URL
https://opencores.org/ocsvn/forwardcom/forwardcom/trunk
Subversion Repositories forwardcom
[/] [forwardcom/] [bintools/] [assem4.cpp] - Rev 101
Go to most recent revision | Compare with Previous | Blame | View Log
/**************************** assem4.cpp ******************************** * Author: Agner Fog * Date created: 2017-04-17 * Last modified: 2021-07-14 * Version: 1.11 * Project: Binary tools for ForwardCom instruction set * Module: assem.cpp * Description: * Module for assembling ForwardCom .as files. * This module contains: * pass3(): Interpretation of code lines. * Copyright 2017-2021 GNU General Public License http://www.gnu.org/licenses ******************************************************************************/ #include "stdafx.h" // Interpret lines. Generate code and data void CAssembler::pass3() { uint16_t last_line_type = 0; // type of preceding line makeFormatLists(); // make formatList3 and formatList4 code_size = cmd.codeSizeOption; // initialize options data_size = cmd.dataSizeOption; section = 0; iLoop = iIf = iSwitch = 0; // index of current high level statements // lines loop for (linei = 1; linei < lines.numEntries()-1; linei++) { tokenB = lines[linei].firstToken; // first token in line tokenN = lines[linei].numTokens; // number of tokens in line if (tokenN == 0 || lines[linei].type == LINE_ERROR || lines[linei].type == LINE_METADEF) continue; lineError = false; switch (lines[linei].type) { case LINE_DATADEF: if (last_line_type == LINE_CODEDEF && (lines[linei].sectionType & SHF_EXEC)) { /* currently, the assembler cannot mix code and data because they are put in different buffers. The only way to hard-code instructions is to put them into a separate section. */ errors.reportLine(ERR_MIX_DATA_AND_CODE); // data definition in code section } break; case LINE_CODEDEF: interpretCodeLine(); if (last_line_type == LINE_DATADEF && !(lines[linei].sectionType & SHF_EXEC)) { errors.reportLine(ERR_MIX_DATA_AND_CODE); // code definition in data section } break; case LINE_METADEF: case LINE_ERROR: continue; case LINE_FUNCTION: interpretFunctionDirective(); break; case LINE_SECTION: interpretSectionDirective(); break; case LINE_ENDDIR: interpretEndDirective(); break; case LINE_OPTIONS: interpretOptionsLine(); break; } last_line_type = lines[linei].type; } while (hllBlocks.numEntries()) { // unfinished block SBlock block = hllBlocks.pop(); errors.report(tokens[block.startBracket].pos, tokens[block.startBracket].stringLength, ERR_BRACKET_BEGIN); } } // extract subsets of formatList (in disasm1.cpp) for multiformat instructions and jump instructions void CAssembler::makeFormatLists() { uint32_t i; for (i = 0; i < formatListSize; i++) { if (formatList[i].category == 3) formatList3.push(formatList[i]); if (formatList[i].category == 4) formatList4.push(formatList[i]); } } // Interpret a line defining code. This covers both assembly style and high level style code void CAssembler::interpretCodeLine() { uint32_t tok; // token index dataType = 0; // data type for current instruction uint32_t nReg = 0; // number of register source operands uint32_t state = 0; /* state during interpretation of line. example: L1: int32 r1 = compare(r2, 5), option = 2 // assembly style L1: int32 r1 = r2 < 5 // same in high level style 0: begin 1: after label 2: after label: 3: after type 4: after destination 5: after destination = (expecting expression or instruction) 6: after expression or instruction() 7: after instruction 8: after instruction( 9: after operand 10: after instruction(), 11: after jump instruction */ SExpression expr; // evaluated expression SCode code; // current instruction code zeroAllMembers(code); // reset code structure if (section == 0) { errors.reportLine(ERR_CODE_WO_SECTION); } // high level instructions with nothing before can be caught already here if (tokens[tokenB].type == TOK_HLL) { interpretHighLevelStatement(); // if, else, switch, for, do, while (){} statements return; } if (tokens[tokenB].type == TOK_OPR && tokens[tokenB].id == '}') { interpretEndBracket(); // end of {} block return; } // interpret line by state machine looping through tokens for (tok = tokenB; tok < tokenB + tokenN; tok++) { SToken token = tokens[tok]; if (token.type == TOK_XPR && expressions[token.value.w].etype & XPR_REG) { // this is an alias for a register. Translate to register token.type = TOK_REG; token.id = expressions[token.value.w].reg1; } if (lineError) break; code.section = section; if (state == 5) { // after '=' if (token.type == TOK_INS) { // instruction if (code.instruction) errors.report(token); // instruction after += etc. code.instruction = token.id; state = 7; } else { // expression after equal sign // interpret expression representing operands and operator expr = expression(tok, tokenB + tokenN - tok, 0); if (lineError) return; if (code.instruction) { // += operator etc. already encountered. combine the operands uint32_t op = code.instruction; code.instruction = 0; code.reg1 = code.dest; // first source operand is same as destination code.etype |= XPR_REG1; code.tokens = 0; expr = op2(op, code, expr); // operation '+' for '+=', etc. code.instruction = 0; code.reg1 = 0; } if (code.etype & XPR_ERROR) { errors.reportLine(code.value.w); // report error } // ordinary '=' goes here if (lineError) return; insertAll(code, expr); tok += expr.tokens - 1; state = 6; } } else if (state == 11) { // interpret jump target expr = expression(tok, tokenB + tokenN - tok, 0); state = 6; if (expr.etype & XPR_REG) { code = code | expr; tok += expr.tokens - 1; } else if (expr.etype & (XPR_INT | XPR_SYM1)) { code.sym5 = expr.sym3 ? expr.sym3 : expr.sym1; code.offset_jump = expr.value.w; if (expr.value.w & 3) errors.report(token.pos, token.stringLength, ERR_JUMP_TARGET_MISALIGN); tok += expr.tokens - 1; code.etype |= XPR_JUMPOS | (expr.etype & ~XPR_IMMEDIATE); } else { errors.report(token.pos, token.stringLength, ERR_EXPECT_JUMP_TARGET); break; } } else if (state == 8 && token.type != TOK_OPT && token.type != TOK_REG) { // expression in parameter list if (token.type == TOK_OPR && token.id == ')') { state = 6; break; // end of parameter list } // interpret any expression, except register or option expr = expression(tok, tokenB + tokenN - tok, 0); tok += expr.tokens - 1; if (code.etype & expr.etype & XPR_INT) { // multiple immediate integer constants if (code.etype & XPR_INT2) { // three integer operands if (code.etype & XPR_OPTIONS) errors.report(token.pos, token.stringLength, ERR_TOO_MANY_OPERANDS); code.optionbits = uint8_t(expr.value.w); code.etype |= XPR_OPTIONS; expr.value.u = 0; } else { // two integer operands if (code.value.u >> 32 != 0) errors.report(token.pos, token.stringLength, ERR_TOO_MANY_OPERANDS); code.value.u = code.value.w | expr.value.u << 32; code.etype |= XPR_INT2; expr.value.u = 0; } } else if (expr.etype & XPR_MEM) { if (expr.etype & XPR_OFFSET) code.offset_mem += expr.offset_mem; //else code.offset += expr.value.i; if (expr.etype & XPR_IMMEDIATE) { // both memory and immediate operands code.value.i = expr.value.i; } } else if (expr.etype & XPR_IMMEDIATE) { code.value.i = expr.value.i; } expr.value.i = 0; code = code | expr; state = 9; } else { switch (token.type) { case TOK_LAB: case TOK_SYM: if (state == 0) { //code.label = token.value.w; code.label = token.id; if (code.label) { int32_t symi = findSymbol(code.label); if (symi > 0) symbols[symi].st_section = section; } state = 1; } else goto ST_ERROR; break; case TOK_OPR: if (token.id == ':' && state == 1) { state = 2; } else if (token.id == '+' && state == 3) { code.dtype |= TYP_PLUS; } else if (token.priority == 15 && state == 4) { // assignment operator state = 5; if (token.id & EQ) { // combined operator and assignment: += -= *= etc. code.reg1 = code.dest; code.etype |= XPR_REG | XPR_REG1; code.instruction = token.id & ~EQ; // temporarily store operator in .instruction } else if (token.id != '=') errors.report(token); } else if (token.id == '=' && state == 11) { state = 12; } else if (token.id == ',' && state == 6) { state = 10; } else if (token.id == ',' && state == 9) { state = 8; } else if (token.id == '(' && state == 7) { state = 8; } else if (token.id == ')' && (state == 8 || state == 9)) { state = 6; } else if (token.id == '[' && (state == 0 || state == 2 || state == 3)) { // interpret memory destination expr = expression(tok, tokenB + tokenN - tok, 0); tok += expr.tokens - 1; insertMem(code, expr); code.dest = 2; state = 4; } else if (token.id == '[' && state == 7 && code.instruction == II_ADDRESS) { // address []. expect memory operand expr = expression(tok, tokenB + tokenN - tok, 0); tok += expr.tokens - 1; insertMem(code, expr); state = 6; } else if ((token.id == '+' + D2 || token.id == '-' + D2) && (state == 3 || state == 4)) { // ++ and -- operators code.instruction = (token.id == '+' + D2) ? II_ADD : II_SUB; // operand is 1, integer or float if (dataType & TYP_FLOAT) { code.value.d = 1.0; code.etype |= XPR_FLT; } else { code.value.i = 1; code.etype |= XPR_INT; } if (state == 3) { // prefix operator. expect register tok++; if (token.type != TOK_REG) errors.report(token); code.dest = token.id; } code.reg1 = code.dest; code.etype |= XPR_REG1; state = 6; } else if (token.id == ';') {} // ignore terminating ';' else goto ST_ERROR; break; case TOK_TYP: if (state == 0 || state == 2) { dataType = code.dtype = token.id; state = 3; } else goto ST_ERROR; break; case TOK_REG: if (state == 0 || state == 2 || state == 3) { code.dest = uint8_t(token.id); state = 4; } else if (state == 8) { if (nReg < 3) { (&code.reg1)[nReg] = (uint8_t)token.id; // insert register in expression code.etype |= XPR_REG1 << nReg++; if ((code.etype & (XPR_INT | XPR_FLT | XPR_MEM)) && code.dest != 2) errors.report(token.pos, token.stringLength, ERR_OPERANDS_WRONG_ORDER); } else errors.report(token.pos, token.stringLength, ERR_TOO_MANY_OPERANDS); state = 9; } else goto ST_ERROR; break; case TOK_XPR: if (token.value.u >= expressions.numEntries()) goto ST_ERROR; // expression not found if (expressions[token.value.w].etype & XPR_MEM) { // this is an alias for a memory operand insertMem(code, expressions[token.value.w]); code.dest = 2; state = 4; } else goto ST_ERROR; break; case TOK_INS: if (state == 0 || state == 2 || state == 3) { // interpret instruction name code.instruction = token.id; state = 7; // expect parenthesis and parameters if (code.instruction & II_JUMP_INSTR) { // Jump or call instruction. The next may be a jump target, a register or a memory operand state = 11; // expect jump target // Check if there is a memory operand for (uint32_t tok2 = tok+1; tok2 < tokenB + tokenN; tok2++) { if (tokens[tok2].type == TOK_OPR && tokens[tok2].id == '[') { // a jump instruction with memory operand is treated as a normal instruction state = 7; break; } } } } else if ((state == 6 || state == 10) && (token.id & II_JUMP_INSTR)) { // second half of jump instruction code.instruction |= token.id; // combine two partial instruction names state = 11; // expect jump target } else goto ST_ERROR; break; case TOK_OPT: // option keyword expr = expression(tok, tokenB + tokenN - tok, 4); // this will read option = value tok += expr.tokens - 1; code.etype |= expr.etype; if (expr.etype & XPR_LIMIT) { code.value.i = expr.value.i; if (expr.value.u >= 0x100000000U) { // limit too high errors.report(tokens[tok - 1].pos, tokens[tok - 1].stringLength, ERR_LIMIT_TOO_HIGH); } } if (expr.etype & (XPR_LENGTH | XPR_BROADC)) code.length = expr.length; if (expr.etype & XPR_MASK) code.mask = expr.mask; if (expr.etype & XPR_FALLBACK) code.fallback = expr.fallback; if (expr.etype & XPR_OPTIONS) code.optionbits = expr.optionbits; if (state == 8) state = 9; else if (state == 6 || state == 10) state = 6; else goto ST_ERROR; break; case TOK_ATT: if (token.id == ATT_ALIGN && state == 0 && tokenN >= 2) { // align n directive code.instruction = II_ALIGN; expr = expression(tok + 1, tokenB + tokenN - tok - 1, 0); tok = tokenB + tokenN; code.value.u = expr.value.u; code.sizeUnknown = 0x80; if ((code.value.u & (code.value.u - 1)) || code.value.u > MAX_ALIGN || (expr.etype & XPR_IMMEDIATE) != XPR_INT || (expr.etype & (XPR_REG|XPR_OPTION|XPR_MEM))) { errors.reportLine(ERR_ALIGNMENT); } } else goto ST_ERROR; break; case TOK_HLL: // high level directive: if, else, while, for, etc. interpretHighLevelStatement(); return; default:; ST_ERROR: errors.report(token); break; } } } if (lineError) return; // check if state machine ends with a finished instruction if (state != 0 && state != 2 && state != 6 && state != 7) { errors.report(tokens[tok-1].pos, tokens[tok-1].stringLength, ERR_UNFINISHED_INSTRUCTION); return; } // move and store instruction has no operator yet if (code.instruction == 0 && code.etype) { if (code.dest == 2) code.instruction = II_STORE; // store to memory else { code.instruction = II_MOVE; // move constant to register if (cmd.optiLevel && (code.etype & XPR_INT) && code.value.i >= 0 && !code.sym3 && (code.dtype & TYP_INT) && (code.dest & REG_R)) { code.dtype |= TYP_PLUS; // optimize to larger type for positive constant because it is zero-extended anyway } } } if (code.instruction) { // a code record with no instruction represents a label only // code record contains instruction if (code.etype & XPR_JUMPOS) mergeJump(code); checkCode1(code); if (lineError) return; // find an instruction variant that fits fitCode(code); if (lineError) return; } // save code structure codeBuffer.push(code); } // Check how many bits are needed to contain immediate constant of an instruction. // The result is returned as bit-flags in code.fitNumX. // The return value is nonzero if the size cannot be resolved yet. int CAssembler::fitConstant(SCode & code) { int64_t value = 0; // the constant or address to fit int64_t valueScaled; // value divided by scale factor double dvalue = 0; // floating point value if needed bool floatType = false; // a floating point type is needed bool floatConst = false; // a floating point constant is provided uint32_t fitNum = 0; // return value uint32_t sym3 = 0, sym4 = 0; // symbols int32_t isym3 = 0, isym4 = 0; // symbol index int32_t uncertainty; // maximum deviance if the value is uncertain int uncertain = 0; // return value int symscale; // scaling of difference between symbols if (code.instruction == II_ALIGN) return 0; // not an instruction if (!(code.etype & (XPR_IMMEDIATE | XPR_SYM1))) return 0; // has no immediate value = value0 = code.value.i; // immediate constant floatType = uint8_t(code.dtype) >= uint8_t(TYP_FLOAT16); // floating point needed floatConst = (code.etype & XPR_FLT) != 0; // floating point provided if (floatType) { if (floatConst) dvalue = code.value.d; else { // Note: We are converting the immediate constant to floating point here in order to find // the optimal representation. We have not identified the instruction yet so we don't know // if it actually needs a floating point constant or an integer. We have saved the original // integer value in value0 so that we can undo the conversion in case an instruction with // floating point type needs an integer operand. dvalue = (double)value; // value as float if (code.etype & XPR_INT) { // convert integer constant to float code.value.d = dvalue; code.etype = (code.etype & ~XPR_IMMEDIATE) | XPR_FLT; floatConst = true; } } if ((code.etype & XPR_FLT) && uint8_t(code.dtype) == uint8_t(TYP_FLOAT32)) { union { // check for overflow in single precision float float f; uint32_t i; } u; u.f = float(code.value.d); if (isinf_f(u.i) && u.f > code.value.d) errors.reportLine(ERR_CONSTANT_TOO_LARGE); } if ((code.etype & XPR_FLT) && uint8_t(code.dtype) == uint8_t(TYP_FLOAT16)) { // check for overflow in half precision float if (isinf_h(double2half(code.value.d) && !isinf_d(code.value.i))) errors.reportLine(ERR_CONSTANT_TOO_LARGE); } } // check for symbols if (code.sym3) { sym3 = code.sym3; sym4 = code.sym4; symscale = code.symscale3; isym3 = findSymbol(sym3); if (isym3 < 1) { code.sizeUnknown = 2; return 2; // should not occur } } if (code.sym3 && !code.sym4 && int32_t(symbols[isym3].st_section) == SECTION_LOCAL_VAR && symbols[isym3].st_type == STT_CONSTANT) { // convert local symbol to constant value = symbols[isym3].st_value; code.value.i = value; code.sym3 = 0; if (cmd.optiLevel && value >= 0 && (code.dtype & TYP_INT) && (code.dest & REG_R)) { code.dtype |= TYP_PLUS; // optimize to larger type for positive constant because it is zero-extended anyway } } else if (sym3) { // there is a symbol if (symbols[isym3].st_unitsize == 0) uncertain = 2; // symbol value is not known yet uint32_t sym3section = symbols[isym3].st_section; // symbol section // determine necessary relocation size if relocation needed uint64_t relSize; // maximum size of relocated address if (symbols[isym3].st_type == STT_CONSTANT) { relSize = 0x10000000; // there is no command line option for the size of absolute symbols. assume 32 bit code.etype |= XPR_INT; } else if (sym3section && symbols[isym3].st_type != STT_CONSTANT) { // local symbol with known section relSize = (sectionHeaders[sym3section].sh_flags & (SHF_EXEC | SHF_IP)) ? code_size : data_size; } else { // external symbol with unknown section. look at symbol attributes relSize = (symbols[isym3].st_other & (STV_EXEC | STV_IP)) ? code_size : data_size; if (!(code.etype & (XPR_MEM | XPR_SYM2))) { errors.reportLine(ERR_CONFLICT_TYPE); // must be memory operand } } if (sym4) { // value is (sym3 - sym4) / scale factor isym4 = findSymbol(sym4); if (isym4 <= 0) { code.sizeUnknown = 2; return 2; // should not occur } code.etype |= XPR_INT; // symbol difference gives an integer if (symbols[isym3].st_unitsize == 0) uncertain = 2; // symbol value is not known yet if (symbols[isym3].st_section != symbols[isym4].st_section || symbols[isym3].st_bind != STB_LOCAL || symbols[isym4].st_bind != STB_LOCAL) { // different sections or not local. relocation needed fitNum = IFIT_RELOC; if (code.symscale1 > 1) relSize /= code.symscale1; // value is scaled if (relSize <= 1 << 7) fitNum |= IFIT_I8; if (relSize <= 1 << 15) fitNum |= IFIT_I16; if (relSize <= (uint64_t)1 << 31) fitNum |= IFIT_I32; code.fitNum = fitNum; code.sizeUnknown = uncertain; return uncertain; } // difference between two local symbols if (pass < 4) { code.fitNum = IFIT_I8 | IFIT_I16 | IFIT_I32; // symbol values are not available yet code.sizeUnknown = 1; return 1; } value += int32_t(uint32_t(symbols[isym3].st_value) - uint32_t(symbols[isym4].st_value)); if (symscale < 1) symscale = 1; valueScaled = value / symscale + code.offset_mem; if (valueScaled >= -(1 << 7) && valueScaled < (1 << 7)) fitNum |= IFIT_I8; if (valueScaled >= -(1 << 15) && valueScaled < (1 << 15)) fitNum |= IFIT_I16; if (valueScaled >= -((int64_t)1 << 31) && valueScaled < ((int64_t)1 << 31)) fitNum |= IFIT_I32; // check if value is certain. uncertainty is stored in high part of st_value uncertainty = (symbols[isym3].st_value >> 32) - (symbols[isym4].st_value >> 32); valueScaled = value / symscale + code.offset_mem + uncertainty; if (symscale > 1) valueScaled /= symscale; // value is scaled if ((valueScaled < -(1 << 7) || valueScaled >= (1 << 7)) && (fitNum & IFIT_I8)) uncertain |= 1; if ((valueScaled < -(1 << 15) || valueScaled >= (1 << 15)) && (fitNum & IFIT_I16)) uncertain |= 1; if ((valueScaled < -((int64_t)1 << 31) || valueScaled >= ((int64_t)1 << 31)) && (fitNum & IFIT_I32)) uncertain |= 1; if (uncertain && (code.fitNum & IFIT_LARGE)) { // choose the larger version if optimization process has convergence problems fitNum = (fitNum & (fitNum - 1)) | IFIT_I32; // remove the lowest set bit uncertain &= ~1; } code.fitNum = fitNum; code.sizeUnknown = uncertain; return uncertain; } // one symbol. must be constant if (sym3section != 0 && symbols[isym3].st_type != STT_CONSTANT && !(code.etype & XPR_MEM)) { errors.reportLine(ERR_MEM_WO_BRACKET); return 1; } if (sym3section && symbols[isym3].st_type != STT_CONSTANT && (sectionHeaders[sym3section].sh_flags & SHF_IP)) { // relative to instruction pointer if (sym3section != code.section || symbols[isym3].st_bind != STB_LOCAL) { // symbol is in different section or not local. relocation needed fitNum = IFIT_RELOC; if (relSize <= 1 << 7) fitNum |= IFIT_I8; // necessary relocation size if (relSize <= 1 << 15) fitNum |= IFIT_I16; if (relSize <= (uint64_t)1 << 31) fitNum |= IFIT_I32; code.fitNum = fitNum; code.sizeUnknown = uncertain; return uncertain; } if (pass < 4) { code.fitNum = IFIT_I8 | IFIT_I16 | IFIT_I32; // symbol values are not available yet code.sizeUnknown = 1; return 1; } // self-relative address to local symbol value = int32_t((uint32_t)symbols[isym3].st_value - (code.address + code.size * 4)); valueScaled = value + code.offset_mem; if (valueScaled >= -(1 << 7) && valueScaled < (1 << 7)) fitNum |= IFIT_I8; if (valueScaled >= -(1 << 15) && valueScaled < (1 << 15)) fitNum |= IFIT_I16; if (valueScaled >= -((int64_t)1 << 31) && valueScaled < ((int64_t)1 << 31)) fitNum |= IFIT_I32; code.fitNum = fitNum; // check if value is certain. uncertainty is stored in high part of st_value and sh_link uncertainty = int32_t((symbols[isym3].st_value >> 32) - sectionHeaders[code.section].sh_link); valueScaled += uncertainty; if ((valueScaled < -(1 << 7) || valueScaled >= (1 << 7)) && (fitNum & IFIT_I8)) uncertain |= 1; if ((valueScaled < -(1 << 15) || valueScaled >= (1 << 15)) && (fitNum & IFIT_I16)) uncertain |= 1; if ((valueScaled < -((int64_t)1 << 31) || valueScaled >= ((int64_t)1 << 31)) && (fitNum & IFIT_I32)) uncertain |= 1; if (uncertain && (code.fitNum & IFIT_LARGE)) { // choose the larger version if optimization process has convergence problems fitNum = (fitNum & (fitNum - 1)) | IFIT_I32; // remove the lowest set bit uncertain &= ~1; } code.fitNum = fitNum; code.sizeUnknown = uncertain; return uncertain; } // symbol is relative to data pointer or external constant. relocation needed fitNum = IFIT_RELOC; if (relSize <= 1 << 7) fitNum |= IFIT_I8; if (relSize <= 1 << 15) fitNum |= IFIT_I16; if (relSize <= (uint64_t)1 << 31) fitNum |= IFIT_I32; code.fitNum = fitNum; code.sizeUnknown = uncertain; return uncertain; } // no symbol. only a constant if (floatType) { // floating point constant code.fitNum = fitFloat(dvalue); if (uint8_t(code.dtype) < uint8_t(TYP_FLOAT64)) code.fitNum |= FFIT_32; code.sizeUnknown = 0; return 0; } // integer constant uint32_t low; // index of lowest set bit uint32_t high; // index of highest set bit fitNum = 0; int nbits; if (value == int64_t(0x8000000000000000)) { // prevent overflow of -value fitNum = 0; } else if (value >= 0) { low = bitScanForward((uint64_t)value); // lowest set bit high = bitScanReverse((uint64_t)value); // highest set bit //if (value < 8) fitNum |= IFIT_I4; //if (value == 8) fitNum |= IFIT_J4; //if (value < 0x10) fitNum |= IFIT_U4; if (value < 0x80) fitNum |= IFIT_I8 | IFIT_I8SHIFT; if (value == 0x80) fitNum |= IFIT_J8; if (value <= 0xFF) fitNum |= IFIT_U8; if (value < 0x8000) fitNum |= IFIT_I16 | IFIT_I16SH16; if (value == 0x8000) fitNum |= IFIT_J16; if (value <= 0xFFFF) fitNum |= IFIT_U16; if (high < 31) fitNum |= IFIT_I32; if (high < 32) fitNum |= IFIT_U32; if (value == 0x80000000U) fitNum |= IFIT_J32; nbits = high - low + 1; if (nbits < 8) fitNum |= IFIT_I8SHIFT; if (nbits < 16) { fitNum |= IFIT_I16SHIFT; if (low >= 16 && high < 31) fitNum |= IFIT_I16SH16; } if (nbits < 32) fitNum |= IFIT_I32SHIFT; if (low >= 32) fitNum |= IFIT_I32SH32; } else { // x < 0 value = -value; low = bitScanForward(value); // lowest set bit high = bitScanReverse(value); // highest set bit //if (value <= 8) fitNum |= IFIT_I4; if (value <= 0x80) fitNum |= IFIT_I8 | IFIT_I8SHIFT; if (value <= 0x8000) fitNum |= IFIT_I16 |IFIT_I16SH16 ; if (value <= 0x80000000U) fitNum |= IFIT_I32; nbits = high - low + 1; if (nbits < 8) fitNum |= IFIT_I8SHIFT; if (nbits < 16) { fitNum |= IFIT_I16SHIFT; if (low >= 16 && high <= 31) fitNum |= IFIT_I16SH16; } if (nbits < 32) fitNum |= IFIT_I32SHIFT; if (low >= 32) fitNum |= IFIT_I32SH32; } code.fitNum = fitNum; code.sizeUnknown = 0; return 0; } // Check how many bits are needed to a relative address or jump offset of an instruction. // This result is returned as bit-flags in codefitAddr, code.fitJump, and code.fitNum // The return value is nonzero if the size cannot be resolved yet. int CAssembler::fitAddress(SCode & code) { int64_t value = 0; // the constant or address to fit int64_t valueScaled; // value divided by scale factor uint32_t fitBits = 0; // bit flags indicating fit int32_t isym1 = 0, isym2 = 0; // symbol index int32_t uncertainty; // maximum deviance if the value is uncertain int uncertain = 0; // return value if (code.instruction == II_ALIGN) return 0; // not an instruction if (!(code.etype & (XPR_OFFSET | XPR_JUMPOS | XPR_MEM))) return 0; // has no address // check address of memory operand if (code.sym1) { // there is a memory operand symbol code.etype |= XPR_OFFSET; value = code.offset_mem; // memory offset isym1 = findSymbol(code.sym1); if (isym1 <= 0) { code.sizeUnknown = 2; return 2; // should not occur } if (symbols[isym1].st_unitsize == 0) uncertain = 2; // symbol value is not known yet uint32_t sym1section = symbols[isym1].st_section; // symbol section if (sym1section < sectionHeaders.numEntries()) { // determine necessary relocation size if relocation needed uint64_t relSize; // maximum size of relocated address if (symbols[isym1].st_type == STT_CONSTANT) { // assume that constant offset is limited by dataSizeOption relSize = data_size; // relocation size for code and constant data } else if (sym1section && !(sectionHeaders[sym1section].sh_flags & (SHF_WRITE | SHF_DATAP | SHF_THREADP))) { relSize = code_size; // relocation size for code and constant data } else if (sym1section) { // local symbol with known section relSize = (sectionHeaders[sym1section].sh_flags & (SHF_EXEC | SHF_IP)) ? code_size : data_size; } else { // external symbol with unknown section. look at symbol attributes relSize = (symbols[isym1].st_other & (STV_EXEC | STV_IP)) ? code_size : data_size; } if (code.sym2) { // value is (sym1 - sym2) / scale factor isym2 = findSymbol(code.sym2); if (isym2 <= 0) { code.sizeUnknown = 2; return 2; // should not occur } if (symbols[isym1].st_unitsize == 0) uncertain = 2; // symbol value is not known yet if (symbols[isym1].st_section != symbols[isym2].st_section || symbols[isym1].st_bind != STB_LOCAL || symbols[isym2].st_bind != STB_LOCAL) { // different sections or not local. relocation needed fitBits = IFIT_RELOC; if (code.symscale1 > 1) relSize /= code.symscale1; // value is scaled if (relSize <= 1 << 7) fitBits |= IFIT_I8; if (relSize <= 1 << 15) fitBits |= IFIT_I16; //if (relSize <= 1 << 23) fitBits |= IFIT_I24; if (relSize <= (uint64_t)1 << 31) fitBits |= IFIT_I32; code.fitAddr = fitBits; code.sizeUnknown += uncertain; //return uncertain; } // difference between two local symbols else if (pass < 4) { code.fitAddr = IFIT_I8 | IFIT_I16 | IFIT_I32; // symbol values are not available yet code.sizeUnknown += 1; uncertain += 1; //return 1; } else { value += int32_t(uint32_t(symbols[isym1].st_value) - uint32_t(symbols[isym2].st_value)); int scale = code.symscale1; if (scale < 1) scale = 1; valueScaled = value / scale + code.offset_mem; if (valueScaled >= -(1 << 7) && valueScaled < (1 << 7)) fitBits |= IFIT_I8; if (valueScaled >= -(1 << 15) && valueScaled < (1 << 15)) fitBits |= IFIT_I16; if (valueScaled >= -((int64_t)1 << 31) && valueScaled < ((int64_t)1 << 31)) fitBits |= IFIT_I32; // check if value is certain. uncertainty is stored in high part of st_value uncertainty = (symbols[isym1].st_value >> 32) - (symbols[isym2].st_value >> 32); valueScaled = value / scale + code.offset_mem + uncertainty; if (code.symscale1 > 1) valueScaled /= code.symscale1; // value is scaled if ((valueScaled < -(1 << 7) || valueScaled >= (1 << 7)) && (fitBits & IFIT_I8)) uncertain |= 1; if ((valueScaled < -(1 << 15) || valueScaled >= (1 << 15)) && (fitBits & IFIT_I16)) uncertain |= 1; if ((valueScaled < -((int64_t)1 << 31) || valueScaled >= ((int64_t)1 << 31)) && (fitBits & IFIT_I32)) uncertain |= 1; if (uncertain && (code.fitAddr & IFIT_LARGE)) { // choose the larger version if optimization process has convergence problems fitBits = (fitBits & (fitBits - 1)) | IFIT_I32; // remove the lowest set bit uncertain &= ~1; } code.fitAddr = fitBits; code.sizeUnknown += uncertain; //return uncertain; } } // one symbol else if (sectionHeaders[sym1section].sh_flags & SHF_IP) { // relative to instruction pointer if (sym1section != code.section || symbols[isym1].st_bind != STB_LOCAL) { // symbol is in different section or not local. relocation needed fitBits = IFIT_RELOC; if (code.etype & XPR_JUMPOS) relSize >>= 2; // value is scaled by 4 if (relSize <= 1 << 7) fitBits |= IFIT_I8; // necessary relocation size if (relSize <= 1 << 15) fitBits |= IFIT_I16; if (relSize <= 1 << 23) fitBits |= IFIT_I24; if (relSize <= (uint64_t)1 << 31) fitBits |= IFIT_I32; code.fitAddr = fitBits; code.sizeUnknown += uncertain; //return uncertain; } else if (pass < 4) { // code.fitBits = IFIT_I16 | IFIT_I32; // symbol values are not available yet code.fitAddr = IFIT_I16 | IFIT_I24 | IFIT_I32; // symbol values are not available yet code.sizeUnknown += 1; uncertain |= 1; //return 1; } else { // self-relative address to local symbol value = int32_t((uint32_t)symbols[isym1].st_value - (code.address + code.size * 4)); valueScaled = value; valueScaled += code.offset_mem; if (valueScaled >= -(1 << 15) && valueScaled < (1 << 15)) fitBits |= IFIT_I16; if (valueScaled >= -(1 << 23) && valueScaled < (1 << 23)) fitBits |= IFIT_I24; if (valueScaled >= -((int64_t)1 << 31) && valueScaled < ((int64_t)1 << 31)) fitBits |= IFIT_I32; code.fitAddr = fitBits; // check if value is certain. uncertainty is stored in high part of st_value and sh_link uncertainty = int32_t((symbols[isym1].st_value >> 32) - sectionHeaders[code.section].sh_link); valueScaled += uncertainty; if ((valueScaled < -(1 << 7) || valueScaled >= (1 << 7)) && (fitBits & IFIT_I8)) uncertain |= 1; if ((valueScaled < -(1 << 15) || valueScaled >= (1 << 15)) && (fitBits & IFIT_I16)) uncertain |= 1; if ((valueScaled < -(1 << 23) || valueScaled >= (1 << 23)) && (fitBits & IFIT_I24)) uncertain |= 1; if ((valueScaled < -((int64_t)1 << 31) || valueScaled >= ((int64_t)1 << 31)) && (fitBits & IFIT_I32)) uncertain |= 1; if (uncertain && (code.fitAddr & IFIT_LARGE)) { // choose the larger version if optimization process has convergence problems fitBits = (fitBits & (fitBits - 1)) | IFIT_I32; // remove the lowest set bit uncertain &= ~1; } code.fitAddr = fitBits; code.sizeUnknown += uncertain; //return uncertain; } } else { // symbol is relative to data pointer. relocation needed fitBits = IFIT_RELOC; if (relSize <= 1 << 7) fitBits |= IFIT_I8; if (relSize <= 1 << 15) fitBits |= IFIT_I16; if (relSize <= (uint64_t)1 << 31) fitBits |= IFIT_I32; code.fitAddr = fitBits; code.sizeUnknown += uncertain; } } } else { // no symbol. only a signed integer constant value = code.offset_mem; fitBits = 0; if (value >= -(int64_t)0x80 && value < 0x80) fitBits |= IFIT_I8; if (value >= -(int64_t)0x8000 && value < 0x8000) fitBits |= IFIT_I16; if (value >= -(int64_t)0x80000000 && value < 0x80000000) fitBits |= IFIT_I32; code.fitAddr = fitBits; } // check jump offset symbol if (code.sym5) { // there is a jump offset symbol value = code.offset_jump; // jump offset fitBits = 0; isym1 = findSymbol(code.sym5); if (isym1 <= 0) { code.sizeUnknown = 2; return 2; // should not occur } // one symbol relative to instruction pointer if (symbols[isym1].st_unitsize == 0) uncertain = 2; // symbol value is not known yet uint32_t sym1section = symbols[isym1].st_section; // symbol section if (sym1section < sectionHeaders.numEntries()) { // determine necessary relocation size if relocation needed uint64_t relSize; // maximum size of relocated address relSize = code_size >> 2; // relocation size for code and constant data, scaled by 4 if (sym1section != code.section || symbols[isym1].st_bind != STB_LOCAL) { // symbol is in different section or not local. relocation needed fitBits = IFIT_RELOC; if (relSize <= 1 << 7) fitBits |= IFIT_I8; // necessary relocation size if (relSize <= 1 << 15) fitBits |= IFIT_I16; if (relSize <= 1 << 23) fitBits |= IFIT_I24; if (relSize <= (uint64_t)1 << 31) fitBits |= IFIT_I32; code.fitJump = fitBits; code.sizeUnknown += uncertain; //return uncertain; } else if (pass < 4) { code.fitJump = IFIT_I16 | IFIT_I24 | IFIT_I32; // symbol values are not available yet code.sizeUnknown += 1; uncertain = 1; //return 1; } else { // self-relative address to local symbol value = int32_t((uint32_t)symbols[isym1].st_value - (code.address + code.size * 4)); valueScaled = value >> 2; // jump address is scaled valueScaled += code.offset_jump; if (valueScaled >= -(1 << 7) && valueScaled < (1 << 7)) fitBits |= IFIT_I8; if (valueScaled >= -(1 << 15) && valueScaled < (1 << 15)) fitBits |= IFIT_I16; if (valueScaled >= -(1 << 23) && valueScaled < (1 << 23)) fitBits |= IFIT_I24; if (valueScaled >= -((int64_t)1 << 31) && valueScaled < ((int64_t)1 << 31)) fitBits |= IFIT_I32; code.fitJump = fitBits; // check if value is certain. uncertainty is stored in high part of st_value and sh_link uncertainty = int32_t((symbols[isym1].st_value >> 32) - sectionHeaders[code.section].sh_link); valueScaled += uncertainty; if ((valueScaled < -(1 << 7) || valueScaled >= (1 << 7)) && (fitBits & IFIT_I8)) uncertain |= 1; if ((valueScaled < -(1 << 15) || valueScaled >= (1 << 15)) && (fitBits & IFIT_I16)) uncertain |= 1; if ((valueScaled < -(1 << 23) || valueScaled >= (1 << 23)) && (fitBits & IFIT_I24)) uncertain |= 1; if ((valueScaled < -((int64_t)1 << 31) || valueScaled >= ((int64_t)1 << 31)) && (fitBits & IFIT_I32)) uncertain |= 1; if (uncertain && (code.fitAddr & IFIT_LARGE)) { // choose the larger version if optimization process has convergence problems fitBits = (fitBits & (fitBits - 1)) | IFIT_I32; // remove the lowest set bit uncertain &= ~1; code.fitJump = fitBits; //code.sizeUnknown += uncertain; } code.sizeUnknown += uncertain; } } } return uncertain; } // find format details in formatList from entry in instructionlist uint32_t findFormat(SInstruction const & listentry, uint32_t imm) { // listentry: record in instructionlist or instructionlistId // imm: immediate operand, if any // make model instruction for lookupFormat STemplate instrModel; instrModel.a.il = listentry.format >> 8; instrModel.a.mode = (listentry.format >> 4) & 7; instrModel.a.ot = (listentry.format >> 5) & 4; if ((listentry.format & ~ 0x12F) == 0x200) { // format 0x200, 0x220, 0x300, 0x320 instrModel.a.mode2 = listentry.format & 7; } else if ((listentry.format & 0xFF0) == 0x270 && listentry.op1 < 8) { instrModel.a.mode2 = listentry.op1 & 7; } else instrModel.a.mode2 = 0; instrModel.a.op1 = listentry.op1; instrModel.b[0] = imm & 0xFF; // look op details for this format (from emulator2.cpp) return lookupFormat(instrModel.q); } // find the smallest representation that the floating point operand fits into int fitFloat(double x) { if (x == 0.) return IFIT_I8 | FFIT_16 | FFIT_32 | FFIT_64; union { double d; struct { uint64_t mantissa: 52; uint64_t exponent: 11; uint64_t sign: 1; } f; } u; u.d = x; int fit = FFIT_64; // check if mantissa fits if ((u.f.mantissa & (((uint64_t)1 << 42) - 1)) == 0) fit |= FFIT_16; if ((u.f.mantissa & (((uint64_t)1 << 29) - 1)) == 0) fit |= FFIT_32; // check if exponent fits, except for infinity or nan if (u.f.exponent != 0x7FF) { int ex = int(u.f.exponent - 0x3FF); if (ex < -14 || ex > 15) fit &= ~FFIT_16; if (ex < -126 || ex > 127) fit &= ~FFIT_32; } // check if x fits into a small integer if (fit & FFIT_16) { int i = int(x); if (i == x && i >= -128 && i < 128) { fit |= IFIT_I8; } } return fit; } // find an instruction variant that fits the code int CAssembler::fitCode(SCode & code) { // return value: // 0: does not fit // 1: fits uint32_t bestInstr = 0; // best fitting instruction variant, index into instructionlistId uint32_t bestSize = 99; // size of best fitting instruction variant SCode codeTemp; // fitted code SCode codeBest; // best fitted code uint32_t instrIndex = 0, ii; // index into instructionlistId uint32_t formatIx = 0; // index into formatList uint32_t isize; // il bits codeBest.category = 0; // find instruction by id SInstruction3 sinstr; // make dummy record with instruction id as parameter to findAll if (code.instruction == II_ALIGN) { return 1; // alignment directive } sinstr.id = code.instruction; int32_t nInstr = instructionlistId.findAll(&instrIndex, sinstr); if (code.etype & (XPR_IMMEDIATE | XPR_OFFSET | XPR_LIMIT | XPR_JUMPOS)) { // there is an immediate constant, offset, or limit. // generate specific error message if large constant cannot fit if ((code.etype & XPR_OFFSET) && !(code.etype & XPR_IMMEDIATE) && !(code.fitAddr & IFIT_I32)) { errors.reportLine(ERR_OFFSET_TOO_LARGE); } //else if ((code.etype & XPR_LIMIT) && !(code.fitBits & (IFIT_U16 | IFIT_U32))) errors.reportLine(ERR_LIMIT_TOO_LARGE); else if ((code.etype & XPR_IMMEDIATE) && !(code.etype & XPR_INT2)) { if (!(code.fitNum & (IFIT_I16 | IFIT_I16SHIFT | IFIT_I32 | IFIT_I32SHIFT | FFIT_16 | FFIT_32)) && (code.etype & XPR_OPTIONS) && code.optionbits) { errors.reportLine(ERR_IMMEDIATE_TOO_LARGE); } } } if (lineError) return 0; // loop through all instruction definitions with same id for (ii = instrIndex; ii < instrIndex + nInstr; ii++) { // category code.instr1 = ii; code.category = instructionlistId[ii].category; // get variant bits from instruction list variant = instructionlistId[ii].variant; // instruction-specific variants switch (instructionlistId[ii].category) { case 1: // single format. find entry in formatList formatIx = findFormat(instructionlistId[ii], code.value.w); code.formatp = formatList + formatIx; if (instructionFits(code, codeTemp, ii)) { // check if smaller than previously found. isize = codeTemp.size; if (isize < bestSize) { bestSize = isize; bestInstr = ii; codeBest = codeTemp; } } break; case 3: // multi-format instructions. search all formats for the best one for (formatIx = 0; formatIx < formatList3.numEntries(); formatIx++) { code.formatp = &formatList3[formatIx]; if (((uint64_t)1 << code.formatp->formatIndex) & instructionlistId[ii].format) { if (instructionFits(code, codeTemp, ii)) { // check if smaller than previously found. category 3 = multiformat preferred isize = codeTemp.size; if (isize < bestSize || (isize == bestSize && codeBest.category != 3)) { bestSize = isize; bestInstr = ii; codeBest = codeTemp; } } } } break; case 4: // jump instructions. search all formats for the best one for (formatIx = 0; formatIx < formatList4.numEntries(); formatIx++) { code.formatp = &formatList4[formatIx]; if (((uint64_t)1 << code.formatp->formatIndex) & instructionlistId[ii].format) { if (jumpInstructionFits(code, codeTemp, ii)) { // check if smaller than previously found. category 3 = multiformat preferred isize = codeTemp.size; if (isize < bestSize) { bestSize = isize; bestInstr = ii; codeBest = codeTemp; } } } } break; default: return 0; // error in list } } if (bestSize > 4) { errors.reportLine(checkCodeE(code)); // find reason why no format fits, and report error return 0; } code = codeBest; // get the best fitting code variant = instructionlistId[bestInstr].variant; // instruction-specific variants checkCode2(code); // check if operands are correct if (lineError) return 0; return 1; } // check if instruction fits into specified format bool CAssembler::instructionFits(SCode const & code, SCode & codeTemp, uint32_t ii) { // code: structure defining all operands and options // codeTemp: fitted code // ii: index into instructionlistId // formatIndex: index into formatList uint32_t shiftCount; // shift count for shifted constant // copy code structure and add details codeTemp = code; codeTemp.category = code.formatp->category; codeTemp.size = (code.formatp->format2 >> 8) & 3; if (codeTemp.size == 0) codeTemp.size = 1; codeTemp.instr1 = ii; if (instructionlistId[ii].opimmediate == OPI_IMPLICIT && !(code.etype & XPR_IMMEDIATE)) { // There is no immediate operand. instructionlistId[ii] has an implicit immediate operand. // Insert implicit operand and see if it fits codeTemp.value.u = instructionlistId[ii].implicit_imm; codeTemp.etype |= XPR_INT; codeTemp.fitNum = 0xFFFFFFFF; } // check vector use bool useVectors = (code.dtype & TYP_FLOAT) || (code.dest & 0xE0) == REG_V || (code.reg1 & 0xE0) == REG_V || (code.reg2 & 0xE0) == REG_V; if (useVectors) { if (!(code.formatp->vect)) return false; // vectors not supported } else if (code.formatp->vect & ~0x10) return false; // vectors provided but not used // requested operand type uint32_t requestOT = code.dtype & 7; if (uint8_t(code.dtype) == uint8_t(TYP_FLOAT16)) { requestOT = TYP_INT16 & 7; // replace pseudo-type TYP_FLOAT16 with TYP_INT16 codeTemp.dtype = TYP_INT16; } // operand type provided by this format uint32_t formatOT = code.formatp->ot; if (formatOT == 0x32) formatOT = 0x12 + (instructionlistId[ii].op1 & 1); // int32 for even op1, int64 for odd op1 if (formatOT == 0x35) formatOT = 0x15 + (instructionlistId[ii].op1 & 1); // float for even op1, double for odd op1 if (formatOT == 0) formatOT = requestOT; // operand type determined by OT field formatOT &= 7; uint32_t scale2 = formatOT; if (scale2 > 4) scale2 -= 3; // operand size = 1 << scale2 if (variant & (VARIANT_D0 | VARIANT_D2)) { // no operand type if (code.dtype == 0 && code.instruction != II_NOP) codeTemp.dtype = formatOT ? formatOT : 3; } else { // check requested operand type if (formatOT <= 3 && requestOT < formatOT && (code.dtype & TYP_PLUS)) { requestOT = formatOT; // request allows bigger type // codeTemp.dtype = formatOT; // prevents merging with subsequent jump with smaller type than formatOT } if (requestOT != formatOT && code.dtype) return false; // requested format type not supported // check if operand type supported by instruction uint32_t optypessupport = useVectors ? (instructionlistId[ii].optypesscalar | instructionlistId[ii].optypesvector) : instructionlistId[ii].optypesgp; optypessupport |= optypessupport >> 8; // include types with optional support if (!(optypessupport & (1 << requestOT))) return false; } // check if there are enough register operands in this format uint8_t opAvail = code.formatp->opAvail; uint8_t numReg = ((opAvail >> 4) & 1) + ((opAvail >> 5) & 1) + ((opAvail >> 6) & 1) + ((opAvail >> 7) & 1); // number of registers available uint8_t numReq = instructionlistId[ii].sourceoperands; // number of registers required for this instruction codeTemp.numOp = numReq; if ((codeTemp.etype & XPR_IMMEDIATE) && numReq) numReq--; if ((codeTemp.etype & XPR_MEM) && numReq) numReq--; if ((codeTemp.etype & (XPR_MASK | XPR_FALLBACK)) && ((code.fallback & 0x1F) != (code.reg1 & 0x1F) || (code.reg1 & 0x1F) == 0x1F)) { numReq += 2; // fallback different from reg1, implies reg1 != destination } else if ((code.etype & XPR_REG1) && code.dest && code.reg1 != code.dest && !(variant & VARIANT_D3)) { numReq++; // reg1 != destination } if (numReq > numReg) return false; // not enough registers in this format // check if mask available if ((code.etype & XPR_MASK) && !(code.formatp->tmplate == 0xA || code.formatp->tmplate == 0xE)) return false; // check option bits if ((code.etype & XPR_OPTIONS) && code.optionbits != 0 && (code.formatp->tmplate != 0xE || !(code.formatp->imm2 & 2)) && (variant & VARIANT_On) && instructionlistId[ii].opimmediate != OPI_INT1688) return false; // only template E has option bits // check memory operand if (code.etype & XPR_MEM) { if (code.formatp->mem == 0) return false; // memory operand requested but not supported if (code.etype & XPR_SYM1) { // has data symbol if (code.etype & XPR_SYM2) { // has difference between two symbols codeTemp.sizeUnknown = 1; } //if (!(code.fitNumX & IFIT_I32)) return false; // assume symbol address requires 32 bits. local symbol difference resolved later when sizeUnknown = 1 } // check index and scale factor if (code.etype & XPR_INDEX) { if (!(code.formatp->mem & 4)) return false; // index not supported if ((code.formatp->scale & 4) && code.scale != -1) return false; // scale factor must be -1 if ((code.formatp->scale & 2) && code.scale != 1 << scale2) return false; // scale factor must match operand type if (!(code.formatp->scale & 6) && code.scale != 1) return false; // scale factor must be 1 } else { // no index requested if (code.formatp->mem & 4) { codeTemp.index = 0x1F; // RT = 0x1F means no index codeTemp.scale = 1 << scale2; } } // check address offset size if (code.etype & (XPR_OFFSET | XPR_SYM1)) { if (!(code.formatp->mem & 0x10)) return false; // format does not support memory offset switch (code.formatp->addrSize) { case 1: if (code.sym1 && !(code.fitAddr & IFIT_I8)) return false; if ((code.base & 0x1F) >= 0x1C && (code.base & 0x1F) != 0x1F) return false; // ip, datap, threadp must have 16 bit offset // no relocation. scale factor depends on operand size if (code.offset_mem & ((1 << scale2) - 1)) return false; // offset is not a multiple of the scale factor if ((code.offset_mem >> scale2) < -0x80 || (code.offset_mem >> scale2) > 0x7F) return false; break; case 2: if (!(code.fitAddr & IFIT_I16)) return false; break; case 4: if (!(code.fitAddr & IFIT_I32)) return false; break; default: return false; } } else if ((code.formatp->addrSize) < 2 && (code.base & 0x1F) >= 0x1C && (code.base & 0x1F) != 0x1F) return false; // fail if limit required and not supported, or supported and not required if (code.etype & XPR_LIMIT) { if (!(code.formatp->mem & 0x20)) return false; // limit not supported by format switch (code.formatp->addrSize) { case 1: if (code.value.u >= 0x100) return false; break; case 2: if (code.value.u >= 0x10000) return false; break; case 4: if (uint64_t(code.value.u) >= 0x100000000U) return false; break; } } else { if (code.formatp->mem & 0x20) return false; // limit provided but not requested } // check length/broadcast/scalar if (code.etype & XPR_SCALAR) { // scalar operand requested if ((code.formatp->vect & 6) != 0) { codeTemp.length = 31; // disable length or broadcast option } } else if (code.etype & XPR_LENGTH) { // vector length specified if ((code.formatp->vect & 2) == 0) return false; // vector length not in this format } else if (code.etype & XPR_BROADC) { // vector broadcast specified if ((code.formatp->vect & 4) == 0) return false; // vector broadcasst not in this format } } else if (code.formatp->mem) return false; // memory operand supported by not requested // check immediate operand //bool isFloat = (code.dtype & TYP_FLOAT32 & 0xF0) != 0; // specified type is float or double or float128 bool hasImmediate = (code.etype & XPR_IMMEDIATE) != 0; // && !(code.etype & (XPR_OFFSET | XPR_LIMIT))); /*if ((variant & VARIANT_M1) && code.formatp->mem && code.formatp->tmplate == 0xE) { // variant M1: immediate operand is in IM3. No further check needed // to do: fail if relocation on immediate return hasImmediate; // succeed if there is an immediate } */ if (hasImmediate) { if (code.formatp->immSize == 0 && instructionlistId[ii].sourceoperands < 4) return false; // immediate not supported // to do: check if relocation // check if size fits. special cases in instruction list switch (instructionlistId[ii].opimmediate) { case OPI_IMPLICIT: // implicit value of immediate operand. Accept explicit value only if same if (codeTemp.value.u != instructionlistId[ii].implicit_imm) return false; break; case OPI_INT8SH: // im2 << im1 if (code.fitNum & (IFIT_I8 | IFIT_I8SHIFT)) { // fits im2 << im1 shiftCount = bitScanForward(codeTemp.value.u); codeTemp.value.u = (codeTemp.value.u >> shiftCount << 8) | shiftCount; codeTemp.fitNum |= IFIT_I16; // make it accepted below break; } return false; case OPI_INT16SH16: // im12 << 16 if (code.fitNum & (IFIT_I16 | IFIT_I16SH16)) { // fits im2 << 16 codeTemp.value.u = codeTemp.value.u >> 16; codeTemp.fitNum |= IFIT_I16; // make it accepted below break; } return false; case OPI_INT32SH32: // im2 << 32 if (code.fitNum & (IFIT_I32 | IFIT_I32SH32)) { // fits im2 << 32 codeTemp.value.u = codeTemp.value.u >> 32; codeTemp.fitNum |= IFIT_I32; // make it accepted below break; } return false; case OPI_UINT8: // 8 bit unsigned integer if (value0 < 0x100 && value0 > -(int64_t)0x80U) return true; return false; case OPI_UINT16: // 16 bit unsigned integer if (value0 < 0x10000 && value0 > -(int64_t)0x8000U) return true; return false; case OPI_UINT32: // 32 bit unsigned integer //if (code.fitNum & IFIT_U32) return true; // this does not work if a float type is specified if (value0 < 0x100000000 && value0 > -(int64_t)0x80000000U) return true; return false; case OPI_INT886: // three integers codeTemp.value.u = (codeTemp.value.w & 0xFF) | (codeTemp.value.u >> 24); return true; case OPI_INT1688: // three integers: 16 + 8 + 8 bits codeTemp.value.u = (codeTemp.value.w & 0xFFFF) | (codeTemp.value.u >> 16 & 0xFF0000) | codeTemp.optionbits << 24; return true; case OPI_OT: // constant of same type as operand type if ((uint8_t(code.dtype) & ~TYP_UNS) <= uint8_t(TYP_INT32) && code.formatp->immSize >= 4) return true; } // check if size fits. general cases switch (code.formatp->immSize) { case 1: if (codeTemp.fitNum & IFIT_I8) break; // fits if ((variant & VARIANT_U0) && (codeTemp.fitNum & IFIT_U8)) break; // unsigned fits if ((codeTemp.dtype & 0x1F) == (TYP_INT8 & 0x1F) && (codeTemp.fitNum & IFIT_U8)) break; // 8 bit size fits unsigned with no sign extension return false; case 2: if (codeTemp.fitNum & (IFIT_I16 | FFIT_16)) break; // fits if ((variant & VARIANT_U0) && (codeTemp.fitNum & IFIT_U16)) break; // unsigned fits if ((codeTemp.dtype & 0x1F) == (TYP_INT16 & 0x1F) && code.formatp->tmplate != 0xC && (codeTemp.fitNum & IFIT_U16)) break; // 16 bit size fits unsigned with no sign extension if ((code.formatp->imm2 & 4) && !(variant & VARIANT_On) && (codeTemp.fitNum & IFIT_I16SHIFT)) { // fits with im2 << im3 shiftCount = bitScanForward(codeTemp.value.u); codeTemp.value.u >>= shiftCount; codeTemp.optionbits = shiftCount; break; } if (variant & VARIANT_H0) break; // half precision fits return false; case 4: if ((code.dtype & 0xFF) == (TYP_FLOAT32 & 0xFF)) break; // float32 must be rounded to fit if (codeTemp.fitNum & (IFIT_I32 | FFIT_32)) break; // fits if ((codeTemp.fitNum & IFIT_U32) && (code.dtype & 0xFF) == (TYP_INT32 & 0xFF)) break; // fits if ((variant & VARIANT_U0) && (codeTemp.fitNum & IFIT_U32)) break; // unsigned fits if (variant & VARIANT_H0) break; // half precision fits if ((codeTemp.dtype & 0x1F) == (TYP_INT32 & 0x1F) && (codeTemp.fitNum & IFIT_U32)) break; // 32 bit size fits unsigned with no sign extension if ((code.formatp->imm2 & 8) && (codeTemp.fitNum & IFIT_I32SHIFT)) { // fits with im4 << im2 shiftCount = bitScanForward(codeTemp.value.u); codeTemp.value.u = ((codeTemp.value.u >> shiftCount) & 0xFFFFFFFF) | ((uint64_t)shiftCount << 32); // store shift count in upper half break; } return false; case 8: break; default:; // other values should not occur in table } } else if ((code.formatp->immSize != 0) && !(code.etype & (XPR_OFFSET | XPR_LIMIT)) && instructionlistId[ii].sourceoperands && code.category != 1) { return false; // immediate operand provided but not required } return true; } // check if instruction fits into specified format bool CAssembler::jumpInstructionFits(SCode const & code, SCode & codeTemp, uint32_t ii) { // code: structure defining all operands and options // codeTemp: fitted code // ii: index into instructionlistId // formatIndex: index into formatList4 //uint8_t offsetSize = 0; // number of bytes to use in relative address //uint8_t immediateSize = 0; // number of bytes to use in immediate operand bool offsetRelocated = false; // relative offset needs relocation //bool immediateRelocated = false; // immediate operand needs relocation codeTemp = code; codeTemp.category = code.formatp->category; codeTemp.size = (code.formatp->format2 >> 8) & 3; codeTemp.instr1 = ii; // check vector use bool useVectors = (code.dtype & TYP_FLOAT) || (code.dest & 0xE0) == REG_V || (code.reg1 & 0xE0) == REG_V; if (useVectors) { if (!(code.formatp->vect)) return false; // vectors not supported } // operand type provided by this format uint32_t formatOT = code.formatp->ot; if (formatOT == 0) formatOT = code.dtype; // operand type determined by OT field formatOT &= 7; // check requested operand type uint32_t requestOT = code.dtype & 7; if (formatOT <= 3 && requestOT < formatOT && (code.dtype & TYP_PLUS)) { requestOT = formatOT; // request allows bigger type codeTemp.dtype = formatOT; } if (requestOT != formatOT && code.dtype) return false; // requested format type not supported // check if operand type supported by instruction uint32_t optypessupport = useVectors ? (instructionlistId[ii].optypesscalar | instructionlistId[ii].optypesvector) : instructionlistId[ii].optypesgp; optypessupport |= optypessupport >> 8; // include types with optional support if (!(optypessupport & (1 << requestOT))) return false; // check if there are enough register operands in this format uint8_t opAvail = code.formatp->opAvail; uint8_t numReg = ((opAvail >> 4) & 1) + ((opAvail >> 5) & 1) + ((opAvail >> 7) & 1); // number of registers available uint8_t numReq = instructionlistId[ii].sourceoperands; // number of registers required for this instruction if ((code.etype & XPR_REG1) && code.dest && code.reg1 != code.dest && numReq > 2) { numReq++; // reg1 != destination, except if no reg2 } if (code.formatp->jumpSize) numReq--; if ((code.etype & (XPR_IMMEDIATE | XPR_MEM)) && numReq) numReq--; if ((code.etype & XPR_INT2) && numReq) numReq--; if (numReq > numReg) return false; // not enough registers in this format // check if correct number of registers specified uint8_t nReg = 0; for (int j = 0; j < 3; j++) nReg += (code.etype & (XPR_REG1 << j)) != 0; if (code.dest && code.dest != code.reg1) nReg++; if (nReg != numReq) return false; // check if mask available if ((code.etype & XPR_MASK) && !(fInstr->tmplate == 0xA || fInstr->tmplate == 0xE)) return false; // self-relative jump offset if (code.etype & XPR_JUMPOS) { if (!(code.formatp->jumpSize)) return false; switch (code.formatp->jumpSize) { case 0: // no offset if (code.offset_jump || offsetRelocated) return false; break; case 1: // 1 byte if (!(code.fitJump & IFIT_I8)) return false; break; case 2: // 2 bytes if (!(code.fitJump & IFIT_I16)) return false; break; case 3: // 3 bytes if (!(code.fitJump & IFIT_I24)) return false; break; case 4: // 4 bytes if (!(code.fitJump & IFIT_I32)) return false; break; } } else { // no self-relative jump offset if (code.formatp->jumpSize) return false; } if (instructionlistId[ii].opimmediate == OPI_IMPLICIT && !(code.etype & XPR_IMMEDIATE)) { // There is no immediate operand. instructionlistId[ii] has an implicit immediate operand. // Insert implicit operand and see if it fits codeTemp.value.u = instructionlistId[ii].implicit_imm; codeTemp.etype |= XPR_INT; codeTemp.fitNum = 0xFFFFFFFF; } // immediate operand if (codeTemp.etype & XPR_IMMEDIATE) { if (code.dtype & TYP_FLOAT) { if (variant & VARIANT_I2) { // immediate should be integer codeTemp.etype = (code.etype & ~XPR_FLT) | XPR_INT; codeTemp.value.i = (int64_t)code.value.d; switch (code.formatp->immSize) { case 0: // no immediate return false; case 1: // 1 byte if (codeTemp.value.i < -0x80 || codeTemp.value.i > 0x7F) return false; break; case 2: // 2 bytes if (codeTemp.value.i < -0x8000 || codeTemp.value.i > 0x7FFF) return false; break; case 4: // 4 bytes if (-codeTemp.value.i > 0x80000000u || codeTemp.value.i > 0x7FFFFFFF) return false; break; } } else { // immediate is floating point or small integer converted to floating point int fit = code.fitNum; if ((code.dtype & 0xFF) <= (TYP_FLOAT32 & 0xFF)) fit |= FFIT_32; switch (code.formatp->immSize) { case 0: // no immediate return false; case 1: // 1 byte if (!(fit & IFIT_I8)) return false; break; case 2: // 2 bytes if (!(fit & FFIT_16)) return false; break; case 4: // 4 bytes if (!(fit & FFIT_32)) return false; break; case 8: // 8 bytes., currently not supported ; } } } else { // immediate integer operand switch (code.formatp->immSize) { case 0: // no immediate return false; case 1: if (codeTemp.fitNum & IFIT_I8) break; // fits if ((codeTemp.dtype & 0x1F) == (TYP_INT8 & 0x1F) && (codeTemp.fitNum & IFIT_U8)) break; // 8 bit size fits unsigned with no sign extension return false; case 2: // 2 bytes if (instructionlistId[ii].opimmediate == OPI_INT1632) { // 16+32 bits if ((codeTemp.value.u >> 32) <= 0xFFFF) break; return false; } if (codeTemp.fitNum & IFIT_I16) break; // fits if ((codeTemp.dtype & 0x1F) == (TYP_INT16 & 0x1F) && (codeTemp.fitNum & IFIT_U16)) break; // 16 bit size fits unsigned with no sign extension return false; case 4: // 4 bytes if (instructionlistId[ii].opimmediate == OPI_2INT16) { // 16+16 bits if (codeTemp.value.w <= 0xFFFF && (codeTemp.value.u >> 32) <= 0xFFFF) break; return false; } if (codeTemp.fitNum & IFIT_I32) break; // fits if ((codeTemp.dtype & 0x1F) == (TYP_INT32 & 0x1F) && (codeTemp.fitNum & IFIT_U32)) break; // 32 bit size fits unsigned with no sign extension return false; case 8: // 8 bytes break; default: // does not fit other sizes return false; } } } else { // no explicit immediate if (code.formatp->immSize && code.instruction != II_JUMP && code.instruction != II_CALL) return false; } // memory operand if (code.etype & XPR_MEM) { if (code.formatp->mem == 0) return false; // memory operand requested but not supported uint32_t scale2 = formatOT; if (scale2 > 4) scale2 -= 3; // operand size = 1 << scale2 if (code.etype & XPR_SYM1) { // has data symbol if (code.etype & XPR_SYM2) { // has difference between two symbols codeTemp.sizeUnknown = 1; } if (!(code.fitAddr & IFIT_I32)) return false; // assume symbol address requires 32 bits. local symbol difference resolved later when sizeUnknown = 1 } // check index and scale factor if (code.etype & XPR_INDEX) { if (!(code.formatp->mem & 4)) return false; // index not supported } else { // no index requested if (code.formatp->mem & 4) { codeTemp.index = 0x1F; // RT = 0x1F means no index codeTemp.scale = 1 << scale2; } } // check address offset size if (code.etype & XPR_OFFSET) { if (!(code.formatp->mem & 0x10)) return false; // format does not support memory offset switch (code.formatp->addrSize) { case 1: // scale factor depends on operand size if (code.offset_mem & ((1 << scale2) - 1)) return false; // offset is not a multiple of the scale factor if ((code.offset_mem >> scale2) < -0x80 || (code.offset_mem >> scale2) > 0x7F) return false; break; case 2: if (!(code.fitAddr & IFIT_I16)) return false; break; case 4: if (!(code.fitAddr & IFIT_I32)) return false; break; default: return false; } } } else if (code.formatp->mem) return false; // memory operand supported by not requested return true; } // Check code for correctness before fitting a format, and fix some code details void CAssembler::checkCode1(SCode & code) { // check code for correctness if (code.etype & XPR_MEM) { // check memory operand bool useVectors = (code.dtype & TYP_FLOAT) != 0 || (code.dest & 0xE0) == REG_V || (code.reg1 & 0xE0) == REG_V; if (useVectors && code.scale == -1) { code.etype |= XPR_LENGTH; code.length = code.index; // index register is also length } int numOpt = ((code.etype & XPR_SCALAR) != 0) + ((code.etype & XPR_LENGTH) != 0) + ((code.etype & XPR_BROADC) != 0); if (numOpt > 1) {errors.reportLine(ERR_CONFLICT_OPTIONS); return;} // conflicting options if (numOpt && !useVectors && !(code.etype & XPR_SCALAR)) {errors.reportLine(ERR_VECTOR_OPTION); return;} // vector option on non-vector operands if (code.etype & XPR_INDEX) { // check scale factor const int dataSizeTable[8] = {1, 2, 4, 8, 16, 4, 8, 16}; // data size for each operant type int8_t scale = code.scale; if (scale != 1 && scale != -1 && scale != dataSizeTable[code.dtype & 7]) errors.reportLine(ERR_SCALE_FACTOR); if (code.scale == -1 && code.length && code.length != code.index) { errors.reportLine(ERR_NEG_INDEX_LENGTH); return; } } if (!(code.etype & XPR_BASE)) { // no base pointer. check if there is a symbol with an implicit base pointer int32_t symi1 = 0; if (code.etype & XPR_SYM1) symi1 = findSymbol(code.sym1); if ((code.etype & XPR_SYM2) || symi1 < 1 || !(symbols[symi1].st_other & STV_SECT_ATTR)) { errors.reportLine(ERR_NO_BASE); } } } // check mask if ((code.etype & XPR_MASK) && (code.mask & 0x1F) > 6) errors.reportLine(ERR_MASK_REGISTER); // check fallback if (code.etype & XPR_MASK) { if (code.fallback == 0) code.fallback = code.reg1 ? code.reg1 : 0x1F; // default fallback is reg1, or 0 if no reg1 if ((code.fallback & 0xE0) == 0) code.fallback |= code.dest & 0xE0; // get type of dest if fallback has no type } // details for unsigned variants if (code.dtype & TYP_UNS) { // an unsigned type is specified switch (code.instruction) { case II_DIV: case II_DIV_EX: case II_MUL_HI: case II_MUL_EX: case II_REM: case II_SHIFT_RIGHT_S: case II_MIN: case II_MAX: code.instruction |= 1; // change to unsigned version break; default:; // other instructions: do nothing } } // handle half precision if (uint8_t(code.dtype) == uint8_t(TYP_FLOAT16)) { switch (code.instruction) { case II_ADD: case II_MUL: case II_DIV: case II_MUL_ADD: code.instruction |= II_ADD_H & 0xFF000; // change to half precision instruction break; case II_SUB: if ((code.etype & XPR_IMMEDIATE) && !(code.etype & (XPR_MEM | XPR_REG2))) { code.instruction = II_ADD_H; code.value.d = - code.value.d; // subtract constant changed to add -constant } else code.instruction = II_SUB_H; break; case II_SUB_REV: if (code.value.i == 0) { // -x code.instruction = II_TOGGLE_BIT; code.value.u = 15; } else errors.reportLine(ERR_WRONG_OPERANDS); break; case II_MOVE: case II_REPLACE: case II_REPLACE_EVEN: case II_REPLACE_ODD: if (code.etype & XPR_INT) { // convert integer to float16 if (abs(code.value.i) > 65504) errors.reportLine(ERR_OVERFLOW); code.value.u = double2half(double(code.value.i)); } else if (code.etype & XPR_FLT) { // convert double to float16 if (code.value.d > 65504. || code.value.d < -65504.) errors.reportLine(ERR_OVERFLOW); code.value.u = double2half(code.value.d); code.etype = (code.etype & ~ XPR_IMMEDIATE) | XPR_INT; } if (code.instruction == II_SUB_H && (code.etype & XPR_IMMEDIATE)) { code.value.w ^= 0x8000; code.instruction &= ~1; // convert sub_h constant to add_h -constant } code.dtype = TYP_INT16; code.fitNum = IFIT_I16 | IFIT_I32; break; case II_STORE: if (code.etype & XPR_INT) code.value.u = double2half(double(code.value.i)); else code.value.u = double2half(code.value.d); code.dtype = TYP_INT16; code.etype = (code.etype & ~ XPR_FLT) | XPR_INT; break; case II_ADD_H: case II_SUB_H: case II_MUL_H: case II_DIV_H: case II_MUL_ADD_H: break; default: // no other instructions support half precision errors.reportLine(ERR_WRONG_OPERANDS); } } // special case instructions switch (code.instruction) { case II_STORE: if ((code.dtype & TYP_FLOAT) && (code.etype & XPR_FLT) && !(code.reg1)) { // store float constant // code.dtype = code.dtype + (TYP_INT32 - TYP_FLOAT32) | TYP_UNS; } } // check size needed for immediate operand and address fitConstant(code); fitAddress(code); if (code.instruction & II_JUMP_INSTR) { // jump instruction code.category = 4; // check register type if (code.dtype && code.reg1) { if ((code.dtype & 0xFF) <= (TYP_FLOAT16 & 0xFF)) { // must use g.p. registers if (code.reg1 & REG_V) errors.reportLine(ERR_WRONG_REG_TYPE); } else { // must use vector registers if (code.reg1 & REG_R) errors.reportLine(ERR_WRONG_REG_TYPE); } } // check if immediate operand too big if (code.etype & XPR_IMMEDIATE) { if (code.dtype & TYP_FLOAT) { if ((code.dtype & 0xFF) >= (TYP_FLOAT64 & 0xFF) && !(code.fitNum & FFIT_32)) errors.reportLine(ERR_TOO_LARGE_FOR_JUMP); } else if (code.dtype & TYP_UNS) { if ((code.dtype & 0x1F) >= (TYP_INT64 & 0x1F) && !(code.fitNum & IFIT_U32)) errors.reportLine(ERR_TOO_LARGE_FOR_JUMP); } else if ((code.dtype & 0x1F) >= (TYP_INT64 & 0x1F) && !(code.fitNum & IFIT_I32)) errors.reportLine(ERR_TOO_LARGE_FOR_JUMP); } } // optimize instruction if (cmd.optiLevel) optimizeCode(code); } // Check register types etc. after fitting a format, and finish code details void CAssembler::checkCode2(SCode & code) { if (code.instruction >= II_ALIGN) return; // not an instruction // check type if (code.dtype == 0) { if ((code.etype & (XPR_INT | XPR_FLT | XPR_REG | XPR_REG1 | XPR_MEM)) && !(variant & (VARIANT_D0 | VARIANT_D2))) { // type not specified if (code.instruction == II_MOVE && code.category == 3 && !(code.etype & (XPR_IMMEDIATE | XPR_MEM))) { // register-to-register move. find appropriate operand type code.dtype = TYP_INT64; // g.p. register. copy whole register ?? if (code.dest & REG_V) code.dtype = TYP_INT8; // vector register. length must be divisible by tpe } else { errors.reportLine(ERR_TYPE_MISSING); // type must be specified return; } } } if (code.etype & XPR_MEM) { // check memory operand if (variant & VARIANT_M0) { // memory destination if (code.etype & XPR_BROADC) { errors.reportLine(ERR_DEST_BROADCAST); return; } } if (code.base >= REG_R + 28 && code.base <= REG_R + 30 && (code.formatp->addrSize) > 1 && pass < 4) { // cannot use r28 - r30 as base pointer with more than 8 bits offset // (we don't get an error message here for a symbol address because the base pointer has not been assigned yet) errors.reportLine(ERR_R28_30_BASE); } // check M1 option /*if (variant & VARIANT_M1) { if (code.formatp->tmplate == 0xE && (code.etype & XPR_MEM) && (code.etype & XPR_INT) && (code.value.i > 63 || code.value.i < -63)) { errors.reportLine(ERR_CONSTANT_TOO_LARGE); return; } if (code.optionbits && (code.etype & XPR_MEM)) { errors.reportLine(ERR_BOTH_MEM_AND_OPTIONS); return; } }*/ } if (lineError) return; // skip additional errors // Make list of operands from available operands. 0=none, 1=immediate, 2=memory, 5=RT, 6=RS, 7=RU, 8=RD uint8_t opAvail = code.formatp->opAvail; // Bit index of available operands int j; // loop counter // check if correct number of registers uint32_t numReq = instructionlistId[code.instr1].sourceoperands; // number of registers required for this instruction if (code.category == 4 && (code.instruction & II_JUMP_INSTR) && (code.etype & XPR_JUMPOS) && numReq) numReq--; if ((code.etype & XPR_IMMEDIATE) && numReq) numReq--; if ((code.etype & XPR_INT2) && numReq) numReq--; if ((code.etype & XPR_MEM) && !(variant & VARIANT_M0) && numReq) numReq--; uint32_t nReg = 0; for (j = 0; j < 3; j++) nReg += (code.etype & (XPR_REG1 << j)) != 0; if (nReg < numReq && !(variant & VARIANT_D3)) errors.reportLine(ERR_TOO_FEW_OPERANDS); else if (nReg > numReq && instructionlistId[code.instr1].opimmediate != 25) { errors.reportLine(ERR_TOO_MANY_OPERANDS); } // count number of available registers in format uint32_t regAvail = 0; opAvail >>= 4; // register operands while (opAvail) { regAvail += opAvail & 1; opAvail >>= 1; } // expected register types uint8_t regType = REG_R; if ((code.formatp->vect & 1) || ((code.formatp->vect & 0x10) && (code.dtype & 4))) regType = REG_V; // check each of up to three source registers for (j = 0; j < 3; j++) { if (code.etype & (XPR_REG1 << j)) { // register j used if (variant & VARIANT_SPECS) { // must be special register if (((&code.reg1)[j] & 0xE0) <= REG_V) errors.reportLine(ERR_WRONG_REG_TYPE); } else if ((variant & (VARIANT_R1 << j)) || ((variant & VARIANT_RL) && (j == 2 || (&code.reg1)[j+1] == 0))) { if (((&code.reg1)[j] & 0xE0) != REG_R) { // this operand must be general purpose register errors.reportLine(ERR_WRONG_REG_TYPE); } } else if (((&code.reg1)[j] & 0xE0) != regType) { // wrong register type errors.reportLine(ERR_WRONG_REG_TYPE); } } if (lineError) return; // skip additional errors } // check destination register if (code.dest) { if (variant & VARIANT_SPECD) { // must be special register if ((code.dest & 0xE0) <= REG_V) errors.reportLine(ERR_WRONG_REG_TYPE); } else if (variant & VARIANT_R0) { if ((code.dest & 0xE0) != REG_R) { // destination must be general purpose register errors.reportLine(ERR_WRONG_REG_TYPE); } } else if ((code.dest & 0xE0) != regType && code.dest != 2) { // wrong register type errors.reportLine(ERR_WRONG_REG_TYPE); } else if ((code.dest == 2) ^ ((variant & VARIANT_M0) != 0)) { // operands in wrong order errors.reportLine(ERR_OPERANDS_WRONG_ORDER); } if (lineError) return; // skip additional errors } if ((variant & (VARIANT_D0 | VARIANT_D1 | VARIANT_D2)) != 0 && code.dest != 0) { // should not have destination errors.reportLine(ERR_NO_DESTINATION); } if ((variant & (VARIANT_D0 | VARIANT_D1)) == 0 && code.dest == 0) { // should have destination errors.reportLine(ERR_MISSING_DESTINATION); } // check mask register if ((code.etype & XPR_FALLBACK) && !(code.etype & XPR_MASK)) { // fallback but no mask code.mask = 7; // no mask } if ((code.etype & (XPR_MASK | XPR_FALLBACK)) && (code.mask & 7) != 7) { // mask used if ((code.mask & 0xE0) != regType) { // wrong type for mask register errors.reportLine(ERR_WRONG_REG_TYPE); } else if ((code.fallback & 0xE0) != regType && (code.fallback & 0x1F) != 0x1F) { // wrong type for fallback registser if ((variant & VARIANT_RL) && code.fallback == code.reg1) { // fallback has been assigned to reg1 in CAssembler::checkCode1, but reg1 is g.p. register code.fallback = 0x5F; } else errors.reportLine(ERR_WRONG_REG_TYPE); } if ((code.etype & XPR_FALLBACK) && (variant & VARIANT_F0)) { // cannot have fallback register errors.reportLine(ERR_CANNOT_HAVEFALLBACK1); } // check if fallback is the right register if (code.etype & XPR_FALLBACK) { if (code.numOp >= 3 && code.fallback != code.reg1) { errors.reportLine(ERR_3OP_AND_FALLBACK); } } } // check scale factor const int dataSizeTable[8] = { 1, 2, 4, 8, 16, 4, 8, 16 }; // data size for each operant type int8_t scale = code.scale; if (scale == 0) scale = 1; if (((code.formatp->scale & 4) && scale != -1) // scale must be -1 || (((code.formatp->scale & 6) == 2) && scale != dataSizeTable[code.dtype & 7]) // scale must match operand type || (((code.formatp->scale & 6) == 0 && scale != 1 && (code.index & 0x1F) != 0x1F))) { // scale must be 1 errors.reportLine(ERR_SCALE_FACTOR); } // check vector length int numOpt = ((code.etype & XPR_SCALAR) != 0) + ((code.etype & XPR_LENGTH) != 0) + ((code.etype & XPR_BROADC) != 0); if (numOpt == 0 && (code.etype & XPR_MEM) && (code.formatp->vect & ~0x10) && !(code.etype & XPR_LIMIT) && !(code.formatp->vect & 0x80)) { errors.reportLine(ERR_LENGTH_OPTION_MISS); return; // missing length option } // check immediate type if ((code.etype & XPR_FLT) && (variant & VARIANT_I2)) { // immediate should be integer code.etype = (code.etype & ~XPR_FLT) | XPR_INT; //code.value.i = (int64_t)code.value.d; code.value.i = value0; } if ((code.etype & XPR_INT) && !(code.etype & (XPR_LIMIT | XPR_INT2))) { // check if value fits specified operand type int ok = 1; switch (code.dtype & 0x1F) { case TYP_INT8 & 0x1F: ok = code.fitNum & (IFIT_I8 | IFIT_U8); break; case TYP_INT16 & 0x1F: ok = code.fitNum & (IFIT_I16 | IFIT_U16); break; case TYP_INT32 & 0x1F: ok = code.fitNum & (IFIT_I32 | IFIT_U32); break; } if (!ok && (instructionlistId[code.instr1].opimmediate & ~0x10) != OPI_INT32) { errors.reportLine(ERR_CONSTANT_TOO_LARGE); } } // check options if ((code.etype & XPR_OPTIONS) && !(variant & VARIANT_On) && code.formatp->category != 4) { errors.reportLine(ERR_CANNOT_HAVE_OPTION); } // details for unsigned variants if (code.dtype & TYP_UNS) { // an unsigned type is specified if ((variant & VARIANT_U3) && code.instruction == II_COMPARE && code.optionbits) code.optionbits |= 8; // unsigned compare } if (section) code.section = section; // insert section } // find reason why no format fits, and return error number uint32_t CAssembler::checkCodeE(SCode & code) { // check fallback if ((code.etype & XPR_FALLBACK) && code.fallback != code.dest) { if (((code.etype & XPR_MEM) && (code.dest & REG_V)) || code.index) return ERR_CANNOT_HAVEFALLBACK2; if (instructionlistId[code.instr1].sourceoperands >= 3) return ERR_3OP_AND_FALLBACK; } // check three-operand instructions if (instructionlistId[code.instr1].sourceoperands >= 3 && code.reg1 != code.dest && (code.etype & XPR_MEM) && ((code.dest & REG_V) || code.index)) { return ERR_3OP_AND_MEM; } return ERR_NO_INSTRUCTION_FIT; // any other reason } // optimize instruction. replace by more efficient instruction if possible void CAssembler::optimizeCode(SCode & code) { // is it a vector instruction? bool hasVector = ((code.dest | code.reg1) & REG_V) != 0; // is it a floating point instruction? bool isFloat = (code.dtype & TYP_FLOAT) != 0; if (code.instruction & II_JUMP_INSTR) { // jump instruction // optimize immediate jump offset operand if ((code.instruction & 0xFF) == II_SUB && (code.etype & XPR_IMMEDIATE) == XPR_INT && code.value.i >= -0x7F && code.value.i <= 0x80 && cmd.optiLevel && ((code.dtype & 0xFF) == (TYP_INT32 & 0xFF) || ((code.dtype & 0xFF) <= (TYP_INT32 & 0xFF) && (code.dtype & TYP_PLUS)))) { // subtract with conditional jump with 8-bit immediate and 8-bit address // should be replaced by addition of the negative constant int32_t isym = 0; if (code.etype & XPR_SYM1) isym = findSymbol(code.sym1); if (isym <= 0 || symbols[isym].st_section == section || code_size <= (1 << 9)) { // we are not sure yet, but chances are good that the address fits an 8-bit field. Replace sub by add code.value.i = -code.value.i; // change sign of immediate constant code.instruction ^= (II_SUB ^ II_ADD); // replace sub with add if ((code.instruction & 0xFFFF00) == II_JUMP_CARRY) code.instruction ^= 0x100; // carry condition is inverted } } if ((code.fitNum & (IFIT_J16 | IFIT_J32) && (code.etype & XPR_IMMEDIATE) == XPR_INT && (code.instruction & 0xFE) == II_ADD)) { // replace add with sub or vice versa code.value.i = -code.value.i; // change sign of immediate constant code.instruction ^= (II_SUB ^ II_ADD); if ((code.instruction & 0xFFFF00) == II_JUMP_CARRY) code.instruction ^= 0x100; // carry condition is inverted code.fitNum |= (code.fitNum & IFIT_J) >> 1; // signal that it fits } } else { // other instruction. optimize immediate operand if ((code.etype & XPR_INT) /* && !(code.etype & (XPR_OFFSET | XPR_LIMIT | XPR_SYM1))*/ ) { if ((code.instruction & 0xFFFFFFFE) == II_ADD && (code.fitNum & IFIT_J8) != 0) { // we can make the instruction smaller by changing the sign of the constant and exchange add and sub // (we don't have to do this for 0x8000 and 0x80000000 because the can be fitted as 1 << x) code.instruction ^= (II_ADD ^ II_SUB); // replace add with sub or vice versa code.value.i = -code.value.i; // change sign of immediate constant code.fitNum |= (code.fitNum & IFIT_J) >> 1; // signal that it fits } else if (code.instruction == II_SUB && (code.fitNum & (IFIT_I16SH16 | IFIT_I16)) && !(code.fitNum & IFIT_I8) && code.value.w != 0x80000000U && code.value.w != 0xFFFF8000U && code.dest == code.reg1 && !hasVector && (((uint8_t)code.dtype == (uint8_t)TYP_INT32) || (((uint8_t)code.dtype < (uint8_t)TYP_INT32) && (code.dtype & TYP_PLUS)))) { code.instruction = II_ADD; // replace sub with add code.value.i = -code.value.i; // change sign of immediate constant } else if (code.instruction == II_SUB && (code.fitNum & IFIT_I8SHIFT) && !(code.fitNum & IFIT_I8) && !isFloat && code.dest == code.reg1 && (((uint8_t)code.dtype >= (uint8_t)TYP_INT32) || (code.dtype & TYP_PLUS))) { code.instruction = II_ADD; // replace sub with add code.value.i = -code.value.i; // change sign of immediate constant code.fitNum &= ~(IFIT_I16 | IFIT_I16SH16 | IFIT_I32SH32); } else if (code.instruction == II_SUB && (code.fitNum & IFIT_I32SH32) && !(code.fitNum & (IFIT_I16SHIFT | IFIT_I32)) && (((uint8_t)code.dtype == (uint8_t)TYP_INT64) || (code.dtype & TYP_PLUS)) && !isFloat) { code.instruction = II_ADD; // replace sub with add code.value.i = -code.value.i; // change sign of immediate constant } else if ((code.instruction == II_MOVE || code.instruction == II_AND) && (code.fitNum & IFIT_U32) && !(code.fitNum & (IFIT_I32 | IFIT_I16SHIFT)) && ((uint8_t)code.dtype == (uint8_t)TYP_INT64) && !hasVector) { code.dtype = TYP_INT32; // changing type to int32 will zero extend } /*else if (code.instruction == II_MOVE && (code.fitNum & IFIT_U16) && !(code.fitNum & IFIT_I16) && ((uint8_t)code.dtype >= (uint8_t)TYP_INT32) && !hasVector && !(code.etype & (XPR_REG | XPR_MEM | XPR_OPTION | XPR_SYM1))) { code.instruction = II_MOVE_U; code.dtype = TYP_INT64; } */ else if (code.instruction == II_OR && (code.value.u & (code.value.u-1)) == 0 && !(code.fitNum & IFIT_I8)) { code.instruction = II_SET_BIT; // OR with a power of 2 code.value.u = bitScanReverse(code.value.u); code.fitNum = IFIT_I8 | IFIT_I16 | IFIT_I32; } else if (code.instruction == II_AND && (~code.value.u & (~code.value.u-1)) == 0 && !(code.fitNum & IFIT_I8)) { code.instruction = II_CLEAR_BIT; // AND with ~(a power of 2) code.value.u = bitScanReverse(~code.value.u); code.fitNum = IFIT_I8 | IFIT_I16 | IFIT_I32; } else if (code.instruction == II_XOR && (code.value.u & (code.value.u-1)) == 0 && !(code.fitNum & IFIT_I8)) { code.instruction = II_TOGGLE_BIT; // XOR with a power of 2 code.value.u = bitScanReverse(code.value.u); code.fitNum = IFIT_I8 | IFIT_I16 | IFIT_I32; } } if ((code.etype & XPR_FLT) && !(code.etype & (XPR_OFFSET | XPR_LIMIT | XPR_SYM1))) { if (code.instruction == II_SUB && (code.fitNum & FFIT_16) && (uint8_t)code.dtype >= (uint8_t)TYP_FLOAT16) { code.instruction = II_ADD; // replace sub with add code.value.d = -code.value.d; // change sign of immediate constant } } } // optimize -float as toggle_bit if (code.instruction == II_SUB_REV && (code.etype & XPR_IMMEDIATE) && (code.dtype & TYP_FLOAT) && code.value.i == 0 && (code.etype & XPR_REG1) && !(code.etype & XPR_REG2)) { // code is -v represented as 0-v. replace by flipping bit uint32_t bits = 1 << (code.dtype & 7); // number of bits in floating point type code.instruction = II_TOGGLE_BIT; code.value.u = bits - 1; code.etype = ((code.etype & ~XPR_IMMEDIATE) | XPR_INT); } // optimize multiply and divide instructions if ((code.instruction == II_MUL || code.instruction == II_DIV) && (code.etype & XPR_IMMEDIATE)) { if (code.dtype & TYP_INT) { // integer multiplication // check if constant is positive and a power of 2 if (code.value.i <= 0 || (code.value.u & (code.value.u - 1))) return; if (code.instruction == II_MUL) { // integer multiplication by power of 2. replace by left shift code.instruction = II_SHIFT_LEFT; code.value.u = bitScanReverse(code.value.u); } else if (code.dtype & TYP_UNS) { // unsigned division by power of 2. replace by right shift // We are not optimizing signed division because this requires multiple instructions and registers code.instruction = II_SHIFT_RIGHT_U; code.value.u = bitScanReverse(code.value.u); } } else if (code.dtype & TYP_FLOAT) { // floating point multiplication or division // check if constant is a power of 2 int shiftCount = 0xFFFFFFFF; // shift count to replace multiplication by power of 2 if ((code.etype & XPR_INT) && code.value.i > 0 && (code.value.u & (code.value.u-1)) == 0) { // positive integer power of 2 shiftCount = bitScanReverse(code.value.u); if (code.instruction == II_DIV) shiftCount = -shiftCount; } else if ((code.etype & XPR_FLT) && code.value.d != 0.) { int32_t exponent = (code.value.u >> 52) & 0x7FF; // exponent field of double if ((code.value.u & ((uint64_t(1) << 52) - 1)) == 0 && exponent != 0 && exponent != 0x7FF) { // value is a power of 2, and not inf, nan, or subnormal shiftCount = exponent - 0x3FF; if (code.instruction == II_DIV) shiftCount = -shiftCount; } } if (shiftCount == (int)0xFFFFFFFF) return; // not a power of 2. cannot optimize if (shiftCount >= 0 || cmd.optiLevel >= 3) { // replace by mul_2pow instruction // use negative powers of 2 only in optimization level 3, because subnormals are ignored code.instruction = II_MUL_2POW; code.value.i = shiftCount; code.etype = (code.etype & ~XPR_IMMEDIATE) | XPR_INT; } else if (code.instruction == II_DIV) { // replace division by power of 2 to multiplication by the reciprocal code.instruction = II_MUL; if (code.etype & XPR_FLT) code.value.d = 1. / code.value.d; else { code.value.d = 1. / double((uint64_t)1 << (-shiftCount)); code.etype = (code.etype & ~XPR_IMMEDIATE) | XPR_FLT; } } } } } void insertMem(SCode & code, SExpression & expr) { // insert memory operand into code structure if (code.value.i && expr.value.i) code.etype |= XPR_ERROR; // both have constants if (expr.etype & XPR_OFFSET) code.offset_mem = expr.offset_mem; else code.offset_mem = expr.value.w; code.etype |= expr.etype; code.tokens += expr.tokens; code.sym1 = expr.sym1; code.sym2 = expr.sym2; code.base = expr.base; code.index = expr.index; code.length = expr.length; code.scale = expr.scale; code.symscale1 = expr.symscale1; code.mask |= expr.mask; code.fallback |= expr.fallback; } void insertAll(SCode & code, SExpression & expr) { // insert everything from expression to code structure, OR'ing all bits for (uint32_t i = 0; i < sizeof(SExpression) / sizeof(uint64_t); i++) { (&code.value.u)[i] |= (&expr.value.u)[i]; } }
Go to most recent revision | Compare with Previous | Blame | View Log