#include "compiler/StackFrame.h"
#include "compiler/CError.h"
#include "compiler/CFunc.h"
#include "compiler/CMachine.h"
#include "compiler/CParser.h"
#include "compiler/CodeGen.h"
#include "compiler/CompilerTools.h"
#include "compiler/PCode.h"
#include "compiler/PCodeInfo.h"
#include "compiler/PCodeUtilities.h"
#include "compiler/RegisterInfo.h"
#include "compiler/objects.h"
#include "compiler/types.h"

#define ALIGN(thing, alignment) ( ~((alignment) - 1) & ((thing) + (alignment) - 1) )
#define ALIGN_REMAINDER(thing, alignment) ( ALIGN(thing, alignment) - (thing) )

Boolean requires_frame;
Boolean makes_call;
Boolean uses_globals;
Boolean dynamic_stack;
Boolean large_stack;
static SInt32 out_param_alignment;
static SInt32 in_parameter_size;
static SInt32 in_param_alignment;
static SInt32 frame_alignment;
static SInt32 alloca_alignment;
static SInt32 linkage_area_size;
static SInt32 parameter_area_size;
static SInt32 parameter_area_size_estimate;
static SInt32 local_data_size;
static SInt32 local_data_limit;
static SInt32 large_data_near_size;
static SInt32 large_data_far_size;
static ObjectList *local_objects[ObjClassMax];
static ObjectList *local_objects_tail[ObjClassMax];
static SInt32 non_volatile_save_offset[RegClassMax];
static SInt32 VRSAVE_save_offset;
static SInt32 LR_save_offset;
static SInt32 nonvolatile_save_size;
static UInt32 vrsave_mask;
static SInt32 frame_size;
static SInt32 frame_size_estimate;
static SInt32 genuine_frame_size;
static Object *dummylocal;
static Boolean dynamic_align_stack;
static Boolean has_varargs;
static Boolean compressing_data_area;
static PCode *setup_caller_sp;
static PCode *align_instr1;
static PCode *align_instr2;
static short vrsave_register;
static PCode *loadvrsave;
static PCode *storevrsave;
Object *dummyvaparam;
void *dummyprofiler;

// forward declarations
static void insert_local_object(UInt8 oclass, Object *obj);
static void compress_data_area(void);
static UInt32 align_bits(UInt32 value, UInt8 bitcount);
static Boolean use_helper_function(char rclass);
static Boolean need_link_register(void);
static void call_helper_function(char *name, char rclass, short effect);
static void save_nonvolatile_FPRs(int reg, SInt32 offset);
static void save_nonvolatile_VRs(int reg, SInt32 offset);
static void restore_nonvolatile_FPRs(int reg, SInt32 offset);
static void restore_nonvolatile_VRs(int reg, SInt32 offset);
static void save_nonvolatile_GPRs(int reg, SInt32 offset);
static void restore_nonvolatile_GPRs(int reg, SInt32 offset);
static void do_allocate_dynamic_stack_space(Boolean isConstantSize, int reg1, int reg2, SInt32 size);

void init_stack_globals(Object *funcobj) {
    char rclass;
    UInt8 oclass;

    requires_frame = 0;
    makes_call = 0;
    uses_globals = 0;
    dynamic_stack = 0;
    large_stack = 0;
    vrsave_register = -1;
    dynamic_align_stack = 0;
    compressing_data_area = 0;
    align_instr1 = NULL;
    align_instr2 = NULL;
    setup_caller_sp = NULL;
    dummyvaparam = NULL;
    loadvrsave = NULL;
    storevrsave = NULL;
    local_data_size = 0;
    local_data_limit = 0x2000;
    large_data_near_size = 0;
    large_data_far_size = 0;
    frame_size_estimate = 0;
    in_parameter_size = 0;
    parameter_area_size = 0;
    parameter_area_size_estimate = 0;
    frame_alignment = 8;
    out_param_alignment = 8;
    in_param_alignment = 8;
    alloca_alignment = 0;
    has_varargs = 0;
    linkage_area_size = 0;
    frame_size = 0;
    genuine_frame_size = 0;
    nonvolatile_save_size = -1;
    VRSAVE_save_offset = -1;
    LR_save_offset = -1;

    for (rclass = 0; rclass < RegClassMax; rclass++)
        non_volatile_save_offset[(char) rclass] = -1;

    dummyprofiler = NULL;
    dummyvaparam = NULL;
    dummylocal = NULL;

    for (oclass = 0; oclass < ObjClassMax; oclass++) {
        local_objects[oclass] = NULL;
        local_objects_tail[oclass] = NULL;
    }
}

void init_frame_sizes(Boolean has_varargs) {
    ObjectList *scan;
    Object *obj;
    SInt32 r30;
    SInt32 align;
    SInt32 mask;

    r30 = in_parameter_size + parameter_area_size_estimate;
    frame_size_estimate = r30 + 2484;
    for (scan = locals; scan; scan = scan->next) {
        obj = scan->object;
        {
            align = CMach_AllocationAlignment(obj->type, obj->qual) - 1;
            mask = ~align;
        }
        frame_size_estimate = (frame_size_estimate + align) & mask;
        frame_size_estimate += obj->type->size;
    }

    if (frame_size_estimate > 0x8000) {
        dynamic_stack = 1;
        large_stack = 1;
        requires_frame = 1;
    }

    local_data_limit = 0x8000 - (r30 + 2484);

    if (dynamic_stack) {
        requires_frame = 1;
        dummylocal = galloc(sizeof(Object));
        memclrw(dummylocal, sizeof(Object));
        dummylocal->type = (Type *) &stvoid;
        dummylocal->otype = OT_OBJECT;
        dummylocal->name = GetHashNameNode("<dummy>");
        dummylocal->datatype = DLOCAL;
        dummylocal->u.var.info = CodeGen_GetNewVarInfo();
        dummylocal->u.var.info->flags |= VarInfoFlag80;
        dummylocal->u.var.info->noregister = 1;
    }

    if (dynamic_stack) {
        retain_register(NULL, RegClass_GPR, 31);
        _FP_ = 31;
    } else {
        _FP_ = 1;
    }
}

void assign_local_memory(Object *obj) {
    // some misassigned registers x.x
    short align;
    VarInfo *vi;

    align = CMach_AllocationAlignment(obj->type, obj->qual);
    if (!compressing_data_area && (obj->u.var.info->flags & VarInfoFlag80))
        return;

    update_frame_align(align);
    if (local_data_size + (ALIGN_REMAINDER(local_data_size, align) + ALIGN(obj->type->size, align)) < local_data_limit) {
        local_data_size = ALIGN(local_data_size, align);
        vi = Registers_GetVarInfo(obj);
        vi->flags &= ~VarInfoFlag2;
        vi->flags |= VarInfoFlag80;
        obj->u.var.uid = local_data_size;
        local_data_size += ALIGN(obj->type->size, align);
        insert_local_object(ObjClass0, obj);
        return;
    }
    if (compressing_data_area || obj->type->size <= 32) {
        large_data_near_size = ALIGN(large_data_near_size, align);
        vi = Registers_GetVarInfo(obj);
        vi->flags &= ~VarInfoFlag2;
        vi->flags |= VarInfoFlag80;
        obj->u.var.uid = 0x8000 + large_data_near_size;
        large_data_near_size += ALIGN(obj->type->size, align);
        insert_local_object(ObjClass1, obj);
    } else {
        large_data_far_size = ALIGN(large_data_far_size, align);
        vi = Registers_GetVarInfo(obj);
        vi->flags &= ~VarInfoFlag2;
        vi->flags |= VarInfoFlag80;
        obj->u.var.uid = 0x10000 + large_data_far_size;
        large_data_far_size += ALIGN(obj->type->size, align);
        insert_local_object(ObjClass2, obj);
    }
}

void assign_locals_to_memory(ObjectList *first) {
    ObjectList *list;
    Object *obj;
    SInt32 i;

    for (i = 1; i < 1024; i <<= 1) {
        for (list = first; list; list = list->next) {
            obj = list->object;
            if (Registers_GetVarInfo(obj)->used) {
                if ((Registers_GetVarInfo(obj) ? Registers_GetVarInfo(obj)->reg : 0) == 0) {
                    if (obj->type->size <= i)
                        assign_local_memory(obj);
                }
            }
        }
    }

    for (list = first; list; list = list->next) {
        obj = list->object;
        if (Registers_GetVarInfo(obj)->used) {
            if ((Registers_GetVarInfo(obj) ? Registers_GetVarInfo(obj)->reg : 0) == 0) {
                assign_local_memory(list->object);
            }
        }

        if (obj->type && IS_TYPE_ARRAY(obj->type) && IS_TYPE_VECTOR(TYPE_POINTER(obj->type)->target))
            has_altivec_arrays = 1;
    }
}

void compute_frame_sizes(void) {
    SInt32 altivec_size;
    SInt32 altivec_offset;

    CError_ASSERT(897, alloca_alignment == 0 || alloca_alignment == frame_alignment);

    update_asm_nonvolatile_registers();
    LR_save_offset = 8;
    non_volatile_save_offset[RegClass_FPR] = -(used_nonvolatile_registers[RegClass_FPR] * 8);
    non_volatile_save_offset[RegClass_GPR] = -(((15 - non_volatile_save_offset[RegClass_FPR]) & ~15) + used_nonvolatile_registers[RegClass_GPR] * 4);
    nonvolatile_save_size = -non_volatile_save_offset[RegClass_GPR];
    non_volatile_save_offset[RegClass_CRFIELD] = 4;
    VRSAVE_save_offset = -1;
    non_volatile_save_offset[RegClass_VR] = -1;

    if (copts.altivec_model) {
        if (vrsave_mask) {
            VRSAVE_save_offset = non_volatile_save_offset[RegClass_GPR] - 4;
            nonvolatile_save_size = nonvolatile_save_size + 4;
        }
        altivec_size = used_nonvolatile_registers[RegClass_VR] * 16;
        if (altivec_size > 0)
            nonvolatile_save_size = ALIGN(nonvolatile_save_size + altivec_size, frame_alignment);
    }

    if (parameter_area_size)
        requires_frame = 1;

    compress_data_area();
    local_data_size = ALIGN(local_data_size, frame_alignment);
    nonvolatile_save_size = ALIGN(nonvolatile_save_size, frame_alignment);
    if (!requires_frame && (local_data_size + nonvolatile_save_size) <= 224) {
        CError_ASSERT(1005, !dynamic_align_stack);
        linkage_area_size = 0;
        frame_size = 0;
        genuine_frame_size = local_data_size + nonvolatile_save_size;
    } else {
        requires_frame = 1;
        if (parameter_area_size < 32)
            parameter_area_size = 32;
        parameter_area_size = ALIGN(parameter_area_size + 24, frame_alignment) - 24;
        if (large_stack) {
            CError_ASSERT(1019, !large_data_far_size);
            large_data_near_size += parameter_area_size;
            parameter_area_size = 0;
        }
        linkage_area_size = 24;
        frame_size = nonvolatile_save_size + (altivec_offset = parameter_area_size + 24 + local_data_size);
        if (copts.altivec_model && used_nonvolatile_registers[RegClass_VR])
            non_volatile_save_offset[RegClass_VR] = altivec_offset;
        frame_size += ALIGN_REMAINDER(frame_size, 16);
        frame_size = ALIGN(frame_size, frame_alignment);
        genuine_frame_size = frame_size;
    }
    if (!large_stack && frame_size > 0x7FFF)
        CError_ErrorTerm(210);
}

static void allocate_new_frame(int reg1, int reg2) {
    if (dynamic_align_stack) {
        CError_ASSERT(1116, reg1 != _CALLER_SP_);
        emitpcode(PC_RLWINM, reg1, 1, 0, align_bits(frame_alignment, 1), 31);
        if (frame_size > 0x7FFF) {
            CError_FATAL(1122);
            return;
        }

        if (frame_size)
            emitpcode(PC_SUBFIC, reg1, reg1, -frame_size);
        else
            emitpcode(PC_SUBFIC, reg1, reg1, -genuine_frame_size);

        if (reg2)
            emitpcode(PC_MR, reg2, 1);

        emitpcode(PC_STWUX, 1, 1, reg1);
    } else {
        if (frame_size > 0x7FFF)
            CError_FATAL(1153);
        else
            emitpcode(PC_STWU, 1, 1, 0, -frame_size);

        if (reg2)
            emitpcode(PC_MR, reg2, 1);
    }
}

void generate_prologue(PCodeBlock *block, Boolean has_varargs) {
    PCodeBlock *save_block;
    Boolean needs_lr;
    Statement *save_statement;
    Statement stmt;
    UInt32 vrsave_low;
    UInt32 vrsave_high;

    save_block = pclastblock;
    needs_lr = need_link_register();
    save_statement = current_statement;
    stmt.sourceoffset = functionbodyoffset;
    current_statement = &stmt;
    pclastblock = block;

    if (setup_caller_sp && setup_caller_sp->block) {
        if (
                setup_caller_sp->op == PC_MR &&
                setup_caller_sp->args[1].kind == PCOp_REGISTER &&
                setup_caller_sp->args[1].arg == RegClass_GPR &&
                setup_caller_sp->args[1].data.reg.reg == _FP_
                )
            CError_FATAL(1197);

        _CALLER_SP_ = setup_caller_sp->args[0].data.reg.reg;
        deletepcode(setup_caller_sp);
        setup_caller_sp = NULL;
    } else if (_CALLER_SP_ != _FP_) {
        _CALLER_SP_ = -1;
    }

    if (align_instr1 && align_instr1->block) {
        deletepcode(align_instr1);
        align_instr1 = NULL;
    }

    if (align_instr2 && align_instr2->block) {
        deletepcode(align_instr2);
        align_instr2 = NULL;
    }

    if (loadvrsave && loadvrsave->block) {
        deletepcode(loadvrsave);
        loadvrsave = NULL;
    }

    if (storevrsave && storevrsave->block) {
        deletepcode(storevrsave);
        storevrsave = NULL;
    }

    if (needs_lr)
        emitpcode(PC_MFLR, 0);

    if (used_nonvolatile_registers[RegClass_CRFIELD]) {
        emitpcode(PC_MFCR, 12);
        emitpcode(PC_STW, 12, 1, 0, non_volatile_save_offset[RegClass_CRFIELD]);
    }

    if (used_nonvolatile_registers[RegClass_FPR])
        save_nonvolatile_FPRs(1, 0);
    if (used_nonvolatile_registers[RegClass_GPR])
        save_nonvolatile_GPRs(1, 0);
    if (needs_lr)
        emitpcode(PC_STW, 0, 1, 0, 8);

    if (frame_size) {
        if (vrsave_mask) {
            emitpcode(PC_MFSPR, 0, 256);
            emitpcode(PC_STW, 0, 1, 0, VRSAVE_save_offset);
            vrsave_register = 0;
        }
        allocate_new_frame(12, (_CALLER_SP_ > 0 && _CALLER_SP_ != _FP_) ? _CALLER_SP_ : 0);
    } else {
        CError_ASSERT(1326, !dynamic_align_stack);
        if (vrsave_mask)
            emitpcode(PC_MFSPR, vrsave_register, 256);
    }

    if (vrsave_mask) {
        vrsave_high = vrsave_mask >> 16;
        vrsave_low = vrsave_mask & 0xFFFF;
        if (vrsave_mask == 0xFFFFFFFF) {
            emitpcode(PC_LI, 0, -1);
        } else {
            if (vrsave_high)
                emitpcode(PC_ORIS, 0, vrsave_register, vrsave_high);
            if (vrsave_low)
                emitpcode(PC_ORI, 0, 0, vrsave_low);
        }
        emitpcode(PC_MTSPR, 256, 0);
    }

    if (used_nonvolatile_registers[RegClass_VR])
        save_nonvolatile_VRs(1, 0);

    if (dynamic_stack)
        emitpcode(PC_MR, 31, 1);

    if (large_stack)
        do_allocate_dynamic_stack_space(1, 11, 0, large_data_near_size);

    block->flags |= fPCBlockFlag1;
    pclastblock = save_block;
    current_statement = save_statement;
}

void generate_epilogue(PCodeBlock *block, Boolean add_blr) {
    PCodeBlock *save_block;
    Boolean needs_lr;
    Statement *save_statement;
    Statement stmt;

    save_block = pclastblock;
    needs_lr = need_link_register();
    save_statement = current_statement;
    if (!save_statement) {
        stmt.sourceoffset = current_linenumber;
        current_statement = &stmt;
    }
    pclastblock = block;

    if (used_nonvolatile_registers[RegClass_VR])
        restore_nonvolatile_VRs(_FP_, 0);

    if (dynamic_align_stack) {
        load_store_register(PC_LWZ, 1, 1, NULL, 0);
        setpcodeflags(fSideEffects);
        if (needs_lr)
            load_store_register(PC_LWZ, 0, 1, 0, 8);
    } else {
        if (needs_lr)
            load_store_register(PC_LWZ, 0, _FP_, 0, frame_size + 8);
        if (frame_size > 0) {
            if (dynamic_stack) {
                load_store_register(PC_LWZ, 1, 1, 0, 0);
                setpcodeflags(fSideEffects);
            } else {
                emitpcode(PC_ADDI, 1, 1, 0, frame_size);
                setpcodeflags(fSideEffects);
            }
        }
    }

    if (used_nonvolatile_registers[RegClass_CRFIELD]) {
        load_store_register(PC_LWZ, 12, 1, NULL, non_volatile_save_offset[RegClass_CRFIELD]);
        emitpcode(PC_MTCRF, 255, 12);
    }

    if (vrsave_mask) {
        if (!requires_frame) {
            emitpcode(PC_MTSPR, 256, vrsave_register);
        } else {
            emitpcode(PC_LWZ, 11, 1, 0, VRSAVE_save_offset);
            emitpcode(PC_MTSPR, 256, 11);
        }
    }

    if (used_nonvolatile_registers[RegClass_FPR])
        restore_nonvolatile_FPRs(1, 0);
    if (needs_lr && !use_helper_function(RegClass_GPR))
        emitpcode(PC_MTLR, 0);

    if (used_nonvolatile_registers[RegClass_GPR])
        restore_nonvolatile_GPRs(1, 0);
    if (needs_lr && use_helper_function(RegClass_GPR))
        emitpcode(PC_MTLR, 0);

    if (add_blr) {
        emitpcode(PC_BLR);
        setpcodeflags(fIsVolatile);
    }

    block->flags |= fPCBlockFlag2;
    pclastblock = save_block;
    current_statement = save_statement;
}

static void load_base_offset(int dest_reg, int base_reg, SInt32 offset) {
    if (offset)
        emitpcode(PC_ADDI, dest_reg, base_reg, 0, offset);
    else
        emitpcode(PC_MR, dest_reg, base_reg);
}

static void save_nonvolatile_FPRs(int reg, SInt32 offset) {
    short i;
    SInt32 o;

    o = offset + non_volatile_save_offset[RegClass_FPR];

    if (!use_helper_function(RegClass_FPR)) {
        for (i = 1; i <= used_nonvolatile_registers[RegClass_FPR]; i++) {
            emitpcode(PC_STFD, 32 - i, reg, NULL, o + (used_nonvolatile_registers[RegClass_FPR] - i) * 8);
            setpcodeflags(fIsVolatile);
        }
    } else {
        load_base_offset(11, reg, o + used_nonvolatile_registers[RegClass_FPR] * 8);
        call_helper_function("__save_fpr_%d", RegClass_FPR, EffectRead);
    }
}

static void save_nonvolatile_VRs(int reg, SInt32 offset) {
    short i;
    SInt32 o;

    o = offset + non_volatile_save_offset[RegClass_VR];

    if (!use_helper_function(RegClass_VR)) {
        for (i = 1; i <= used_nonvolatile_registers[RegClass_VR]; i++) {
            emitpcode(PC_LI, 0, o + (used_nonvolatile_registers[RegClass_VR] - i) * 16);
            emitpcode(PC_STVX, 32 - i, reg, 0);
            setpcodeflags(fIsVolatile);
        }
    } else {
        load_base_offset(0, reg, o + used_nonvolatile_registers[RegClass_VR] * 16);
        call_helper_function("__savev%d", RegClass_VR, EffectRead);
    }
}

static void restore_nonvolatile_FPRs(int reg, SInt32 offset) {
    short i;
    SInt32 o;

    o = offset + non_volatile_save_offset[RegClass_FPR];

    if (!use_helper_function(RegClass_FPR)) {
        for (i = 1; i <= used_nonvolatile_registers[RegClass_FPR]; i++) {
            load_store_register(PC_LFD, 32 - i, reg, NULL, o + (used_nonvolatile_registers[RegClass_FPR] - i) * 8);
            setpcodeflags(fIsVolatile);
        }
    } else {
        load_base_offset(11, reg, o + used_nonvolatile_registers[RegClass_FPR] * 8);
        call_helper_function("__restore_fpr_%d", RegClass_FPR, EffectWrite);
    }
}

static void restore_nonvolatile_VRs(int reg, SInt32 offset) {
    short i;
    SInt32 o;

    o = offset + non_volatile_save_offset[RegClass_VR];

    if (!use_helper_function(RegClass_VR)) {
        for (i = 1; i <= used_nonvolatile_registers[RegClass_VR]; i++) {
            emitpcode(PC_LI, 0, o + (used_nonvolatile_registers[RegClass_VR] - i) * 16);
            setpcodeflags(fIsVolatile);
            emitpcode(PC_LVX, 32 - i, reg, 0);
            setpcodeflags(fIsVolatile);
        }
    } else {
        load_base_offset(0, reg, o + used_nonvolatile_registers[RegClass_VR] * 16);
        call_helper_function("__restv%d", RegClass_VR, EffectWrite);
    }
}

static void save_nonvolatile_GPRs(int reg, SInt32 offset) {
    int i;
    SInt32 o;

    o = offset + non_volatile_save_offset[RegClass_GPR];

    if (!use_helper_function(RegClass_GPR)) {
        if (copts.use_lmw_stmw && ((used_nonvolatile_registers[RegClass_GPR] > 4) || (copts.optimize_for_size && (used_nonvolatile_registers[RegClass_GPR] > 1)))) {
            emitpcode(PC_STMW, used_nonvolatile_registers[RegClass_GPR] - 1, 32 - used_nonvolatile_registers[RegClass_GPR], reg, 0, o);
        } else {
            for (i = 1; i <= used_nonvolatile_registers[RegClass_GPR]; i++) {
                emitpcode(PC_STW, 32 - i, reg, 0, o + (used_nonvolatile_registers[RegClass_GPR] - i) * 4);
            }
        }
    } else {
        load_base_offset(11, reg, o + used_nonvolatile_registers[RegClass_GPR] * 4);
        call_helper_function("__savegpr_%d", RegClass_GPR, EffectRead);
    }
}

static void restore_nonvolatile_GPRs(int reg, SInt32 offset) {
    int i;
    SInt32 o;

    o = offset + non_volatile_save_offset[RegClass_GPR];

    if (!use_helper_function(RegClass_GPR)) {
        if (copts.use_lmw_stmw && ((used_nonvolatile_registers[RegClass_GPR] > 4) || (copts.optimize_for_size && (used_nonvolatile_registers[RegClass_GPR] > 1)))) {
            emitpcode(PC_LMW, used_nonvolatile_registers[RegClass_GPR] - 1, 32 - used_nonvolatile_registers[RegClass_GPR], reg, 0, o);
            setpcodeflags(fIsVolatile);
        } else {
            for (i = 1; i <= used_nonvolatile_registers[RegClass_GPR]; i++) {
                emitpcode(PC_LWZ, 32 - i, reg, 0, o + (used_nonvolatile_registers[RegClass_GPR] - i) * 4);
                setpcodeflags(fIsVolatile);
            }
        }
    } else {
        load_base_offset(11, reg, o + used_nonvolatile_registers[RegClass_GPR] * 4);
        call_helper_function("__restgpr_%d", RegClass_GPR, EffectWrite);
    }
}

static void do_allocate_dynamic_stack_space(Boolean isConstantSize, int reg1, int reg2, SInt32 size) {
    load_store_register(PC_LWZ, reg2, 1, NULL, 0);
    if (isConstantSize) {
        size = ALIGN(size, frame_alignment);
        if (size < 0x8000) {
            emitpcode(PC_STWU, reg2, 1, 0, -size);
        } else {
            emitpcode(PC_LIS, reg1, 0, (short) HIGH_PART(-size));
            if (-size)
                emitpcode(PC_ADDI, reg1, reg1, 0, LOW_PART(-size));
            emitpcode(PC_STWUX, reg2, 1, reg1);
            setpcodeflags(fIsVolatile | fSideEffects);
        }
    } else {
        emitpcode(PC_STWUX, reg2, 1, reg1);
        setpcodeflags(fIsVolatile | fSideEffects);
    }
}

void allocate_dynamic_stack_space(Boolean isConstantSize, int reg1, int reg2, SInt32 size) {
    if (copts.altivec_model)
        update_frame_align(16);
    do_allocate_dynamic_stack_space(isConstantSize, reg1, reg2, size);
    add_immediate(reg1, 1, dummylocal, 0);
}

#ifdef __MWERKS__
#pragma options align=mac68k
#endif
typedef struct Traceback {
    UInt8 x0;
    UInt8 x1;
    UInt8 x2;
    UInt8 x3;
    UInt8 x4;
    UInt8 x5;

    UInt8 x6_0 : 2;
    UInt8 x6_1 : 1; // set to 1
    UInt8 x6_2 : 5;

    UInt8 x7_0 : 1;
    UInt8 x7_1 : 1; // set to 1
    UInt8 has_dynamic_stack : 1; // set to 1 if dynamic_stack
    UInt8 x7_3 : 3;
    UInt8 uses_CRs : 1; // set to 1 if CRs used
    UInt8 needs_link_register : 1; // set to 1 if link register used

    UInt8 has_frame_size : 1; // set to 1 if frame_size is nonzero
    UInt8 x8_1 : 1; // set to 0
    UInt8 used_FPRs : 6; // stores non-volatile FPRs used

    UInt8 x9_0 : 1; // set to 0
    UInt8 x9_1 : 1; // set to 1 if VRs or vrsave used
    UInt8 used_GPRs : 6; // stores non-volatile GPRs used

    UInt8 xA;
    UInt8 xB;

    SInt32 funcsize;
    SInt16 namelen;
    char name[0];
} Traceback;

typedef struct TracebackExtra {
    UInt8 used_VRs : 6;
    UInt8 has_vrsave_mask : 1;
    UInt8 is_varargs : 1;
    UInt8 vec_arg_count : 7;
    UInt8 has_vrsave_mask_or_used_VRs : 1;
} TracebackExtra;
#ifdef __MWERKS__
#pragma options align=reset
#endif

char *generate_traceback(SInt32 funcsize, char *funcname, SInt32 *tbsize, Object *func) {
    char *work;
    short namelen;
    Traceback *buf;
    SInt32 bufsize;

    namelen = strlen(funcname);
    bufsize = ALIGN(sizeof(Traceback) + namelen + (dynamic_stack ? 1 : 0) + ((used_nonvolatile_registers[RegClass_VR] || vrsave_mask) ? sizeof(TracebackExtra) : 0), 4);
    buf = lalloc(bufsize);
    memclrw(buf, bufsize);

    buf->x4 = 0;
    buf->x5 = copts.cplusplus ? 9 : 0;
    buf->x6_1 = 1;
    buf->x7_1 = 1;
    if (dynamic_stack)
        buf->has_dynamic_stack = 1;
    if (used_nonvolatile_registers[RegClass_CRFIELD])
        buf->uses_CRs = 1;
    if (need_link_register())
        buf->needs_link_register = 1;
    if (frame_size)
        buf->has_frame_size = 1;
    buf->used_FPRs = used_nonvolatile_registers[RegClass_FPR];
    buf->used_GPRs = used_nonvolatile_registers[RegClass_GPR];
    buf->x8_1 = 0;
    buf->x9_0 = 0;
    buf->x9_1 = (used_nonvolatile_registers[RegClass_VR] || vrsave_mask) != 0;
    buf->funcsize = funcsize;
    buf->namelen = namelen;

    work = buf->name;
    strcpy(work, funcname);
    work += namelen;
    if (dynamic_stack) {
        *(work++) = 31;
    }

    if (vrsave_mask || used_nonvolatile_registers[RegClass_VR]) {
        TracebackExtra *extra;
        Boolean is_varargs;
        int vec_count;
        FuncArg *args, *scan;
        Type *type;

        extra = (TracebackExtra *) work;
        vec_count = 0;
        args = TYPE_FUNC(func->type)->args;
        scan = args;
        while (scan && scan != &elipsis)
            scan = scan->next;
        is_varargs = scan == &elipsis;
        while (args) {
            if ((type = args->type) && IS_TYPE_VECTOR(type))
                vec_count++;
            args = args->next;
        }
        extra->used_VRs = used_nonvolatile_registers[RegClass_VR];
        extra->has_vrsave_mask = vrsave_mask != 0;
        extra->is_varargs = is_varargs;
        extra->vec_arg_count = vec_count;
        extra->has_vrsave_mask_or_used_VRs = vrsave_mask || used_nonvolatile_registers[RegClass_VR];
    }

    *tbsize = bufsize;
    return (char *) buf;
}

static SInt32 localsbase(void) {
    SInt32 size = parameter_area_size;
    if (frame_size || dynamic_align_stack)
        size += linkage_area_size;
    else
        size -= genuine_frame_size;
    return size;
}

static SInt32 parametersbase(int flag) {
    if (flag)
        return 24;

    return frame_size ? (genuine_frame_size + 24) : 24;
}

void check_dynamic_aligned_frame(void) {
    PCode *pc;

    if (used_nonvolatile_registers[RegClass_VR]) {
        update_frame_align(16);
        requires_frame = 1;
    }

    if (frame_alignment > in_param_alignment) {
        dynamic_align_stack = 1;
        requires_frame = 1;
        CError_ASSERT(2091, !has_varargs || _CALLER_SP_ != -1);
        CError_ASSERT(2096, _CALLER_SP_ != _FP_);
        if (setup_caller_sp && setup_caller_sp->block) {
            align_instr1 = makepcode(PC_RLWINM, 12, 1, 0, 5, 31);
            insertpcodebefore(setup_caller_sp, align_instr1);
            align_instr2 = makepcode(PC_STWUX, 1, 1, 12);
            insertpcodeafter(setup_caller_sp, align_instr2);
        }
    } else {
        dynamic_align_stack = 0;
        if (setup_caller_sp && setup_caller_sp->block) {
            pc = makepcode(PC_MR, _CALLER_SP_, _FP_);
            insertpcodebefore(setup_caller_sp, pc);
            deletepcode(setup_caller_sp);
            setup_caller_sp = pc;
        }
        _CALLER_SP_ = _FP_;
    }

    vrsave_mask = 0;
    if (copts.altivec_model) {
        vrsave_mask = colored_vrs_as_vrsave(pcbasicblocks);
        if (!requires_frame && vrsave_mask) {
            vrsave_register = 11;
            loadvrsave = makepcode(PC_LWZ, 11, 1, 0, -4);
            appendpcode(prologue, loadvrsave);
            storevrsave = makepcode(PC_STW, 11, 1, 0, -4);
            appendpcode(epilogue, storevrsave);
        }
    }
}

void move_varargs_to_memory(void) {
    short reg;

    has_varargs = 1;
    dummyvaparam = galloc(sizeof(Object));
    memclrw(dummyvaparam, sizeof(Object));

    dummyvaparam->type = (Type *) &stvoid;
    dummyvaparam->otype = OT_OBJECT;
    dummyvaparam->name = GetHashNameNode("<vaparam>");
    dummyvaparam->datatype = DLOCAL;
    dummyvaparam->u.var.info = CodeGen_GetNewVarInfo();
    dummyvaparam->u.var.uid = 0;
    dummyvaparam->u.var.info->noregister = 1;
    Registers_GetVarInfo(dummyvaparam)->flags = (Registers_GetVarInfo(dummyvaparam)->flags & ~VarInfoFlag1) | VarInfoFlag1;

    for (reg = last_argument_register[RegClass_GPR] + 1; (int)reg <= 10; reg++) {
        emitpcode(PC_STW, reg, local_base_register(dummyvaparam), dummyvaparam, (reg - 3) * 4);
        setpcodeflags(fPCodeFlag20 | fPCodeFlag8000);
    }
}

void assign_arguments_to_memory(Object *func, UInt8 mysteryFlag, Boolean hasVarargs) {
    // almost matches except for the not/andc issue
    SInt32 pos;
    ObjectList *list;
    Object *obj;
    Type *type;
    short reg;
    SInt32 chk;
    Boolean flag;

    pos = 0;
    reg = 2;

    for (list = arguments; list; list = list->next) {
        obj = list->object;
        type = obj->type;
        if (!IS_TYPE_VECTOR(type)) {
            obj->datatype = DLOCAL;
            obj->u.var.info = CodeGen_GetNewVarInfo();
            if (IS_TYPE_ARRAY(type) || IS_TYPE_NONVECTOR_STRUCT(type) || IS_TYPE_CLASS(type) ||
                    IS_TYPE_12BYTES_MEMBERPOINTER(type)) {
                chk = CMach_ArgumentAlignment(type);
                if (chk > 4) {
                    pos = ALIGN(pos, chk);
                    update_in_param_align(chk);
                }
            }
            obj->u.var.uid = pos;
            Registers_GetVarInfo(obj)->flags = (Registers_GetVarInfo(obj)->flags & ~VarInfoFlag1) | VarInfoFlag1;
            if (!copts.little_endian && (IS_TYPE_INT(obj->type) || IS_TYPE_ENUM(obj->type)) && obj->type->size < 4)
                obj->u.var.uid += 4 - obj->type->size;
            pos += type->size;
            pos = ALIGN(pos, 4);
        } else {
            obj->u.var.info = CodeGen_GetNewVarInfo();
            obj->u.var.uid = 0;
            obj->datatype = DLOCAL;
            flag = 1;
            if (reg <= 13)
                flag = hasVarargs;
            if (flag) {
                pos = ALIGN(pos + 24, 16) - 24;
                obj->u.var.uid = pos;
                pos += 16;
                update_in_param_align(16);
                Registers_GetVarInfo(obj)->flags = (Registers_GetVarInfo(obj)->flags & ~VarInfoFlag1) | VarInfoFlag1;
            } else {
                assign_local_memory(obj);
                Registers_GetVarInfo(obj)->flags = Registers_GetVarInfo(obj)->flags & ~VarInfoFlag1;
            }
            reg++;
        }
    }

    in_parameter_size = (in_parameter_size < pos) ? pos : in_parameter_size;
    CError_ASSERT(2408, !dummyvaparam);
}

SInt32 set_out_param_displ(SInt32 a, Type *type, Boolean flag, SInt32 *outvar, SInt32 b) {
    // does not match due to errant andc
    SInt32 argAlign;

    if (!flag && !b) {
        *outvar = 0;
        return a;
    }

    if (IS_TYPE_VECTOR(type)) {
        update_out_param_align(16);
        a = ALIGN(a + 16 + 24, 16) - 24;
    } else if (IS_TYPE_ARRAY(type) || IS_TYPE_NONVECTOR_STRUCT(type) || IS_TYPE_CLASS(type) || IS_TYPE_12BYTES_MEMBERPOINTER(type)) {
        argAlign = CMach_ArgumentAlignment(type);
        if (argAlign > 4) {
            a = ALIGN(a + 24, argAlign) - 24;
            update_in_param_align(argAlign);
        }
    }

    *outvar = a;
    a = ALIGN(a + b, 4);
    return a;
}

SInt32 out_param_displ_to_offset(SInt32 displ) {
    return displ + 24;
}

Boolean needs_frame(void) {
    return (frame_size > 224) || requires_frame;
}

void update_out_param_size(SInt32 size) {
    if (size < 32)
        size = 32;
    if (parameter_area_size < size)
        parameter_area_size = size;
}

void estimate_out_param_size(SInt32 size) {
    if (parameter_area_size_estimate < size)
        parameter_area_size_estimate = size;
}

void update_out_param_align(SInt32 align) {
    if (out_param_alignment < align)
        out_param_alignment = align;
    update_frame_align(align);
}

void update_in_param_align(SInt32 align) {
    if (in_param_alignment < align)
        in_param_alignment = align;
}

void update_frame_align(SInt32 align) {
    if (frame_alignment < align)
        frame_alignment = align;
}

SInt32 local_offset_32(Object *obj) {
    short align;
    SInt32 offset;

    if (obj->u.var.info->flags & VarInfoFlag1)
        align = CMach_ArgumentAlignment(obj->type);
    else
        align = CMach_AllocationAlignment(obj->type, obj->qual);

    offset = obj->u.var.uid;
    if (offset > 0x7FFF)
        offset = 0x8000 - offset - ALIGN(obj->type->size, align);

    if (obj->u.var.info->flags & VarInfoFlag1)
        return offset + parametersbase(local_base_register(obj) != _FP_);
    else
        return offset + localsbase();
}

SInt32 local_offset_lo(Object *obj, SInt32 offset) {
    SInt32 combo = offset + local_offset_32(obj);
    return LOW_PART(combo);
    //return (SInt16) (offset + local_offset_32(obj));
}

SInt32 local_offset_ha(Object *obj, SInt32 offset) {
    SInt32 combo = offset + local_offset_32(obj);
    return HIGH_PART(combo);
    //return (SInt16) ((combo >> 16) + ((combo & 0x8000) >> 15));
}

SInt32 local_offset_16(Object *obj) {
    SInt32 offset32 = local_offset_32(obj);
    SInt16 offset16 = (SInt16) offset32;
    CError_ASSERT(2662, offset32 == offset16);
    return offset16;
}

Boolean local_is_16bit_offset(Object *obj) {
    SInt32 offset32 = local_offset_32(obj);
    SInt16 offset16 = (SInt16) offset32;
    return offset32 == offset16;
}

int local_base_register(Object *obj) {
    PCode *pc;

    if (obj->u.var.info->flags & VarInfoFlag1) {
        if (coloring && _CALLER_SP_ == -1) {
            _CALLER_SP_ = used_virtual_registers[RegClass_GPR]++;
            pc = makepcode(PC_LWZ, _CALLER_SP_, 1, 0, 0);
            setup_caller_sp = pc;
            appendpcode(prologue, pc);
        }
        return _CALLER_SP_;
    } else {
        return _FP_;
    }
}

static UInt32 align_bits(UInt32 value, UInt8 bitcount) {
    UInt32 base = bitcount != 0;
    switch (value) {
        case 0x0002: return base + 30;
        case 0x0004: return base + 29;
        case 0x0008: return base + 28;
        case 0x0010: return base + 27;
        case 0x0020: return base + 26;
        case 0x0040: return base + 25;
        case 0x0080: return base + 24;
        case 0x0100: return base + 23;
        case 0x0200: return base + 22;
        case 0x0400: return base + 21;
        case 0x0800: return base + 20;
        case 0x1000: return base + 19;
        case 0x2000: return base + 18;
        default:
            CError_FATAL(2754);
            return base + 27;
    }
}

Boolean is_large_frame(void) {
    CError_ASSERT(2769, frame_size != -1);
    return large_stack;
}

void no_frame_for_asm(void) {
    frame_size = 0;
}

Boolean can_add_displ_to_local(Object *obj, SInt32 displ) {
    if (obj->datatype != DLOCAL)
        return 0;

    if (local_offset_32(obj) == (short) local_offset_32(obj))
        if ((displ + local_offset_32(obj)) == (short) (displ + local_offset_32(obj)))
            return 1;

    return 0;
}

SInt32 get_alloca_alignment(void) {
    SInt32 align = frame_alignment;
    if (copts.altivec_model)
        align = ALIGN(align, 16);

    if (!alloca_alignment)
        alloca_alignment = align;
    else
        CError_ASSERT(2825, alloca_alignment == align);

    return align_bits(align, 0);
}

static Boolean use_helper_function(char rclass) {
    if (copts.no_register_save_helpers)
        return 0;

    switch (rclass) {
        case RegClass_GPR:
            if (copts.use_lmw_stmw)
                return 0;
            return (used_nonvolatile_registers[RegClass_GPR] > 4) || (copts.optimize_for_size && used_nonvolatile_registers[RegClass_GPR] > 2);
        case RegClass_FPR:
            return (used_nonvolatile_registers[RegClass_FPR] > 3) || (copts.optimize_for_size && used_nonvolatile_registers[RegClass_FPR] > 2);
        case RegClass_VR:
            return (used_nonvolatile_registers[RegClass_VR] > 3) || (copts.optimize_for_size && used_nonvolatile_registers[RegClass_VR] > 2);
        default:
            CError_FATAL(2862);
            return 0;
    }
}

static Boolean need_link_register(void) {
    if (copts.codegen_pic && uses_globals)
        return 1;

    if (makes_call)
        return 1;

    return use_helper_function(RegClass_FPR) || use_helper_function(RegClass_GPR) || use_helper_function(RegClass_VR);
}

static void call_helper_function(char *name, char rclass, short effect) {
    char str[32];
    Object *func;
    NameSpace *save_scope;
    PCode *pc;
    int extra_args;
    PCodeArg *arg;
    short i;

    extra_args = 1;
    if (rclass == RegClass_VR)
        extra_args = 2;

    sprintf(str, name, 32 - used_nonvolatile_registers[rclass]);

    save_scope = cscope_current;
    cscope_current = cscope_root;
    func = CParser_NewRTFunc(&stvoid, NULL, 2, 0);
    cscope_current = save_scope;

    func->name = GetHashNameNodeExport(str);

    pc = makepcode(PC_BL, extra_args + used_nonvolatile_registers[rclass], func, 0);
    for (i = 1, arg = &pc->args[1]; i <= used_nonvolatile_registers[rclass]; i++, arg++) {
        arg->kind = PCOp_REGISTER;
        arg->arg = rclass;
        arg->data.reg.reg = n_real_registers[rclass] - i;
        arg->data.reg.effect = effect;
    }

    if (rclass == RegClass_VR) {
        arg[1].kind = PCOp_REGISTER;
        arg[1].arg = RegClass_GPR;
        arg[1].data.reg.reg = 12;
        arg[1].data.reg.effect = EffectWrite;
        arg[2].kind = PCOp_REGISTER;
        arg[2].arg = RegClass_GPR;
        arg[2].data.reg.reg = 0;
        arg[2].data.reg.effect = EffectRead;
    } else {
        arg[1].kind = PCOp_REGISTER;
        arg[1].arg = RegClass_GPR;
        arg[1].data.reg.reg = 11;
        arg[1].data.reg.effect = EffectRead;
    }

    appendpcode(pclastblock, pc);
    setpcodeflags(fSideEffects);
}

static SInt32 nearest_power_of_two(SInt32 n) {
    SInt32 power = 1;
    do {
        power <<= 1;
    } while (power && power < n);

    CError_ASSERT(2933, power != 0);
    return power;
}

static void compress_data_area(void) {
    // doesn't quite match
    SInt32 r0;
    SInt32 r7;
    ObjectList *list;
    Object *obj;
    PCodeBlock *block;
    PCode *pc;
    int i;

    compressing_data_area = 1;

    if (large_stack) {
        r0 = 0;
    } else {
        r0 = parameter_area_size;
        if (r0 < 32)
            r0 = 32;
    }
    r7 = ALIGN(r0 + 24, frame_alignment) - 24;
    local_data_limit = 0x8000 - ALIGN(24 + in_parameter_size + nonvolatile_save_size + r7, frame_alignment);

    if (local_objects_tail[ObjClass0]) {
        if (local_objects[ObjClass1]) {
            local_objects_tail[ObjClass0]->next = local_objects[ObjClass1];
            local_objects_tail[ObjClass0] = local_objects_tail[ObjClass1];
        }
        if (local_objects[ObjClass2]) {
            local_objects_tail[ObjClass0]->next = local_objects[ObjClass2];
            local_objects_tail[ObjClass0] = local_objects_tail[ObjClass2];
        }
    } else if (local_objects_tail[ObjClass1]) {
        local_objects[ObjClass0] = local_objects[ObjClass1];
        local_objects_tail[ObjClass0] = local_objects_tail[ObjClass1];
        if (local_objects[ObjClass2]) {
            local_objects_tail[ObjClass0]->next = local_objects[ObjClass2];
            local_objects_tail[ObjClass0] = local_objects_tail[ObjClass2];
        }
    } else {
        local_objects[ObjClass0] = local_objects[ObjClass2];
        local_objects_tail[ObjClass0] = local_objects_tail[ObjClass2];
    }

    for (list = local_objects[ObjClass0]; list; list = list->next)
        Registers_GetVarInfo(list->object)->used = 0;

    for (block = pcbasicblocks; block; block = block->nextBlock) {
        for (pc = block->firstPCode; pc; pc = pc->nextPCode) {
            for (i = 0; i < pc->argCount; i++) {
                if (pc->args[i].kind == PCOp_MEMORY && pc->args[i].data.mem.obj && pc->args[i].data.mem.obj->datatype == DLOCAL)
                    Registers_GetVarInfo(pc->args[i].data.mem.obj)->used = 1;
            }
        }
    }

    local_data_size = 0;
    large_data_near_size = 0;
    large_data_far_size = 0;

    for (list = local_objects[ObjClass0]; list; list = list->next) {
        obj = list->object;
        if (Registers_GetVarInfo(obj)->used)
            assign_local_memory(obj);
    }
}

static void insert_local_object(UInt8 oclass, Object *obj) {
    ObjectList *list;

    if (!compressing_data_area) {
        list = lalloc(sizeof(ObjectList));
        memclrw(list, sizeof(ObjectList));
        list->object = obj;
        if (!local_objects[oclass])
            local_objects[oclass] = list;
        if (local_objects_tail[oclass])
            local_objects_tail[oclass]->next = list;
        local_objects_tail[oclass] = list;
    }
}