U0 OptPass3(CLex *lx,COptReg *reg_offsets) { CHashClass *tempc,*tempc1,*tempc2; CLexCodeMisc *lb; CIntermediateCode *tempi,*tempi1,*tempi2,*tempi_next, *tempil1,*tempil2; I64 code,i,j,l,member_cnt,used_reg_mask=0; CMemberLst *tempm; COptMemberVar *mv=NULL; CPrsStk *ps=lx->ps; ps->ptr=0; ps->ptr2=0; for (i=0;ihtc.fun) { member_cnt=lx->htc.fun->member_cnt; if (Bt(&lx->htc.fun->flags,Ff_DOT_DOT_DOT)) member_cnt+=2; mv=CAlloc(member_cnt*sizeof(COptMemberVar)); member_cnt=0; tempm=lx->htc.fun->member_lst; while (tempm) { tempc=OptClassFwd(tempm->member_class); if (0<=tempm->regreg)) PrintWarn("Reg in use\n $$LK,\"FL:%s,%d\"$$ '%s' in '%s'\n", lx->cur_lfn->name,lx->cur_lfn->line_num,tempm->str,lx->htc.fun->str); reg_offsets[tempm->reg].offset=tempm->offset; reg_offsets[tempm->reg].m=tempm; } else if (tempc->raw_type!=RT_F64 && tempm->reg!=REG_NONE || tempm->reg==REG_ALLOC) { if (tempm->reg==REG_ALLOC) mv[member_cnt].score=MAX_I64/2; //big but not too big mv[member_cnt].offset_start=tempm->offset; mv[member_cnt].offset_end=tempm->offset+MaxI64(1,tempm->size); mv[member_cnt++].m=tempm; } else { if (tempm->reg==REG_ALLOC) PrintWarn("Can't reg var\n $$LK,\"FL:%s,%d\"$$ '%s' in '%s'\n", lx->cur_lfn->name,lx->cur_lfn->line_num,tempm->str,lx->htc.fun->str); } tempm=tempm->next; } } else member_cnt=0; tempi=lx->lcc.lcc_root.next; while (code=tempi->ic_code) { tempi_next=tempi->next; if (code==IC_NOP1) OptFree(tempi); else { if (tempil1=OptLag(tempi)) { if (!(tempil2=OptLag(tempil1))) tempil2=&opt_nop; } else tempil1=tempil2=&opt_nop; tempc=tempi->ic_class; nobound_switch (cmp_arg_cnt[code]) { case 9: ps->ptr-=tempi->ic_data>>3; break; case 2: tempi2=PrsPop(ps); if (tempi2->ic_flags&ICF_R_TO_F64) tempc2=sys_internal_types[RT_F64]; else if (tempi2->ic_flags & ICF_R_TO_INT) tempc2=sys_internal_types[RT_I64]; else tempc2=tempi->t.a2c; tempi->a2.type=TY_STK+CmpRawType(tempc2); tempi->a2.type_pointed=CmpRawTypePointed(tempc2); tempi->ic_flags|=ICF_A2_WAS_STK; case 1: tempi1=PrsPop(ps); if (tempi1->ic_flags&ICF_R_TO_F64) tempc1=sys_internal_types[RT_F64]; else if (tempi1->ic_flags & ICF_R_TO_INT) tempc1=sys_internal_types[RT_I64]; else tempc1=tempi->t.a1c; tempi->a1.type=TY_STK+CmpRawType(tempc1); tempi->a1.type_pointed=CmpRawTypePointed(tempc1); tempi->ic_flags|=ICF_A1_WAS_STK; break; case 0: //nobound_switch break; } if (cmp_result_cnt[code]) { tempi->r.type=TY_STK+CmpRawType(tempc); tempi->r.type_pointed=CmpRawTypePointed(tempc); tempi->ic_flags|=ICF_R_WAS_STK; PrsPush(ps,tempi); } nobound_switch (code) { case IC_IMM_F64: tempi->ic_flags&=~ICF_R_TO_F64; if (tempi->ic_flags&ICF_R_TO_INT) { tempi->ic_data=ToI64(tempi->ic_data(F64)); tempi->ic_flags&=~ICF_R_TO_INT; tempi->ic_code=IC_IMM_U64; } break; case IC_IMM_U64: tempi->ic_flags&=~ICF_R_TO_INT; if (tempi->ic_flags&ICF_R_TO_F64) { tempi->ic_data(F64)=ToF64(tempi->ic_data); tempi->ic_flags&=~ICF_R_TO_F64; tempi->ic_code=IC_IMM_F64; } break; case IC_RBP: tempi->ic_code=IC_REG; tempi->a1.reg=REG_RBP; break; case IC_DEREF: tempi->a1.type_pointed=tempi->r.type&RT_MASK; if (tempi1->ic_code==IC_LEA) { tempi->ic_flags|=tempi1->ic_flags; tempi->a1.reg=tempi1->a1.reg; tempi->a1.disp=tempi1->a1.disp; tempi->a1.type=TY_DISP+tempi->r.type&RT_MASK; tempi->a1.type_pointed=tempi->r.type_pointed; tempi->ic_code=IC_MOV; OptFree(tempi1); if (tempi->a1.reg==REG_RBP) for (i=0;ia1.disp) { mv[i].lea_balance--; mv[i].score++; break; } } else if (tempil1->ic_code==IC_ADD_CONST) { if (tempil2->ic_code==IC_REG) { tempi->ic_flags|=tempil2->ic_flags|tempil1->ic_flags; tempi->a1.reg=tempil2->a1.reg; tempi->a1.disp=tempi->ic_data; tempi->a1.type=TY_DISP+tempi->r.type&RT_MASK; tempi->a1.type_pointed=tempi->r.type_pointed; tempi->ic_code=IC_MOV; OptFree(tempil2); OptFree(tempil1); } } break; case IC__PP: case IC__MM: case IC_PP_: case IC_MM_: tempi->r.type_pointed=CmpRawTypePointed(tempi->t.a1c); case IC_DEREF_PP: case IC_DEREF_MM: if (tempi1->ic_code==IC_LEA && tempi1->a1.type&TY_MASK==TY_DISP && tempi1->a1.reg==REG_RBP) for (i=0;ia1.disp) { mv[i].lea_balance--; mv[i].score++; break; } tempi->a1.type_pointed=tempi->r.type&RT_MASK; break; case IC_MUL: case IC_DIV: if (tempc->raw_type==RT_F64) { CmpF2PushPop(tempi,tempi1,tempi2); break; } break; case IC_ADD: if (tempc->raw_type==RT_F64) { CmpF2PushPop(tempi,tempi1,tempi2); break; } if (OptFixupBinaryOp2(&tempi1,&tempi2)) { tempi->ic_flags|=tempi1->ic_flags; if (tempi->t.a1t!=tempi2) tempi->t.a1c=tempi->t.a2c; tempi->ic_data=tempi1->ic_data; tempi->ic_code=IC_ADD_CONST; tempi->a1.type_pointed=tempi->r.type&RT_MASK; tempi->a2.type=TY_NULL; OptFree(tempi1); if (tempil2->ic_code==IC_REG && tempil2->a1.reg==REG_RBP) { tempi->ic_flags|=tempil2->ic_flags; tempi->ic_code=IC_LEA; tempi->a1.reg=REG_RBP; tempi->a1.type=TY_DISP+tempi->a1.type&RT_MASK; tempi->a1.disp=tempi->ic_data; for (i=0;iic_dataraw_type==RT_F64) { CmpF2PushPop(tempi,tempi1,tempi2); break; } if (tempi2->ic_code==IC_IMM_U64) { tempi->ic_flags|=tempi2->ic_flags; tempi->ic_data=tempi2->ic_data; tempi->ic_code=IC_SUB_CONST; tempi->a2.type=TY_NULL; OptFree(tempi2); } break; case IC_LESS: case IC_GREATER: case IC_LESS_EQUAL: case IC_GREATER_EQUAL: if (tempi->ic_flags&ICF_USE_F64) CmpF2PushPop(tempi,tempi1,tempi2); break; case IC_MUL_EQUAL: case IC_DIV_EQUAL: case IC_ADD_EQUAL: case IC_SUB_EQUAL: if (tempc->raw_type==RT_F64) CmpF1PushPop(tempi,tempi2); case IC_ASSIGN_PP: case IC_ASSIGN_MM: case IC_ASSIGN: case IC_SHL_EQUAL: case IC_SHR_EQUAL: case IC_MOD_EQUAL: case IC_AND_EQUAL: case IC_OR_EQUAL: case IC_XOR_EQUAL: if (tempi1->ic_code==IC_LEA && tempi1->a1.type&TY_MASK==TY_DISP && tempi1->a1.reg==REG_RBP) for (i=0;ia1.disp) { mv[i].lea_balance--; mv[i].score++; break; } tempi->a1.type_pointed=tempi->r.type&RT_MASK; tempi->r.type_pointed=CmpRawTypePointed(tempi->t.a1c); break; case IC_RETURN_VAL: case IC_RETURN_VAL2: if (tempc) { if (tempc->raw_type==RT_F64 && tempil1->ic_class->raw_type!=RT_F64) tempil1->ic_flags|=ICF_R_TO_F64; else if (tempc->raw_type!=RT_F64 && tempil1->ic_class->raw_type==RT_F64) tempil1->ic_flags|=ICF_R_TO_INT; } break; case IC_SQR: case IC_ABS: case IC_SQRT: case IC_SIN: case IC_COS: case IC_TAN: case IC_ATAN: if (tempc->raw_type==RT_F64) CmpF1PushPop(tempi,tempi1); break; case IC_NOBOUND_SWITCH: case IC_SWITCH: lb=OptLabelFwd(tempi->ic_data(CLexCodeMisc *)->dft); lb->use_cnt++; break; case IC_BR_NOT_EQUAL: case IC_BR_EQUAL_EQUAL: if ((tempi1->ic_code==IC_IMM_U64 || tempi1->ic_code==IC_IMM_F64) && !tempi1->ic_data) { OptFree(tempi1); MemCpy(&tempi->a1,&tempi->a2,sizeof(CICArg)); tempi->a2.type=TY_NULL; if (code==IC_BR_EQUAL_EQUAL) code=tempi->ic_code=IC_BR_ZERO; else code=tempi->ic_code=IC_BR_NOT_ZERO; tempi1=tempi2; tempc1=tempc2; } else if ((tempi2->ic_code==IC_IMM_U64 || tempi2->ic_code==IC_IMM_F64) && !tempi2->ic_data) { OptFree(tempi2); tempi->a2.type=TY_NULL; if (code==IC_BR_EQUAL_EQUAL) code=tempi->ic_code=IC_BR_ZERO; else code=tempi->ic_code=IC_BR_NOT_ZERO; } else goto here1; case IC_BR_ZERO: case IC_BR_NOT_ZERO: if (tempi1->ic_code==IC_IMM_U64 || tempi1->ic_code==IC_IMM_F64) { if (code==IC_BR_ZERO ^^ tempi1->ic_data) { OptFree(tempi1); tempi->a1.type=TY_NULL; tempi->ic_code=IC_JMP; } else { OptFree(tempi1); tempi=OptFree(tempi); break; } } goto here1; case IC_BR_AND_ZERO: if (tempi1->ic_code==IC_IMM_U64) { i=Bsr(tempi1->ic_data); if (0<=i==Bsf(tempi1->ic_data)) { tempi1->ic_data=i; tempi->ic_flags|=ICF_NOT_ADDR; tempi->ic_code=IC_BR_NOT_BT; goto here1; } } if (tempi2->ic_code==IC_IMM_U64) { i=Bsr(tempi2->ic_data); if (0<=i==Bsf(tempi2->ic_data)) { tempi2->ic_data=i; tempi->ic_flags|=ICF_SWAP|ICF_NOT_ADDR; tempi->ic_code=IC_BR_NOT_BT; } } goto here1; case IC_BR_AND_NOT_ZERO: if (tempi1->ic_code==IC_IMM_U64) { i=Bsr(tempi1->ic_data); if (0<=i==Bsf(tempi1->ic_data)) { tempi1->ic_data=i; tempi->ic_flags|=ICF_NOT_ADDR; tempi->ic_code=IC_BR_BT; goto here1; } } if (tempi2->ic_code==IC_IMM_U64) { i=Bsr(tempi2->ic_data); if (0<=i==Bsf(tempi2->ic_data)) { tempi2->ic_data=i; tempi->ic_flags|=ICF_SWAP|ICF_NOT_ADDR; tempi->ic_code=IC_BR_BT; } } goto here1; case IC_BR_MM_ZERO: case IC_BR_MM_NOT_ZERO: if (tempi1->ic_code==IC_LEA && tempi1->a1.type&TY_MASK==TY_DISP && tempi1->a1.reg==REG_RBP) for (i=0;ia1.disp) { mv[i].lea_balance--; mv[i].score++; break; } tempi->a1.type_pointed=CmpRawType(tempc); goto here1; case IC_BR_GREATER_EQUAL: case IC_BR_LESS_EQUAL: case IC_BR_GREATER: case IC_BR_LESS: if (tempi->ic_flags&ICF_USE_F64) CmpF2PushPop(tempi,tempi1,tempi2); case IC_BR_CARRY: case IC_BR_NOT_CARRY: case IC_BR_GREATER_EQUAL2: case IC_BR_LESS_EQUAL2: case IC_BR_GREATER2: case IC_BR_LESS2: case IC_BR_NOT_EQUAL2: case IC_BR_EQUAL_EQUAL2: case IC_GET_LABEL: case IC_BR_BT: case IC_BR_BTS: case IC_BR_BTR: case IC_BR_BTC: case IC_BR_NOT_BT: case IC_BR_NOT_BTS: case IC_BR_NOT_BTR: case IC_BR_NOT_BTC: case IC_BR_LBTS: case IC_BR_LBTR: case IC_BR_LBTC: case IC_BR_NOT_LBTS: case IC_BR_NOT_LBTR: case IC_BR_NOT_LBTC: case IC_JMP: case IC_SUB_CALL: here1: if (tempi->ic_flags&ICF_PUSH_CMP) lb=tempi->ic_data; else lb=OptLabelFwd(tempi->ic_data); lb->use_cnt++; break; case IC_NOP1: tempi=OptFree(tempi); break; case IC_NOP2: ps->ptr+=tempi->ic_data; break; case IC_SHL_CONST: case IC_SHR_CONST: case IC_ENTER: case IC_ADD_RSP: case IC_ADD_RSP1: case IC_CALL: case IC_CALL_INDIRECT: case IC_CALL_INDIRECT2: case IC_CALL_EXTERN: case IC_CALL_IMPORT: case IC_PUSH: case IC_POP: case IC_INVLPG: case IC_CLFLUSH: case IC_GET_RFLAGS: case IC_CARRY: case IC_RDTSC: case IC_HPET: case IC_SET_RFLAGS: case IC_GET_RBP: case IC_SET_RBP: case IC_GET_RSP: case IC_GET_RAX: case IC_SET_RSP: case IC_SET_RAX: case IC_ABS_ADDR: case IC_HEAP_GLBL: case IC_ADDR_IMPORT: case IC_TYPE: case IC_BT: case IC_BTS: case IC_BTR: case IC_BTC: case IC_LBTS: case IC_LBTR: case IC_LBTC: case IC_BSF: case IC_BSR: case IC_PREEMPT: case IC_SIGN_I64: case IC_TOUPPER: case IC_TO_I64: case IC_TO_F64: case IC_TO_BOOL: case IC_ABS_I64: case IC_MIN_I64: case IC_MAX_I64: case IC_MIN_U64: case IC_MAX_U64: case IC_MOD_U64: case IC_SQR_I64: case IC_SQR_U64: case IC_SWAP_U8: case IC_SWAP_U16: case IC_SWAP_U32: case IC_SWAP_U64: case IC_QUE_INIT: case IC_QUE_INS: case IC_QUE_INS_REV: case IC_QUE_REM: case IC_IN_U32: case IC_IN_U16: case IC_IN_U8: case IC_STRLEN: case IC_OUT_U32: case IC_OUT_U16: case IC_OUT_U8: case IC_ASM: case IC_STR_CONST: case IC_FS: case IC_GS: case IC_MOV_FS: case IC_MOV_GS: case IC_RIP: case IC_PUSH_CMP: case IC_REG: case IC_COM: case IC_HOLYC_TYPECAST: case IC_NOT: case IC_UNARY_MINUS: case IC_POWER: case IC_SHL: case IC_SHR: case IC_MOD: case IC_AND: case IC_OR: case IC_XOR: case IC_EQUAL_EQUAL: case IC_NOT_EQUAL: case IC_AND_AND: case IC_OR_OR: case IC_XOR_XOR: case IC_LEAVE: case IC_RET: case IC_ADDR: case IC_END: case IC_END_EXP: case IC_CALL_END: case IC_CALL_END2: case IC_CALL_START: case IC_PUSH_REGS: case IC_POP_REGS: case IC_LABEL: break; default: "Pass:%d Missing IC handler\n",lx->pass; ICPut(tempi); LexExcept(lx,"Compiler Optimization Error at "); } } tempi=tempi_next; } if (lx->htc.fun) { lx->htc.fun->used_reg_mask=lx->htc.fun->used_reg_mask&~(REG_VARS_MASK|REG_NON_PTR_MASK)|used_reg_mask; if (!Bt(&lx->opts,OPTf_NO_REG_VAR) && !(lx->flags & LF_NO_REG_OPT)) { QSort(mv,member_cnt,sizeof(COptMemberVar),&OptMVCompare); while (member_cnt && !mv[member_cnt-1].score) member_cnt--; j=0; for (i=0;ireg==REG_ALLOC) PrintWarn("Can't reg var\n $$LK,\"FL:%s,%d\"$$ '%s' in '%s'\n", lx->cur_lfn->name,lx->cur_lfn->line_num,mv[i].m->str,lx->htc.fun->str); } } if (j>0) { if (Bt(&lx->flags,Lf_OPT_TRACE_PRESENT)) "Fun:%s\n",lx->htc.fun->str; if (j>num_reg_vars) { l=0; for (i=0;imember_class); if (!tempc->ptr_cnt && !tempm->dim.next) { while (lhtc.fun->used_reg_mask,to_non_ptr_vars_map[l])) l++; if (lreg=to_non_ptr_vars_map[l++]; reg_offsets[tempm->reg].offset=mv[i].offset_start; reg_offsets[tempm->reg].m=tempm; if (Bt(&lx->flags,Lf_OPT_TRACE_PRESENT)) "Reg %Z Var \"%s\" %08X[RBP]\n",tempm->reg,"ST_U64_REGS", tempm->str,reg_offsets[tempm->reg].offset; mv[i].offset_start=0; //flag as reg var if (tempm->size<8 && !StrIMatch("Bool",tempm->member_class->str) && tempm->member_class->type&HTT_INTERNAL_TYPE) PrintWarn("Using 64-bit reg var.\n " "$$LK,\"FL:%s,%d\"$$ '%s' in '%s'\n", lx->cur_lfn->name,lx->cur_lfn->line_num, tempm->str,lx->htc.fun->str); } } } } l=0; for (i=0;idim.next|| tempm->offset>0 && StrCmp(tempm->str,"argv"))) { while (lhtc.fun->used_reg_mask,to_reg_vars_map[l])) l++; if (lreg=to_reg_vars_map[l++]; reg_offsets[tempm->reg].offset=mv[i].offset_start; reg_offsets[tempm->reg].m=tempm; if (Bt(&lx->flags,Lf_OPT_TRACE_PRESENT)) "Reg %Z Var \"%s\" %08X[RBP]\n",tempm->reg,"ST_U64_REGS", tempm->str,reg_offsets[tempm->reg].offset; if (tempm->size<8 && !StrIMatch("Bool",tempm->member_class->str) && tempm->member_class->type&HTT_INTERNAL_TYPE) PrintWarn("Using 64-bit reg var.\n " "$$LK,\"FL:%s,%d\"$$ '%s' in '%s'\n", lx->cur_lfn->name,lx->cur_lfn->line_num, tempm->str,lx->htc.fun->str); } } } } } Free(mv); } if (ps->ptr>2) { "Pass:%d Stk:%08X\n",lx->pass,ps->ptr; LexExcept(lx,"Compiler Optimization Error at "); } }