gglassembler.cpp

Android 一些工具

        }
        if (mask & Z_WRITE) {
            if (mask == Z_WRITE) {
                // only z-write asked, cc is meaningless
                ic = AL;
            }
            MOV(AL, 0, depth, reg_imm(z, LSR, 16));
            STRH(ic, depth, zbase);
        }
    }
}

void GGLAssembler::build_iterate_z(const fragment_parts_t& parts)
{
    const needs_t& needs = mBuilderContext.needs;
    if ((mDepthTest != GGL_ALWAYS) || GGL_READ_NEEDS(P_MASK_Z, needs.p)) {
        Scratch scratches(registerFile());
        int dzdx = scratches.obtain();
        CONTEXT_LOAD(dzdx, generated_vars.dzdx);    // stall
        ADD(AL, 0, parts.z.reg, parts.z.reg, dzdx); 
    }
}

void GGLAssembler::build_iterate_f(const fragment_parts_t& parts)
{
    const needs_t& needs = mBuilderContext.needs;
    if (GGL_READ_NEEDS(P_FOG, needs.p)) {
        Scratch scratches(registerFile());
        int dfdx = scratches.obtain();
        int f = scratches.obtain();
        CONTEXT_LOAD(f,     generated_vars.f);
        CONTEXT_LOAD(dfdx,  generated_vars.dfdx);   // stall
        ADD(AL, 0, f, f, dfdx);
        CONTEXT_STORE(f,    generated_vars.f);
    }
}

// ---------------------------------------------------------------------------

void GGLAssembler::build_logic_op(pixel_t& pixel, Scratch& regs)
{
    const needs_t& needs = mBuilderContext.needs;
    const int opcode = GGL_READ_NEEDS(LOGIC_OP, needs.n) | GGL_CLEAR;
    if (opcode == GGL_COPY)
        return;
    
    comment("logic operation");

    pixel_t s(pixel);
    if (!(pixel.flags & CORRUPTIBLE)) {
        pixel.reg = regs.obtain();
        pixel.flags |= CORRUPTIBLE;
    }
    
    pixel_t d(mDstPixel);
    switch(opcode) {
    case GGL_CLEAR:         MOV(AL, 0, pixel.reg, imm(0));          break;
    case GGL_AND:           AND(AL, 0, pixel.reg, s.reg, d.reg);    break;
    case GGL_AND_REVERSE:   BIC(AL, 0, pixel.reg, s.reg, d.reg);    break;
    case GGL_COPY:                                                  break;
    case GGL_AND_INVERTED:  BIC(AL, 0, pixel.reg, d.reg, s.reg);    break;
    case GGL_NOOP:          MOV(AL, 0, pixel.reg, d.reg);           break;
    case GGL_XOR:           EOR(AL, 0, pixel.reg, s.reg, d.reg);    break;
    case GGL_OR:            ORR(AL, 0, pixel.reg, s.reg, d.reg);    break;
    case GGL_NOR:           ORR(AL, 0, pixel.reg, s.reg, d.reg);
                            MVN(AL, 0, pixel.reg, pixel.reg);       break;
    case GGL_EQUIV:         EOR(AL, 0, pixel.reg, s.reg, d.reg);
                            MVN(AL, 0, pixel.reg, pixel.reg);       break;
    case GGL_INVERT:        MVN(AL, 0, pixel.reg, d.reg);           break;
    case GGL_OR_REVERSE:    // s | ~d == ~(~s & d)
                            BIC(AL, 0, pixel.reg, d.reg, s.reg);
                            MVN(AL, 0, pixel.reg, pixel.reg);       break;
    case GGL_COPY_INVERTED: MVN(AL, 0, pixel.reg, s.reg);           break;
    case GGL_OR_INVERTED:   // ~s | d == ~(s & ~d)
                            BIC(AL, 0, pixel.reg, s.reg, d.reg);
                            MVN(AL, 0, pixel.reg, pixel.reg);       break;
    case GGL_NAND:          AND(AL, 0, pixel.reg, s.reg, d.reg);
                            MVN(AL, 0, pixel.reg, pixel.reg);       break;
    case GGL_SET:           MVN(AL, 0, pixel.reg, imm(0));          break;
    };        
}

// ---------------------------------------------------------------------------

static uint32_t find_bottom(uint32_t val)
{
    uint32_t i = 0;
    while (!(val & (3<<i)))
        i+= 2;
    return i;
}

static void normalize(uint32_t& val, uint32_t& rot)
{
    rot = 0;
    while (!(val&3)  || (val & 0xFC000000)) {
        uint32_t newval;
        newval = val >> 2;
        newval |= (val&3) << 30;
        val = newval;
        rot += 2;
        if (rot == 32) {
            rot = 0;
            break;
        }
    }
}

void GGLAssembler::build_and_immediate(int d, int s, uint32_t mask, int bits)
{
    uint32_t rot;
    uint32_t size = ((bits>=32) ? 0 : (1LU << bits)) - 1;
    mask &= size;

    if (mask == size) {
        if (d != s)
            MOV( AL, 0, d, s);
        return;
    }
    
    int negative_logic = !isValidImmediate(mask);
    if (negative_logic) {
        mask = ~mask & size;
    }
    normalize(mask, rot);

    if (mask) {
        while (mask) {
            uint32_t bitpos = find_bottom(mask);
            int shift = rot + bitpos;
            uint32_t m = mask & (0xff << bitpos);
            mask &= ~m;
            m >>= bitpos;
            int32_t newMask =  (m<<shift) | (m>>(32-shift));
            if (!negative_logic) {
                AND( AL, 0, d, s, imm(newMask) );
            } else {
                BIC( AL, 0, d, s, imm(newMask) );
            }
            s = d;
        }
    } else {
        MOV( AL, 0, d, imm(0));
    }
}		

void GGLAssembler::build_masking(pixel_t& pixel, Scratch& regs)
{
    if (!mMasking)
        return;

    comment("color mask");

    pixel_t fb(mDstPixel);
    pixel_t s(pixel);
    if (!(pixel.flags & CORRUPTIBLE)) {
        pixel.reg = regs.obtain();
        pixel.flags |= CORRUPTIBLE;
    }

    int mask = 0;
    for (int i=0 ; i<4 ; i++) {
        const int component_mask = 1<<i;
        const int h = fb.format.c[i].h;
        const int l = fb.format.c[i].l;
        if (h && (!(mMasking & component_mask))) {
            mask |= ((1<<(h-l))-1) << l;
        }
    }

    // There is no need to clear the masked components of the source
    // (unless we applied a logic op), because they're already zeroed 
    // by contruction (masked components are not computed)

    if (mLogicOp) {
        const needs_t& needs = mBuilderContext.needs;
        const int opcode = GGL_READ_NEEDS(LOGIC_OP, needs.n) | GGL_CLEAR;
        if (opcode != GGL_CLEAR) {
            // clear masked component of source
            build_and_immediate(pixel.reg, s.reg, mask, fb.size());
            s = pixel;
        }
    }

    // clear non masked components of destination
    build_and_immediate(fb.reg, fb.reg, ~mask, fb.size()); 

    // or back the channels that were masked
    if (s.reg == fb.reg) {
         // this is in fact a MOV
        if (s.reg == pixel.reg) {
            // ugh. this in in fact a nop
        } else {
            MOV(AL, 0, pixel.reg, fb.reg);
        }
    } else {
        ORR(AL, 0, pixel.reg, s.reg, fb.reg);
    }
}

// ---------------------------------------------------------------------------

void GGLAssembler::base_offset(
        const pointer_t& d, const pointer_t& b, const reg_t& o)
{
    switch (b.size) {
    case 32:
        ADD(AL, 0, d.reg, b.reg, reg_imm(o.reg, LSL, 2));
        break;
    case 24:
        if (d.reg == b.reg) {
            ADD(AL, 0, d.reg, b.reg, reg_imm(o.reg, LSL, 1));
            ADD(AL, 0, d.reg, d.reg, o.reg);
        } else {
            ADD(AL, 0, d.reg, o.reg, reg_imm(o.reg, LSL, 1));
            ADD(AL, 0, d.reg, d.reg, b.reg);
        }
        break;
    case 16:
        ADD(AL, 0, d.reg, b.reg, reg_imm(o.reg, LSL, 1));
        break;
    case 8:
        ADD(AL, 0, d.reg, b.reg, o.reg);
        break;
    }
}

// ----------------------------------------------------------------------------
// cheezy register allocator...
// ----------------------------------------------------------------------------

void RegisterAllocator::reset()
{
    mRegs.reset();
}

int RegisterAllocator::reserveReg(int reg)
{
    return mRegs.reserve(reg);
}

int RegisterAllocator::obtainReg()
{
    return mRegs.obtain();
}

void RegisterAllocator::recycleReg(int reg)
{
    mRegs.recycle(reg);
}

RegisterAllocator::RegisterFile& RegisterAllocator::registerFile()
{
    return mRegs;
}

// ----------------------------------------------------------------------------

RegisterAllocator::RegisterFile::RegisterFile()
    : mRegs(0), mTouched(0), mStatus(0)
{
    reserve(ARMAssemblerInterface::SP);
    reserve(ARMAssemblerInterface::PC);
}

RegisterAllocator::RegisterFile::RegisterFile(const RegisterFile& rhs)
    : mRegs(rhs.mRegs), mTouched(rhs.mTouched)
{
}

RegisterAllocator::RegisterFile::~RegisterFile()
{
}

bool RegisterAllocator::RegisterFile::operator == (const RegisterFile& rhs) const
{
    return (mRegs == rhs.mRegs);
}

void RegisterAllocator::RegisterFile::reset()
{
    mRegs = mTouched = mStatus = 0;
    reserve(ARMAssemblerInterface::SP);
    reserve(ARMAssemblerInterface::PC);
}

int RegisterAllocator::RegisterFile::reserve(int reg)
{
    LOG_ALWAYS_FATAL_IF(isUsed(reg),
                        "reserving register %d, but already in use",
                        reg);
    mRegs |= (1<<reg);
    mTouched |= mRegs;
    return reg;
}

void RegisterAllocator::RegisterFile::reserveSeveral(uint32_t regMask)
{
    mRegs |= regMask;
    mTouched |= regMask;
}

int RegisterAllocator::RegisterFile::isUsed(int reg) const
{
    LOG_ALWAYS_FATAL_IF(reg>=16, "invalid register %d", reg);
    return mRegs & (1<<reg);
}

int RegisterAllocator::RegisterFile::obtain()
{
    const char priorityList[14] = {  0,  1, 2, 3, 
                                    12, 14, 4, 5, 
                                     6,  7, 8, 9,
                                    10, 11 };
    const int nbreg = sizeof(priorityList);
    int i, r;
    for (i=0 ; i<nbreg ; i++) {
        r = priorityList[i];
        if (!isUsed(r)) {
            break;
        }
    }
    // this is not an error anymore because, we'll try again with
    // a lower optimization level.
    //LOGE_IF(i >= nbreg, "pixelflinger ran out of registers\n");
    if (i >= nbreg) {
        mStatus |= OUT_OF_REGISTERS;
        // we return SP so we can more easily debug things
        // the code will never be run anyway.
        return ARMAssemblerInterface::SP; 
    }
    reserve(r);
    return r;
}

bool RegisterAllocator::RegisterFile::hasFreeRegs() const
{
    return ((mRegs & 0xFFFF) == 0xFFFF) ? false : true;
}

int RegisterAllocator::RegisterFile::countFreeRegs() const
{
    int f = ~mRegs & 0xFFFF;
    // now count number of 1
   f = (f & 0x5555) + ((f>>1) & 0x5555);
   f = (f & 0x3333) + ((f>>2) & 0x3333);
   f = (f & 0x0F0F) + ((f>>4) & 0x0F0F);
   f = (f & 0x00FF) + ((f>>8) & 0x00FF);
   return f;
}

void RegisterAllocator::RegisterFile::recycle(int reg)
{
    LOG_FATAL_IF(!isUsed(reg),
            "recycling unallocated register %d",
            reg);
    mRegs &= ~(1<<reg);
}

void RegisterAllocator::RegisterFile::recycleSeveral(uint32_t regMask)
{
    LOG_FATAL_IF((mRegs & regMask)!=regMask,
            "recycling unallocated registers "
            "(recycle=%08x, allocated=%08x, unallocated=%08x)",
            regMask, mRegs, mRegs&regMask);
    mRegs &= ~regMask;
}

uint32_t RegisterAllocator::RegisterFile::touched() const
{
    return mTouched;
}

// ----------------------------------------------------------------------------

}; // namespace android