Files
TempleOS-Rijn/Compiler/OptPass789A.CPP
T
2018-09-08 21:47:32 +02:00

1156 lines
32 KiB
C++

I64 OptPass789A(CCmpCtrl *cc,COptReg *reg_offsets,U8 *buf,CDbgInfo **_dbg)
{/*cc->pass==7 is first time
cc->pass==8 is second time
cc->pass==9 is third time
cc->pass==9 is fourth time and repeated until size stops shrinking
size is now known
cc->pass==10 is final pass, code is placed into buf.
*/
CIntermediateCode *tempi,*tempi_next;
I64 i,cnt,num_lines=cc->max_line+1-cc->min_line,ip=0,ip2;
U8 *ptr,saved_a1_a2_r[3*sizeof(CICArg)];
CCodeMisc *lb;
CAOT *tempaot;
CAOTAbsAddr *tempa;
CAOTImportExport *tempie;
CAOTHeapGlbl *temphg;
CAOTHeapGlblRef *temphgr;
CDbgInfo *dbg_info;
CAOTCtrl *aotc=cc->aotc;
Bool short_jmp;
CHashClass *tempc;
CHashFun *tempf;
CHashGlblVar *tempg;
CExternUsage *tempeu;
if (_dbg) {
*_dbg=dbg_info=CAlloc(offset(CDbgInfo.body)+sizeof(U32)*(num_lines+1));
dbg_info->min_line=cc->min_line;
dbg_info->max_line=cc->max_line;
if (cc->flags&CCF_AOT_COMPILE)
dbg_info->body[0]=aotc->ip;
else
dbg_info->body[0]=buf;
} else
dbg_info=NULL;
if (Bt(&cc->flags,CCf_PASS_TRACE_PRESENT) &&
Bt(&cc->saved_pass_trace,cc->pass))
"$$BK,1$$$$LTRED$$$$IV,1$$This code gets merged together and patched.\n"
"$$FG$$$$IV,0$$$$BK,0$$";
cc->last_float_op_ic=NULL;
tempi=&cc->coc.coc_root;
tempi->ic_last_start=-1;
tempi->ic_cnt=0;
tempi=tempi->next;
while (tempi->ic_code) {
tempi_next=tempi->next;
if (tempi->ic_flags&ICF_PASS_TRACE && Bt(&cc->saved_pass_trace,cc->pass))
ICPut(cc,tempi);
ip2=ip;
if (cc->flags&CCF_AOT_COMPILE)
ip2+=aotc->ip;
else
ip2+=buf;
cc->cur_ic_float_op_num=0;
if (!(tempi->ic_flags &ICF_CODE_FINAL)) {
tempi->ic_flags=tempi->ic_flags&
~(ICF_PREV_DELETED|ICF_DONT_RESTORE)|ICF_CODE_FINAL;
if (cc->pass==7)
cc->dont_push_float=Btr(&tempi->ic_flags,ICf_DONT_PUSH_FLOAT0);
MemCpy(saved_a1_a2_r,&tempi->a1,3*sizeof(CICArg));
tempi->ic_cnt=0;
tempi->ic_last_start=-1;
if (tempi->a2.type.mode) {
if (tempi->ic_flags & ICF_A2_TO_F64) {
ICFCvt(cc,tempi,REG_RAX,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,FALSE,CN_A2,ip2);
tempi->a2.type=MDF_REG+RT_I64;
tempi->a2.reg=REG_RAX;
tempi->a2.disp=0;
} else if (tempi->ic_flags & ICF_A2_TO_INT) {
ICFCvt(cc,tempi,REG_RAX,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,TRUE,CN_A2,ip2);
tempi->a2.type=MDF_REG+RT_I64;
tempi->a2.reg=REG_RAX;
tempi->a2.disp=0;
}
}
if (tempi->a1.type.mode) {
if (tempi->ic_flags & ICF_A1_TO_F64) {
ICFCvt(cc,tempi,REG_RDX,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,FALSE,CN_A1,ip2);
tempi->a1.type=MDF_REG+RT_I64;
tempi->a1.reg=REG_RDX;
tempi->a1.disp=0;
} else if (tempi->ic_flags & ICF_A1_TO_INT) {
ICFCvt(cc,tempi,REG_RDX,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,TRUE,CN_A1,ip2);
tempi->a1.type=MDF_REG+RT_I64;
tempi->a1.reg=REG_RDX;
tempi->a1.disp=0;
}
}
switch [tempi->ic_code] {
start:
start:
case IC_ABS_ADDR:
ICU16(tempi,0xB848);
ICU64(tempi,tempi->ic_data);
if (buf && cc->flags&CCF_AOT_COMPILE &&
!(cc->flags&(CCF_NO_ABSS|CCF_ASM_EXPRESSIONS))) {
tempa=CAlloc(sizeof(CAOTAbsAddr));
tempa->next=aotc->abss;
tempa->type=AAT_ADD_U64;
aotc->abss=tempa;
tempa->ip=ip2+tempi->ic_cnt-8;
}
break;
case IC_HEAP_GLBL:
ICU16(tempi,0xB848);
ICU64(tempi,0);
temphg=tempi->ic_data;
if (buf && cc->flags&CCF_AOT_COMPILE &&
//TODO:is this necessary--flags?
!(cc->flags&(CCF_NO_ABSS|CCF_ASM_EXPRESSIONS))) {
temphgr=CAlloc(sizeof(CAOTHeapGlblRef));
temphgr->next=temphg->references;
temphg->references=temphgr;
temphgr->ip=ip2+tempi->ic_cnt-8;
}
break;
case IC_ADDR_IMPORT:
ICU8(tempi,0xB8);
ICU32(tempi,0);
if (buf && !(cc->flags&CCF_NO_ABSS)) {
tempg=tempi->ic_data;
tempie=CAlloc(sizeof(CAOTImportExport));
tempie->type=IET_IMM_U32;
tempie->ip=ip2+tempi->ic_cnt-4;
tempie->next=tempg->ie_lst;
tempg->ie_lst=tempie;
}
ICU24(tempi,0xC06348);
break;
case IC_RIP:
ICU16(tempi,0xB848);
ICU64(tempi,ip2+tempi->ic_cnt-2);
if (cc->flags&CCF_AOT_COMPILE && buf &&!(cc->flags&CCF_NO_ABSS)) {
tempa=CAlloc(sizeof(CAOTAbsAddr));
tempa->next=aotc->abss;
tempa->type=AAT_ADD_U64;
aotc->abss=tempa;
tempa->ip=ip2+tempi->ic_cnt-8;
}
break;
end:
ICMov(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
MDF_REG+RT_I64,REG_RAX,0,ip2);
break;
case IC_BR_CARRY:
ICFlagBranch(tempi,ip,0x72820F,buf);
break;
case IC_BR_NOT_CARRY:
ICFlagBranch(tempi,ip,0x73830F,buf);
break;
case IC_BR_ZERO:
ICTestAndBranch(tempi,ip,0x74840F,buf,ip2);
break;
case IC_BR_NOT_ZERO:
ICTestAndBranch(tempi,ip,0x75850F,buf,ip2);
break;
case IC_BR_MM_ZERO:
ICPreIncDec(tempi,SLASH_OP_DEC,ip2);
ICFlagBranch(tempi,ip,0x74840F,buf);
break;
case IC_BR_MM_NOT_ZERO:
ICPreIncDec(tempi,SLASH_OP_DEC,ip2);
ICFlagBranch(tempi,ip,0x75850F,buf);
break;
case IC_BR_EQU_EQU:
ICCmpAndBranch(tempi,FALSE,ip,0x74840F,0x74840F,
0x74840F,0x74840F,buf,ip2);
break;
case IC_BR_EQU_EQU2:
ICCmpAndBranch(tempi,TRUE,ip,0x74840F,0x74840F,
0x74840F,0x74840F,buf,ip2);
break;
case IC_BR_NOT_EQU:
ICCmpAndBranch(tempi,FALSE,ip,0x75850F,0x75850F,
0x75850F,0x75850F,buf,ip2);
break;
case IC_BR_NOT_EQU2:
ICCmpAndBranch(tempi,TRUE,ip,0x75850F,0x75850F,
0x75850F,0x75850F,buf,ip2);
break;
case IC_BR_LESS:
if (tempi->ic_flags&ICF_USE_F64)
ICFCmpAndBranch(cc,tempi, ip,0x72820F,0x77870F,buf,ip2);
else
ICCmpAndBranch(tempi,FALSE,ip,0x72820F,0x7C8C0F,
0x77870F,0x7F8F0F,buf,ip2);
break;
case IC_BR_LESS2:
ICCmpAndBranch(tempi,TRUE,ip,0x72820F,0x7C8C0F,
0x77870F,0x7F8F0F,buf,ip2);
break;
case IC_BR_GREATER_EQU:
if (tempi->ic_flags&ICF_USE_F64)
ICFCmpAndBranch(cc,tempi, ip,0x73830F,0x76860F,buf,ip2);
else
ICCmpAndBranch(tempi,FALSE,ip,0x73830F,0x7D8D0F,
0x76860F,0x7E8E0F,buf,ip2);
break;
case IC_BR_GREATER_EQU2:
ICCmpAndBranch(tempi,TRUE,ip,0x73830F,0x7D8D0F,
0x76860F,0x7E8E0F,buf,ip2);
break;
case IC_BR_GREATER:
if (tempi->ic_flags&ICF_USE_F64)
ICFCmpAndBranch(cc,tempi, ip,0x77870F,0x72820F,buf,ip2);
else
ICCmpAndBranch(tempi,FALSE,ip,0x77870F,0x7F8F0F,
0x72820F,0x7C8C0F,buf,ip2);
break;
case IC_BR_GREATER2:
ICCmpAndBranch(tempi,TRUE,ip,0x77870F,0x7F8F0F,
0x72820F,0x7C8C0F,buf,ip2);
break;
case IC_BR_LESS_EQU:
if (tempi->ic_flags&ICF_USE_F64)
ICFCmpAndBranch(cc,tempi, ip,0x76860F,0x73830F,buf,ip2);
else
ICCmpAndBranch(tempi,FALSE,ip,0x76860F,0x7E8E0F,
0x73830F,0x7D8D0F,buf,ip2);
break;
case IC_BR_LESS_EQU2:
ICCmpAndBranch(tempi,TRUE,ip,0x76860F,0x7E8E0F,
0x73830F,0x7D8D0F,buf,ip2);
break;
case IC_BR_BT:
ICBrBitOps(tempi,ip,0xA30F,0x20BA0F,0x72820F,buf,ip2);
break;
case IC_BR_BTS:
ICBrBitOps(tempi,ip,0xAB0F,0x28BA0F,0x72820F,buf,ip2);
break;
case IC_BR_BTR:
ICBrBitOps(tempi,ip,0xB30F,0x30BA0F,0x72820F,buf,ip2);
break;
case IC_BR_BTC:
ICBrBitOps(tempi,ip,0xBB0F,0x38BA0F,0x72820F,buf,ip2);
break;
case IC_BR_NOT_BT:
ICBrBitOps(tempi,ip,0xA30F,0x20BA0F,0x73830F,buf,ip2);
break;
case IC_BR_NOT_BTS:
ICBrBitOps(tempi,ip,0xAB0F,0x28BA0F,0x73830F,buf,ip2);
break;
case IC_BR_NOT_BTR:
ICBrBitOps(tempi,ip,0xB30F,0x30BA0F,0x73830F,buf,ip2);
break;
case IC_BR_NOT_BTC:
ICBrBitOps(tempi,ip,0xBB0F,0x38BA0F,0x73830F,buf,ip2);
break;
case IC_BR_AND_ZERO:
ICAndBranch(tempi,ip,0x74840F,buf,ip2);
break;
case IC_BR_AND_NOT_ZERO:
ICAndBranch(tempi,ip,0x75850F,buf,ip2);
break;
case IC_SUB_CALL:
lb=OptLabelFwd(tempi->ic_data);
ICU8(tempi,0xE8);
ICU32(tempi,lb->addr-(ip+5));
break;
case IC_JMP:
lb=OptLabelFwd(tempi->ic_data);
short_jmp=ToBool(tempi->ic_flags&ICF_SHORT_JMP);
if (!buf && lb->addr!=INVALID_PTR &&
MIN_I8+5<lb->addr-ip<MAX_I8-5)
short_jmp=TRUE;
if (short_jmp) {
tempi->ic_flags|=ICF_SHORT_JMP;
i=lb->addr-(ip+2);
if (buf || i)
ICU16(tempi,i<<8+0xEB);
else
tempi->ic_code=IC_NOP1;
} else {
i=lb->addr-(ip+5);
ICU8(tempi,0xE9);
ICU32(tempi,i);
}
break;
case IC_LABEL:
lb=tempi->ic_data;
lb->addr=ip;
if (lb->flags&CMF_POP_CMP) {
ICAddRSP(tempi,-8,FALSE);
ICAddRSP(tempi,8,FALSE);
}
if (lb->type==CMT_ASM_LABEL)
lb->addr+=lb->ip;
break;
case IC_STR_CONST:
case IC_GET_LABEL:
lb=tempi->ic_data;
if (cc->flags&CCF_AOT_COMPILE)
i=lb->addr+aotc->ip;
else
i=lb->addr+buf;
ICLea(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
MDF_RIP_DISP32+RT_PTR,0,i,cc,buf,ip2);
break;
case IC_ASM:
tempaot=tempi->ic_data;
tempi->ic_cnt+=tempaot->aot_U8s;
if (buf) {
MemCpy(buf+ip,tempaot->buf,tempaot->aot_U8s);
Free(tempaot->buf);
tempaot->buf=buf;
tempaot->ip=ip;
tempaot->ip2=ip2;
if (cc->flags&CCF_AOT_COMPILE)
CmpFixUpAOTAsm(cc,tempaot);
else
CmpFixUpJITAsm(cc,tempaot);
cnt=tempi->ic_cnt;
goto op789A_skip_copy;
}
break;
case IC_CALL:
i=tempi->ic_data-(ip2+5);
if (!(MIN_I32<=i<=MAX_I32) && !(cc->flags&CCF_AOT_COMPILE)) {
ICU16(tempi,0xBB48);
ICU64(tempi,tempi->ic_data);
ICU16(tempi,0xD3FF);
} else {
ICU8(tempi,0xE8);
ICU32(tempi,i);
}
break;
case IC_CALL_EXTERN: //Only for static modules
ICU8(tempi,0xE8);
ICU32(tempi,0);
if (buf) {
tempf=tempi->ic_data;
tempeu=CAlloc(sizeof(CExternUsage));
tempeu->next=tempf->ext_lst;
tempf->ext_lst=tempeu;
tempeu->ip=ip2+1;
}
break;
case IC_CALL_INDIRECT2:
ICU16(tempi,0xBB48);
if (cc->flags&CCF_AOT_COMPILE) i=ip2+tempi->ic_cnt;
ICU64(tempi,tempi->ic_data);
ICU16(tempi,0x13FF);
if (buf && cc->flags&CCF_AOT_COMPILE&& !(cc->flags&CCF_NO_ABSS)) {
tempa=CAlloc(sizeof(CAOTAbsAddr));
tempa->next=aotc->abss;
tempa->type=AAT_ADD_U64;
aotc->abss=tempa;
tempa->ip=i;
}
break;
case IC_CALL_IMPORT:
if (GetOption(OPTf_USE_IMM64)) {
ICU16(tempi,0xBB48);
ICU64(tempi,0);
if (buf) {
tempf=tempi->ic_data;
tempie=CAlloc(sizeof(CAOTImportExport));
tempie->type=IET_IMM_I64;
tempie->ip=ip2+tempi->ic_cnt-8;
tempie->next=tempf->ie_lst;
tempf->ie_lst=tempie;
}
ICU16(tempi,0xD3FF);
} else {
ICU8(tempi,0xE8);
ICU32(tempi,0);
if (buf) {
tempf=tempi->ic_data;
tempie=CAlloc(sizeof(CAOTImportExport));
tempie->type=IET_REL_I32;
tempie->ip=ip2+tempi->ic_cnt-4;
tempie->next=tempf->ie_lst;
tempf->ie_lst=tempie;
}
}
break;
end:
tempi->ic_flags&=~ICF_CODE_FINAL;
break;
case IC_LEAVE:
if (cc->htc.fun) {
if (Bt(&cc->htc.fun->flags,Ff_INTERRUPT))
ICPopRegs(tempi,REGG_CLOBBERED|cc->htc.fun->used_reg_mask&
(REGG_LOCAL_VARS|REGG_LOCAL_NON_PTR_VARS|REGG_STK_TEMP));
else
ICPopRegs(tempi,cc->htc.fun->used_reg_mask&
(REGG_LOCAL_VARS|REGG_LOCAL_NON_PTR_VARS));
}
if (tempi->ic_data<=MAX_I16) {
if (tempi->ic_data)
ICU8(tempi,0xC9); //LEAVE
else
ICU8(tempi,0x5D); //POP RBP
} else {
ICAddRSP(tempi,tempi->ic_data);
ICU8(tempi,0x5D); //POP RBP
}
if (cc->htc.fun && Bt(&cc->htc.fun->flags,Ff_INTERRUPT)) {
if (Bt(&cc->htc.fun->flags,Ff_HASERRCODE))
ICAddRSP(tempi,8);
ICU16(tempi,0xCF48);
} else if (cc->htc.fun && cc->htc.fun->arg_cnt &&
(Bt(&cc->htc.fun->flags,Ff_RET1) ||
Bt(&cc->htc.fun->flags,Ff_ARGPOP)) &&
!Bt(&cc->htc.fun->flags,Ff_NOARGPOP)) {
ICU8(tempi,0xC2);
ICU16(tempi,cc->htc.fun->arg_cnt<<3);
} else
ICU8(tempi,0xC3);
break;
case IC_RET:
ICU8(tempi,0xC3);
break;
case IC_FS:
ICZero(tempi,REG_RAX);
ICU32(tempi,0x8B4864);
break;
case IC_GS:
ICZero(tempi,REG_RAX);
ICU32(tempi,0x8B4865);
break;
case IC_MOV_FS:
ICZero(tempi,REG_RAX);
ICU8(tempi,0x64);
//It's ugly to use ic_class here
ICMov(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
MDF_DISP+CmpRawType(tempi->ic_class),REG_RAX,tempi->ic_data,
ip2);
break;
case IC_MOV_GS:
ICZero(tempi,REG_RAX);
ICU8(tempi,0x65);
//It's ugly to use ic_class here
ICMov(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
MDF_DISP+CmpRawType(tempi->ic_class),REG_RAX,tempi->ic_data,
ip2);
break;
case IC_HOLYC_TYPECAST:
ICMov(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
break;
case IC_COM:
ICUnaries(tempi,SLASH_OP_NOT,ip2);
break;
case IC_NOT:
ICNot(tempi,ip2);
break;
case IC_UNARY_MINUS:
if (tempi->r.type.raw_type==RT_F64)
ICFUnaryMinus(cc,tempi,buf,ip2);
else
ICUnaries(tempi,SLASH_OP_NEG,ip2);
break;
case IC_ADDR:
case IC_MOV:
ICMov(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
break;
case IC_DEREF:
ICDeref(tempi,ip2);
break;
case IC_DEREF_PP:
ICDerefPostIncDec(tempi,SLASH_OP_INC,ip2);
break;
case IC_DEREF_MM:
ICDerefPostIncDec(tempi,SLASH_OP_DEC,ip2);
break;
case IC__PP:
if (tempi->ic_flags&ICF_USE_INT)
ICPostIncDec(tempi,SLASH_OP_INC,ip2);
else
ICFPostIncDec(cc,tempi,CMP_TEMPLATE_INC,ip2);
break;
case IC__MM:
if (tempi->ic_flags&ICF_USE_INT)
ICPostIncDec(tempi,SLASH_OP_DEC,ip2);
else
ICFPostIncDec(cc,tempi,CMP_TEMPLATE_DEC,ip2);
break;
case IC_PP_:
if (tempi->ic_flags&ICF_USE_INT)
ICPreIncDec(tempi,SLASH_OP_INC,ip2);
else
ICFPreIncDec(cc,tempi,CMP_TEMPLATE_INC,ip2);
break;
case IC_MM_:
if (tempi->ic_flags&ICF_USE_INT)
ICPreIncDec(tempi,SLASH_OP_DEC,ip2);
else
ICFPreIncDec(cc,tempi,CMP_TEMPLATE_DEC,ip2);
break;
case IC_LEA:
ICLea(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,cc,buf,ip2);
break;
case IC_POWER:
ICFPow(cc,tempi,buf,ip2);
break;
case IC_SHL:
ICShift(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,
0xE0D1E0D3E0C1,0xE0D1E0D3E0C1,ip2);
break;
case IC_SHR:
ICShift(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,
0xE8D1E8D3E8C1,0xF8D1F8D3F8C1,ip2);
break;
case IC_MUL:
if (tempi->ic_flags&ICF_USE_INT)
ICMul(tempi,ip2);
else
ICFMul(cc,tempi,buf,ip2);
break;
case IC_DIV:
if (tempi->ic_flags&ICF_USE_INT)
ICDiv(tempi,ip2);
else
ICFDiv(cc,tempi,buf,ip2);
break;
case IC_MOD:
if (tempi->ic_flags&ICF_USE_INT)
ICMod(tempi,ip2);
else
ICFMod(cc,tempi,ip2);
break;
case IC_AND:
ICAddEct(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,0x23,ip2);
break;
case IC_OR:
ICAddEct(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,0x0B,ip2);
break;
case IC_XOR:
ICAddEct(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,0x33,ip2);
break;
case IC_ADD:
if (tempi->ic_flags&ICF_USE_INT)
ICAddEct(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,0x03,ip2);
else
ICFAdd(cc,tempi,buf,ip2);
break;
case IC_SUB:
if (tempi->ic_flags&ICF_USE_INT)
ICSub(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,ip2);
else
ICFSub(cc,tempi,buf,ip2);
break;
case IC_EQU_EQU:
ICCmp(tempi,0x75,0x75,ip2);
break;
case IC_NOT_EQU:
ICCmp(tempi,0x74,0x74,ip2);
break;
case IC_LESS:
if (tempi->ic_flags&ICF_USE_INT)
ICCmp(tempi,0x73,0x7D,ip2);
else
ICFCmp(cc,tempi,CMP_TEMPLATE_LESS,ip2);
break;
case IC_GREATER_EQU:
if (tempi->ic_flags&ICF_USE_INT)
ICCmp(tempi,0x72,0x7C,ip2);
else
ICFCmp(cc,tempi,CMP_TEMPLATE_GREATER_EQU,ip2);
break;
case IC_GREATER:
if (tempi->ic_flags&ICF_USE_INT)
ICCmp(tempi,0x76,0x7E,ip2);
else
ICFCmp(cc,tempi,CMP_TEMPLATE_GREATER,ip2);
break;
case IC_LESS_EQU:
if (tempi->ic_flags&ICF_USE_INT)
ICCmp(tempi,0x77,0x7F,ip2);
else
ICFCmp(cc,tempi,CMP_TEMPLATE_LESS_EQU,ip2);
break;
case IC_AND_AND:
ICAndAnd(tempi,ip2);
break;
case IC_OR_OR:
ICOrOr(tempi,ip2);
break;
case IC_XOR_XOR:
ICXorXor(tempi,ip2);
break;
case IC_ASSIGN:
ICAssign(tempi,ip2);
break;
case IC_ASSIGN_PP:
ICAssignPostIncDec(tempi,SLASH_OP_INC,ip2);
break;
case IC_ASSIGN_MM:
ICAssignPostIncDec(tempi,SLASH_OP_DEC,ip2);
break;
case IC_SHL_EQU:
ICShiftEqu(tempi,tempi->a1_type_pointed_to,
tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,
0xE0D1E0D3E0C1,0xE0D1E0D3E0C1,ip2);
break;
case IC_SHR_EQU:
ICShiftEqu(tempi,tempi->a1_type_pointed_to,
tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,
0xE8D1E8D3E8C1,0xF8D1F8D3F8C1,ip2);
break;
case IC_MUL_EQU:
if (tempi->ic_flags&ICF_USE_INT)
ICMulEqu(tempi,ip2);
else
ICFOpEqu(cc,tempi,SLASH_OP_FMUL,buf,ip2);
break;
case IC_DIV_EQU:
if (tempi->ic_flags&ICF_USE_INT)
ICDivEqu(tempi,FALSE,ip2);
else
ICFOpEqu(cc,tempi,SLASH_OP_FDIV,buf,ip2);
break;
case IC_MOD_EQU:
if (tempi->ic_flags&ICF_USE_INT)
ICDivEqu(tempi,TRUE,ip2);
else
ICFModEqu(cc,tempi,ip2);
break;
case IC_AND_EQU:
ICAndEqu(tempi,ip2);
break;
case IC_OR_EQU:
ICOrEqu(tempi,ip2);
break;
case IC_XOR_EQU:
ICXorEqu(tempi,ip2);
break;
case IC_ADD_EQU:
if (tempi->ic_flags&ICF_USE_INT)
ICAddSubEctEqu(tempi,tempi->a1_type_pointed_to,
tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,
0x010000000003,ip2);
else
ICFOpEqu(cc,tempi,SLASH_OP_FADD,buf,ip2);
break;
case IC_SUB_EQU:
if (tempi->ic_flags&ICF_USE_INT)
ICAddSubEctEqu(tempi,tempi->a1_type_pointed_to,
tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,
0x29000000052B,ip2);
else
ICFOpEqu(cc,tempi,SLASH_OP_FSUB,buf,ip2);
break;
case IC_SHL_CONST:
ICShift(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
MDF_IMM+RT_I64,0,tempi->ic_data,
0xE0D1E0D3E0C1,0xE0D1E0D3E0C1,ip2);
break;
case IC_SHR_CONST:
ICShift(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,
MDF_IMM+RT_I64,0,tempi->ic_data,
0xE8D1E8D3E8C1,0xF8D1F8D3F8C1,ip2);
break;
case IC_ADD_CONST:
ICAddSubEctImm(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,tempi->ic_data,
0x0003,ip2);
break;
case IC_SUB_CONST:
ICAddSubEctImm(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,tempi->ic_data,
0x052B,ip2);
break;
case IC_ENTER:
ICU32(tempi,0xEC8B4855);
if (tempi->ic_data)
ICAddRSP(tempi,-tempi->ic_data,FALSE);
if (cc->htc.fun) {
if (Bt(&cc->htc.fun->flags,Ff_INTERRUPT))
ICPushRegs(tempi,REGG_CLOBBERED|cc->htc.fun->used_reg_mask&
(REGG_LOCAL_VARS|REGG_LOCAL_NON_PTR_VARS|REGG_STK_TEMP));
else {
if (sys_var_init_flag && i)
ICLocalVarInit(tempi);
ICPushRegs(tempi,cc->htc.fun->used_reg_mask
&(REGG_LOCAL_VARS|REGG_LOCAL_NON_PTR_VARS));
}
for (i=0;i<NUM_REGS;i++)
if (reg_offsets[i]>0 && reg_offsets[i].offset!=MAX_I64) {
tempc=OptClassFwd(reg_offsets[i].m->member_class);
ICMov(tempi,MDF_REG+RT_I64,i,0,MDF_DISP+tempc->raw_type,
REG_RBP,reg_offsets[i].offset,ip2);
}
}
break;
case IC_ADD_RSP:
ICAddRSP(tempi,tempi->ic_data);
break;
case IC_CALL_INDIRECT:
if (MIN_I8<=tempi->ic_data<=MAX_I8) {
ICU24(tempi,0x2454FF); //CALL disp[RSP]
ICU8(tempi,tempi->ic_data);
} else {
ICU24(tempi,0x2494FF); //CALL disp[RSP]
ICU32(tempi,tempi->ic_data);
}
break;
case IC_PUSH:
ICPush(tempi,tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
break;
case IC_POP:
ICU8(tempi,0x58);
break;
case IC_INVLPG:
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICU24(tempi,0x38010F);
break;
case IC_CLFLUSH:
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICU24(tempi,0x38AE0F);
break;
case IC_GET_RFLAGS:
ICU8(tempi,0x9C);
ICPop(tempi,MDF_REG+RT_I64,REG_RAX,0,ip2);
break;
case IC_CARRY:
ICU24(tempi,0xC0920F); //SETC AL
ICU24(tempi,0x01E083); //AND EAX,1
break;
case IC_RDTSC:
ICCopyTemplate(cc,tempi,CMP_TEMPLATE_RDTSC,TRUE,FALSE,FALSE,CN_MAIN);
break;
case IC_HPET:
ICCopyTemplate(cc,tempi,CMP_TEMPLATE_HPET,TRUE,FALSE,FALSE,CN_MAIN);
break;
case IC_SET_RFLAGS:
ICPush(tempi,tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICU8(tempi,0x9D);
break;
case IC_GET_RBP:
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
MDF_REG+RT_I64,REG_RBP,0,ip2);
break;
case IC_SET_RBP:
ICMov(tempi,MDF_REG+RT_I64,REG_RBP,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
break;
case IC_GET_RSP:
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
MDF_REG+RT_I64,REG_RSP,0,ip2);
break;
case IC_SET_RSP:
ICMov(tempi,MDF_REG+RT_I64,REG_RSP,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
break;
case IC_RETURN_VAL:
case IC_SET_RAX:
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
break;
case IC_RETURN_VAL2:
case IC_GET_RAX:
break;
case IC_BT:
ICBitOps(tempi,&tempi->a1,&tempi->a2,tempi->next,
0xA30F,0x20BA0F,ip2);
break;
case IC_BTS:
case IC_LBTS:
ICBitOps(tempi,&tempi->a1,&tempi->a2,tempi->next,
0xAB0F,0x28BA0F,ip2);
break;
case IC_BTR:
case IC_LBTR:
ICBitOps(tempi,&tempi->a1,&tempi->a2,tempi->next,
0xB30F,0x30BA0F,ip2);
break;
case IC_BTC:
case IC_LBTC:
ICBitOps(tempi,&tempi->a1,&tempi->a2,tempi->next,
0xBB0F,0x38BA0F,ip2);
break;
case IC_BSF:
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICU32(tempi,0xC0BC0F48);
ICU16(tempi,0x0375);
ICU24(tempi,0xD0F748);
break;
case IC_BSR:
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICU32(tempi,0xC0BD0F48);
ICU16(tempi,0x0375);
ICU24(tempi,0xD0F748);
break;
case IC_SIGN_I64:
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICCopyTemplate(cc,tempi,CMP_TEMPLATE_SIGN_I64,
TRUE,FALSE,FALSE,CN_MAIN);
break;
case IC_TOUPPER:
ICToUpper(tempi,ip2);
break;
case IC_TO_I64:
ICToI64(cc,tempi,ip2);
break;
case IC_TO_F64:
ICToF64(cc,tempi,ip2);
break;
case IC_TO_BOOL:
ICToBool(cc,tempi,ip2);
break;
case IC_SQR:
ICFTemplateFun(cc,tempi,CMP_TEMPLATE_SQR,ip2);
break;
case IC_ABS:
ICFTemplateFun(cc,tempi,CMP_TEMPLATE_ABS,ip2);
break;
case IC_SQRT:
ICFTemplateFun(cc,tempi,CMP_TEMPLATE_SQRT,ip2);
break;
case IC_SIN:
ICFTemplateFun(cc,tempi,CMP_TEMPLATE_SIN,ip2);
break;
case IC_COS:
ICFTemplateFun(cc,tempi,CMP_TEMPLATE_COS,ip2);
break;
case IC_TAN:
ICFTemplateFun(cc,tempi,CMP_TEMPLATE_TAN,ip2);
break;
case IC_ATAN:
ICFTemplateFun(cc,tempi,CMP_TEMPLATE_ATAN,ip2);
break;
case IC_ABS_I64:
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICU24(tempi,0xC08548);
ICU16(tempi,0x0379);
ICU24(tempi,0xD8F748);
break;
case IC_MIN_I64:
ICMinMax(tempi,0x4F,ip2);
break;
case IC_MAX_I64:
ICMinMax(tempi,0x4C,ip2);
break;
case IC_MIN_U64:
ICMinMax(tempi,0x47,ip2);
break;
case IC_MAX_U64:
ICMinMax(tempi,0x42,ip2);
break;
case IC_MOD_U64:
ICModU64(tempi,ip2);
break;
case IC_SQR_I64:
ICSqr(tempi,SLASH_OP_IMUL,ip2);
break;
case IC_SQR_U64:
ICSqr(tempi,SLASH_OP_MUL,ip2);
break;
case IC_SWAP_U8:
case IC_SWAP_U16:
case IC_SWAP_U32:
case IC_SWAP_I64:
ICSwap(tempi,ip2);
break;
case IC_QUE_INIT:
ICQueInit(tempi,ip2);
break;
case IC_QUE_INS:
ICQueIns(tempi,ip2);
break;
case IC_QUE_INS_REV:
ICQueInsRev(tempi,ip2);
break;
case IC_QUE_REM:
ICQueRem(tempi,ip2);
break;
case IC_STRLEN:
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICCopyTemplate(cc,tempi,CMP_TEMPLATE_STRLEN,TRUE,FALSE,FALSE,CN_MAIN);
break;
case IC_IN_U32:
if (tempi->a1.type&MDF_IMM) {
ICU16(tempi,0xC033);
if (tempi->a1.disp<=MAX_U8)
ICU16(tempi,0xE5+tempi->a1.disp<<8);
else {
ICU32(tempi,0xBA00+OC_OP_SIZE_PREFIX+tempi->a1.disp<<16);
ICU8(tempi,0xED);
}
} else {
ICMov(tempi,MDF_REG+RT_I64,REG_RDX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICU16(tempi,0xC033);
ICU8(tempi,0xED);
}
break;
case IC_IN_U16:
if (tempi->a1.type&MDF_IMM) {
ICU16(tempi,0xC033);
if (tempi->a1.disp<=MAX_U8)
ICU24(tempi,0xE500+OC_OP_SIZE_PREFIX+tempi->a1.disp<<16);
else {
ICU32(tempi,0xBA00+OC_OP_SIZE_PREFIX+tempi->a1.disp<<16);
ICU16(tempi,0xED00+OC_OP_SIZE_PREFIX);
}
} else {
ICMov(tempi,MDF_REG+RT_I64,REG_RDX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICU16(tempi,0xC033);
ICU16(tempi,0xED00+OC_OP_SIZE_PREFIX);
}
break;
case IC_IN_U8:
if (tempi->a1.type&MDF_IMM) {
ICU16(tempi,0xC033);
if (tempi->a1.disp<=MAX_U8)
ICU16(tempi,0xE4+tempi->a1.disp<<8);
else {
ICU32(tempi,0xBA00+OC_OP_SIZE_PREFIX+tempi->a1.disp<<16);
ICU8(tempi,0xEC);
}
} else {
ICMov(tempi,MDF_REG+RT_I64,REG_RDX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICU16(tempi,0xC033);
ICU8(tempi,0xEC);
}
break;
case IC_OUT_U32:
if (tempi->a2.type&MDF_IMM) {
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
if (tempi->a2.disp<=MAX_U8)
ICU16(tempi,0xE7+tempi->a2.disp<<8);
else {
ICU32(tempi,0xBA00+OC_OP_SIZE_PREFIX+tempi->a2.disp<<16);
ICU8(tempi,0xEF);
}
} else {
ICMov(tempi,MDF_REG+RT_I64,REG_RDX,0,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,ip2);
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICU8(tempi,0xEF);
}
break;
case IC_OUT_U16:
if (tempi->a2.type&MDF_IMM) {
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
if (tempi->a2.disp<=MAX_U8)
ICU24(tempi,0xE700+OC_OP_SIZE_PREFIX+tempi->a2.disp<<16);
else {
ICU32(tempi,0xBA00+OC_OP_SIZE_PREFIX+tempi->a2.disp<<16);
ICU16(tempi,0xEF00+OC_OP_SIZE_PREFIX);
}
} else {
ICMov(tempi,MDF_REG+RT_I64,REG_RDX,0,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,ip2);
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICU16(tempi,0xEF00+OC_OP_SIZE_PREFIX);
}
break;
case IC_OUT_U8:
if (tempi->a2.type&MDF_IMM) {
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
if (tempi->a2.disp<=MAX_U8)
ICU16(tempi,0xE6+tempi->a2.disp<<8);
else {
ICU32(tempi,0xBA00+OC_OP_SIZE_PREFIX+tempi->a2.disp<<16);
ICU8(tempi,0xEE);
}
} else {
ICMov(tempi,MDF_REG+RT_I64,REG_RDX,0,
tempi->a2.type,tempi->a2.reg,tempi->a2.disp,ip2);
ICMov(tempi,MDF_REG+RT_I64,REG_RAX,0,
tempi->a1.type,tempi->a1.reg,tempi->a1.disp,ip2);
ICU8(tempi,0xEE);
}
break;
case IC_NOBOUND_SWITCH:
ICSwitch(tempi,ip,TRUE,cc,buf,ip2);
break;
case IC_SWITCH:
ICSwitch(tempi,ip,FALSE,cc,buf,ip2);
break;
case IC_NOP1:
case IC_NOP2:
OptFree(tempi);
goto op789A_next;
case IC_CALL_START:
case IC_PUSH_REGS:
ICPushRegs(tempi,tempi->ic_data);
break;
case IC_CALL_END:
ICPopRegs(tempi,tempi->ic_data);
if (tempi->r.type.mode)
ICMov(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
MDF_REG+RT_I64,REG_RAX,0,ip2);
break;
case IC_POP_REGS:
ICPopRegs(tempi,tempi->ic_data);
break;
case IC_PUSH_CMP:
case IC_CALL_END2:
case IC_END:
case IC_ADD_RSP1:
break;
default:
"Pass:%d Missing IC handler\n",cc->pass;
ICPut(cc,tempi);
LexExcept(cc,"Compiler Optimization Error at ");
}
if (tempi->r.type.mode) {
if (tempi->ic_flags & ICF_R_TO_F64) {
if (tempi->ic_code==IC_PUSH_CMP) {
ICU24(tempi,0x242CDF); //FILD U64 [RSP]
ICU24(tempi,0x241CDD); //FSTP U64 [RSP]
} else {
ICFCvt(cc,tempi,REG_RAX,tempi->r.type,tempi->r.reg,tempi->r.disp,
FALSE,CN_R,ip2);
if (!Bt(&tempi->ic_flags,ICf_DONT_POP_FLOAT0+
cc->cur_ic_float_op_num-1))
ICMov(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
MDF_REG+RT_I64,REG_RAX,0,ip2);
}
} else if (tempi->ic_flags & ICF_R_TO_INT) {
ICFCvt(cc,tempi,REG_RAX,tempi->r.type,tempi->r.reg,tempi->r.disp,
TRUE,CN_R,ip2);
ICMov(tempi,tempi->r.type,tempi->r.reg,tempi->r.disp,
MDF_REG+RT_I64,REG_RAX,0,ip2);
}
}
}
cnt=tempi->ic_cnt;
if (tempi->ic_flags&ICF_DEL_PREV_INS) {
if (cc->pass>8)
cnt=tempi->ic_last_start;
tempi->ic_flags&=~ICF_DEL_PREV_INS;
}
if (cnt && buf)
MemCpy(buf+ip,tempi->ic_body,cnt);
op789A_skip_copy:
if (dbg_info && cc->min_line<=tempi->ic_line<=cc->max_line) {
i=tempi->ic_line-cc->min_line;
if (!dbg_info->body[i])
dbg_info->body[i]=ip2;
}
if (tempi->ic_flags&ICF_PASS_TRACE &&
Bt(&cc->saved_pass_trace,cc->pass) && cnt) {
"$$RED$$";
if (buf)
Un(buf+ip,cnt,64);
else
Un(tempi->ic_body,cnt,64);
"$$FG$$";
}
if (!(tempi->ic_flags&(ICF_CODE_FINAL|ICF_DONT_RESTORE)))
MemCpy(&tempi->a1,saved_a1_a2_r,3*sizeof(CICArg));
ip+=cnt;
if (tempi->ic_cnt>=IC_BODY_SIZE && tempi->ic_code!=IC_ASM)
throw('Compiler');
op789A_next:
tempi=tempi_next;
}
lb=cc->coc.coc_next_misc;
while (lb!=&cc->coc.coc_next_misc) {
switch (lb->type) {
case CMT_STR_CONST:
lb->addr=ip;
if (buf)
MemCpy(buf+ip,lb->str,lb->st_len);
ip+=lb->st_len;
break;
case CMT_JMP_TABLE:
lb->addr=ip;
ptr=buf+lb->addr;
if (lb->flags&(CMF_I8_JMP_TABLE|CMF_U8_JMP_TABLE)) {
if (buf)
for (i=0;i<lb->range;i++)
*ptr++=lb->jmp_table[i]->addr-lb->begin->addr;
ip+=lb->range;
} else if (lb->flags&(CMF_I16_JMP_TABLE|CMF_U16_JMP_TABLE)) {
if (buf)
for (i=0;i<lb->range;i++)
*ptr(U16 *)++=lb->jmp_table[i]->addr-lb->begin->addr;
ip+=lb->range<<1;
} else {
if (buf)
for (i=0;i<lb->range;i++) {
if (cc->flags&CCF_AOT_COMPILE && !(cc->flags&CCF_NO_ABSS)) {
tempa=CAlloc(sizeof(CAOTAbsAddr));
tempa->next=aotc->abss;
tempa->type=AAT_ADD_U32;
aotc->abss=tempa;
tempa->ip=aotc->ip+lb->addr+i<<2;
*ptr(U32 *)++=lb->jmp_table[i]->addr+aotc->ip;
} else
*ptr(U32 *)++=lb->jmp_table[i]->addr+buf;
}
ip+=lb->range<<2;
}
break;
case CMT_FLOAT_CONSTS:
lb->addr=ip;
if (buf)
MemCpy(buf+lb->addr,lb->float_consts,lb->num_consts*sizeof(F64));
ip+=lb->num_consts*sizeof(F64);
break;
}
lb=lb->next;
}
if (dbg_info) {
if (cc->flags&CCF_AOT_COMPILE)
dbg_info->body[num_lines]=ip+aotc->ip;
else
dbg_info->body[num_lines]=ip+buf;
}
return ip;
}