is being pointed to!
And Address Size prefix affects the decoding of the value of the pointer(indirection)!

Anyways, you have to write the size explicitly when it's uncleared from reading the operands.

Take in account second operand type, because if we know what is the second operand type (in case one exists),
the prefix is useless(I.E: ADD [bx], ax).
However, we have such cases where you need the prefix anyways(I.E: MOVZX ebx, byte [edi]), which means that you have
to add a prefix when the destination variable size is greater than the source variable size...
*/
void str_indirection_text(_WString* s, _OpType ot, _OpType ot2, _DecodeType dt, _PrefixState* ps)
{
	switch (ot)
	{
		case OT_MM32:
			strcat_WSN(s, TEXT_32_BITS);
		return;
		case OT_MM64:
			strcat_WSN(s, TEXT_64_BITS);
		return;
		case OT_XMM32:
			strcat_WSN(s, TEXT_32_BITS);
		return;
		case OT_XMM64:
			strcat_WSN(s, TEXT_64_BITS);
		return;
		case OT_XMM128:
			strcat_WSN(s, TEXT_128_BITS);
		return;
		case OT_FPUM16:
			strcat_WSN(s, TEXT_16_BITS);
		return;
		case OT_FPUM32:
			strcat_WSN(s, TEXT_32_BITS);
		return;
		case OT_FPUM64:
			strcat_WSN(s, TEXT_64_BITS);
		return;
		case OT_FPUM80:
			strcat_WSN(s, TEXT_80_BITS);
		return;
		case OT_R32M16:
			strcat_WSN(s, TEXT_16_BITS);
		return;
		case OT_RFULL_M16:
			// MOV OT_RFULL_M16, SREG eats operand size prefix without any effect.
			// This is because stupid assemblers insert the 0x66 prefix for this instruction anyways.
			ps->usedPrefixes |= (ps->totalPrefixes & INST_PRE_OP_SIZE);
			strcat_WSN(s, TEXT_16_BITS);
		return;
                default:
                return;
	}
	switch (ot2)
	{ // Used by MOVSX/MOVZX, in a case like: MOVZX EBP, *BYTE* [EBX].
		case OT_REG_FULL:
			// Special case - notice that the destination is 32 or 64 bits and the source is RM(!) 8.
			if (ot == OT_RM8) {
				strcat_WSN(s, TEXT_8_BITS);
			}
		return;
		case OT_REG32_64:
			// Special case - notice that the destination is 32 or 64 bits and the source is RM(!) 8.
			if (ot == OT_RM16) {
				strcat_WSN(s, TEXT_16_BITS);
			}
		return;
                default:
                return;
	}
	if (ot != OT_MEM1616) { // No need if it's [16:16], in a case like: BOUND EAX, [EBX].
		// Rolls, Shifts, Bit Operations.
		if (ot2 == OT_NONE || ot2 == OT_IMM8 || ot2 == OT_IMM_FULL || ot2 == OT_IMM32 || ot2 == OT_SEIMM8 || ot2 == OT_REGCL || ot2 == OT_CONST1) {
			switch (ot)
			{
				case OT_RM8:
					strcat_WSN(s, TEXT_8_BITS);
				return;
				case OT_RM_FULL:
					ps->usedPrefixes |= (ps->totalPrefixes & INST_PRE_OP_SIZE);
					if (dt == Decode32Bits) {
						strcat_WSN(s, TEXT_32_BITS);
						return;
					} else if (dt == Decode64Bits) {
						ps->usedPrefixes |= INST_PRE_REX;
						strcat_WSN(s, TEXT_64_BITS);
						return;
					}
				/* FALL THROUGH CUZ dt==Decoded16Bits @-<----*/
				case OT_RM16:
					ps->usedPrefixes |= (ps->totalPrefixes & INST_PRE_OP_SIZE);
					strcat_WSN(s, TEXT_16_BITS);
				return;
				case OT_RM32:
					ps->usedPrefixes |= (ps->totalPrefixes & INST_PRE_OP_SIZE);
					strcat_WSN(s, TEXT_32_BITS);
				return;
                                default:
                                return;
			}
		}
	}
}


REDUCE Historical
REDUCE Sourceforge Project | Historical SVN Repository | GitHub Mirror | SourceHut Mirror | NotABug Mirror | Chisel Mirror | Chisel RSS ]