eden/src/shader_recompiler/ir_opt/lower_int64_to_int32.cpp

// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
// SPDX-License-Identifier: GPL-3.0-or-later

// SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later

#include <utility>

#include "shader_recompiler/exception.h"
#include "shader_recompiler/frontend/ir/basic_block.h"
#include "shader_recompiler/frontend/ir/ir_emitter.h"
#include "shader_recompiler/frontend/ir/program.h"
#include "shader_recompiler/frontend/ir/value.h"
#include "shader_recompiler/ir_opt/passes.h"

namespace Shader::Optimization {
namespace {
std::pair<IR::U32, IR::U32> Unpack(IR::IREmitter& ir, const IR::Value& packed) {
    if (packed.IsImmediate()) {
        const u64 value{packed.U64()};
        return {
            ir.Imm32(static_cast<u32>(value)),
            ir.Imm32(static_cast<u32>(value >> 32)),
        };
    } else {
        return std::pair<IR::U32, IR::U32>{
            ir.CompositeExtract(packed, 0u),
            ir.CompositeExtract(packed, 1u),
        };
    }
}

void IAdd64To32(IR::Block& block, IR::Inst& inst) {
    if (inst.HasAssociatedPseudoOperation()) {
        throw NotImplementedException("IAdd64 emulation with pseudo instructions");
    }
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    const auto [a_lo, a_hi]{Unpack(ir, inst.Arg(0))};
    const auto [b_lo, b_hi]{Unpack(ir, inst.Arg(1))};

    const IR::U32 ret_lo{ir.IAdd(a_lo, b_lo)};
    const IR::U32 carry{ir.Select(ir.GetCarryFromOp(ret_lo), ir.Imm32(1u), ir.Imm32(0u))};

    const IR::U32 ret_hi{ir.IAdd(ir.IAdd(a_hi, b_hi), carry)};
    inst.ReplaceUsesWith(ir.CompositeConstruct(ret_lo, ret_hi));
}

void ISub64To32(IR::Block& block, IR::Inst& inst) {
    if (inst.HasAssociatedPseudoOperation()) {
        throw NotImplementedException("ISub64 emulation with pseudo instructions");
    }
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    const auto [a_lo, a_hi]{Unpack(ir, inst.Arg(0))};
    const auto [b_lo, b_hi]{Unpack(ir, inst.Arg(1))};

    const IR::U32 ret_lo{ir.ISub(a_lo, b_lo)};
    const IR::U1 underflow{ir.IGreaterThan(ret_lo, a_lo, false)};
    const IR::U32 underflow_bit{ir.Select(underflow, ir.Imm32(1u), ir.Imm32(0u))};

    const IR::U32 ret_hi{ir.ISub(ir.ISub(a_hi, b_hi), underflow_bit)};
    inst.ReplaceUsesWith(ir.CompositeConstruct(ret_lo, ret_hi));
}

void INeg64To32(IR::Block& block, IR::Inst& inst) {
    if (inst.HasAssociatedPseudoOperation()) {
        throw NotImplementedException("INeg64 emulation with pseudo instructions");
    }
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    auto [lo, hi]{Unpack(ir, inst.Arg(0))};
    lo = ir.BitwiseNot(lo);
    hi = ir.BitwiseNot(hi);

    lo = ir.IAdd(lo, ir.Imm32(1));

    const IR::U32 carry{ir.Select(ir.GetCarryFromOp(lo), ir.Imm32(1u), ir.Imm32(0u))};
    hi = ir.IAdd(hi, carry);

    inst.ReplaceUsesWith(ir.CompositeConstruct(lo, hi));
}

void ShiftLeftLogical64To32(IR::Block& block, IR::Inst& inst) {
    if (inst.HasAssociatedPseudoOperation()) {
        throw NotImplementedException("ShiftLeftLogical64 emulation with pseudo instructions");
    }
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    const auto [lo, hi]{Unpack(ir, inst.Arg(0))};
    const IR::U32 shift{inst.Arg(1)};

    const IR::U32 shifted_lo{ir.ShiftLeftLogical(lo, shift)};
    const IR::U32 shifted_hi{ir.ShiftLeftLogical(hi, shift)};

    const IR::U32 inv_shift{ir.ISub(shift, ir.Imm32(32))};
    const IR::U1 is_long{ir.IGreaterThanEqual(inv_shift, ir.Imm32(0), true)};
    const IR::U1 is_zero{ir.IEqual(shift, ir.Imm32(0))};

    const IR::U32 long_ret_lo{ir.Imm32(0)};
    const IR::U32 long_ret_hi{ir.ShiftLeftLogical(lo, inv_shift)};

    const IR::U32 shift_complement{ir.ISub(ir.Imm32(32), shift)};
    const IR::U32 lo_extract{ir.BitFieldExtract(lo, shift_complement, shift, false)};
    const IR::U32 short_ret_lo{shifted_lo};
    const IR::U32 short_ret_hi{ir.BitwiseOr(shifted_hi, lo_extract)};

    const IR::U32 zero_ret_lo{lo};
    const IR::U32 zero_ret_hi{hi};

    const IR::U32 non_zero_lo{ir.Select(is_long, long_ret_lo, short_ret_lo)};
    const IR::U32 non_zero_hi{ir.Select(is_long, long_ret_hi, short_ret_hi)};

    const IR::U32 ret_lo{ir.Select(is_zero, zero_ret_lo, non_zero_lo)};
    const IR::U32 ret_hi{ir.Select(is_zero, zero_ret_hi, non_zero_hi)};
    inst.ReplaceUsesWith(ir.CompositeConstruct(ret_lo, ret_hi));
}

void ShiftRightLogical64To32(IR::Block& block, IR::Inst& inst) {
    if (inst.HasAssociatedPseudoOperation()) {
        throw NotImplementedException("ShiftRightLogical64 emulation with pseudo instructions");
    }
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    const auto [lo, hi]{Unpack(ir, inst.Arg(0))};
    const IR::U32 shift{inst.Arg(1)};

    const IR::U32 shifted_lo{ir.ShiftRightLogical(lo, shift)};
    const IR::U32 shifted_hi{ir.ShiftRightLogical(hi, shift)};

    const IR::U32 inv_shift{ir.ISub(shift, ir.Imm32(32))};
    const IR::U1 is_long{ir.IGreaterThanEqual(inv_shift, ir.Imm32(0), true)};
    const IR::U1 is_zero{ir.IEqual(shift, ir.Imm32(0))};

    const IR::U32 long_ret_hi{ir.Imm32(0)};
    const IR::U32 long_ret_lo{ir.ShiftRightLogical(hi, inv_shift)};

    const IR::U32 shift_complement{ir.ISub(ir.Imm32(32), shift)};
    const IR::U32 short_hi_extract{ir.BitFieldExtract(hi, ir.Imm32(0), shift)};
    const IR::U32 short_ret_hi{shifted_hi};
    const IR::U32 short_ret_lo{
        ir.BitFieldInsert(shifted_lo, short_hi_extract, shift_complement, shift)};

    const IR::U32 zero_ret_lo{lo};
    const IR::U32 zero_ret_hi{hi};

    const IR::U32 non_zero_lo{ir.Select(is_long, long_ret_lo, short_ret_lo)};
    const IR::U32 non_zero_hi{ir.Select(is_long, long_ret_hi, short_ret_hi)};

    const IR::U32 ret_lo{ir.Select(is_zero, zero_ret_lo, non_zero_lo)};
    const IR::U32 ret_hi{ir.Select(is_zero, zero_ret_hi, non_zero_hi)};
    inst.ReplaceUsesWith(ir.CompositeConstruct(ret_lo, ret_hi));
}

void ShiftRightArithmetic64To32(IR::Block& block, IR::Inst& inst) {
    if (inst.HasAssociatedPseudoOperation()) {
        throw NotImplementedException("ShiftRightArithmetic64 emulation with pseudo instructions");
    }
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    const auto [lo, hi]{Unpack(ir, inst.Arg(0))};
    const IR::U32 shift{inst.Arg(1)};

    const IR::U32 shifted_lo{ir.ShiftRightLogical(lo, shift)};
    const IR::U32 shifted_hi{ir.ShiftRightArithmetic(hi, shift)};

    const IR::U32 sign_extension{ir.ShiftRightArithmetic(hi, ir.Imm32(31))};

    const IR::U32 inv_shift{ir.ISub(shift, ir.Imm32(32))};
    const IR::U1 is_long{ir.IGreaterThanEqual(inv_shift, ir.Imm32(0), true)};
    const IR::U1 is_zero{ir.IEqual(shift, ir.Imm32(0))};

    const IR::U32 long_ret_hi{sign_extension};
    const IR::U32 long_ret_lo{ir.ShiftRightArithmetic(hi, inv_shift)};

    const IR::U32 shift_complement{ir.ISub(ir.Imm32(32), shift)};
    const IR::U32 short_hi_extract(ir.BitFieldExtract(hi, ir.Imm32(0), shift));
    const IR::U32 short_ret_hi{shifted_hi};
    const IR::U32 short_ret_lo{
        ir.BitFieldInsert(shifted_lo, short_hi_extract, shift_complement, shift)};

    const IR::U32 zero_ret_lo{lo};
    const IR::U32 zero_ret_hi{hi};

    const IR::U32 non_zero_lo{ir.Select(is_long, long_ret_lo, short_ret_lo)};
    const IR::U32 non_zero_hi{ir.Select(is_long, long_ret_hi, short_ret_hi)};

    const IR::U32 ret_lo{ir.Select(is_zero, zero_ret_lo, non_zero_lo)};
    const IR::U32 ret_hi{ir.Select(is_zero, zero_ret_hi, non_zero_hi)};
    inst.ReplaceUsesWith(ir.CompositeConstruct(ret_lo, ret_hi));
}

void IAbs64To32(IR::Block& block, IR::Inst& inst) {
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    const auto [lo, hi]{Unpack(ir, inst.Arg(0))};

    const IR::U32 neg_lo{ir.IAdd(ir.BitwiseNot(lo), ir.Imm32(1))};
    const IR::U32 carry{IR::U32{ir.Select(ir.GetCarryFromOp(neg_lo), ir.Imm32(1u), ir.Imm32(0u))}};
    const IR::U32 neg_hi{ir.IAdd(ir.BitwiseNot(hi), carry)};

    const IR::U1 is_negative{ir.INotEqual(ir.BitwiseAnd(hi, ir.Imm32(0x80000000u)), ir.Imm32(0u))};
    const IR::U32 ret_lo{IR::U32{ir.Select(is_negative, neg_lo, lo)}};
    const IR::U32 ret_hi{IR::U32{ir.Select(is_negative, neg_hi, hi)}};
    inst.ReplaceUsesWith(ir.CompositeConstruct(ret_lo, ret_hi));
}

void SelectU64To32(IR::Block& block, IR::Inst& inst) {
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    const IR::U1 condition{inst.Arg(0)};
    const auto [true_lo, true_hi]{Unpack(ir, inst.Arg(1))};
    const auto [false_lo, false_hi]{Unpack(ir, inst.Arg(2))};

    const IR::U32 ret_lo{IR::U32{ir.Select(condition, true_lo, false_lo)}};
    const IR::U32 ret_hi{IR::U32{ir.Select(condition, true_hi, false_hi)}};
    inst.ReplaceUsesWith(ir.CompositeConstruct(ret_lo, ret_hi));
}

void UndefU64To32(IR::Block& block, IR::Inst& inst) {
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    inst.ReplaceUsesWith(ir.CompositeConstruct(ir.Imm32(0u), ir.Imm32(0u)));
}

void ConvertU64U32To32(IR::Block& block, IR::Inst& inst) {
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    inst.ReplaceUsesWith(ir.CompositeConstruct(IR::U32{inst.Arg(0)}, ir.Imm32(0u)));
}

void ConvertU32U64To32(IR::Block& block, IR::Inst& inst) {
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    inst.ReplaceUsesWith(Unpack(ir, inst.Arg(0)).first);
}

void IntToFloat64To32(IR::Block& block, IR::Inst& inst, bool is_signed, size_t dest_bitsize) {
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    const auto [lo, hi]{Unpack(ir, inst.Arg(0))};
    const IR::F32 low{ir.ConvertUToF(32, 32, lo)};
    const IR::F32 high{is_signed ? IR::F32{ir.ConvertSToF(32, 32, hi)}
                                 : IR::F32{ir.ConvertUToF(32, 32, hi)}};
    const IR::F32 combined{ir.FPFma(high, ir.Imm32(4294967296.0f), low)};
    if (dest_bitsize == 32) {
        inst.ReplaceUsesWith(combined);
    } else {
        inst.ReplaceUsesWith(ir.FPConvert(dest_bitsize, combined));
    }
}

void FloatToInt64To32(IR::Block& block, IR::Inst& inst, bool is_signed, size_t src_bitsize) {
    IR::IREmitter ir(block, IR::Block::InstructionList::s_iterator_to(inst));
    const IR::F32 value{src_bitsize == 32 ? IR::F32{inst.Arg(0)}
                                          : IR::F32{ir.FPConvert(32, IR::F16F32F64{inst.Arg(0)})}};
    const IR::F32 high_f{ir.FPFloor(ir.FPMul(value, ir.Imm32(1.0f / 4294967296.0f)))};
    const IR::U32 hi{is_signed ? IR::U32{ir.ConvertFToS(32, high_f)}
                               : IR::U32{ir.ConvertFToU(32, high_f)}};
    const IR::F32 low_f{ir.FPFma(high_f, ir.FPNeg(ir.Imm32(4294967296.0f)), value)};
    const IR::U32 lo{IR::U32{ir.ConvertFToU(32, low_f)}};
    inst.ReplaceUsesWith(ir.CompositeConstruct(lo, hi));
}

void Lower(IR::Block& block, IR::Inst& inst) {
    switch (inst.GetOpcode()) {
    case IR::Opcode::PackUint2x32:
    case IR::Opcode::UnpackUint2x32:
        return inst.ReplaceOpcode(IR::Opcode::Identity);
    case IR::Opcode::IAdd64:
        return IAdd64To32(block, inst);
    case IR::Opcode::ISub64:
        return ISub64To32(block, inst);
    case IR::Opcode::INeg64:
        return INeg64To32(block, inst);
    case IR::Opcode::ShiftLeftLogical64:
        return ShiftLeftLogical64To32(block, inst);
    case IR::Opcode::ShiftRightLogical64:
        return ShiftRightLogical64To32(block, inst);
    case IR::Opcode::ShiftRightArithmetic64:
        return ShiftRightArithmetic64To32(block, inst);
    case IR::Opcode::SharedAtomicExchange64:
        return inst.ReplaceOpcode(IR::Opcode::SharedAtomicExchange32x2);
    case IR::Opcode::GlobalAtomicIAdd64:
        return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicIAdd32x2);
    case IR::Opcode::GlobalAtomicSMin64:
        return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicSMin32x2);
    case IR::Opcode::GlobalAtomicUMin64:
        return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicUMin32x2);
    case IR::Opcode::GlobalAtomicSMax64:
        return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicSMax32x2);
    case IR::Opcode::GlobalAtomicUMax64:
        return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicUMax32x2);
    case IR::Opcode::GlobalAtomicAnd64:
        return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicAnd32x2);
    case IR::Opcode::GlobalAtomicOr64:
        return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicOr32x2);
    case IR::Opcode::GlobalAtomicXor64:
        return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicXor32x2);
    case IR::Opcode::GlobalAtomicExchange64:
        return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicExchange32x2);
    case IR::Opcode::StorageAtomicIAdd64:
        return inst.ReplaceOpcode(IR::Opcode::StorageAtomicIAdd32x2);
    case IR::Opcode::StorageAtomicSMin64:
        return inst.ReplaceOpcode(IR::Opcode::StorageAtomicSMin32x2);
    case IR::Opcode::StorageAtomicUMin64:
        return inst.ReplaceOpcode(IR::Opcode::StorageAtomicUMin32x2);
    case IR::Opcode::StorageAtomicSMax64:
        return inst.ReplaceOpcode(IR::Opcode::StorageAtomicSMax32x2);
    case IR::Opcode::StorageAtomicUMax64:
        return inst.ReplaceOpcode(IR::Opcode::StorageAtomicUMax32x2);
    case IR::Opcode::StorageAtomicAnd64:
        return inst.ReplaceOpcode(IR::Opcode::StorageAtomicAnd32x2);
    case IR::Opcode::StorageAtomicOr64:
        return inst.ReplaceOpcode(IR::Opcode::StorageAtomicOr32x2);
    case IR::Opcode::StorageAtomicXor64:
        return inst.ReplaceOpcode(IR::Opcode::StorageAtomicXor32x2);
    case IR::Opcode::StorageAtomicExchange64:
        return inst.ReplaceOpcode(IR::Opcode::StorageAtomicExchange32x2);
    case IR::Opcode::BitCastU64F64:
        return inst.ReplaceOpcode(IR::Opcode::UnpackDouble2x32);
    case IR::Opcode::BitCastF64U64:
        return inst.ReplaceOpcode(IR::Opcode::PackDouble2x32);
    case IR::Opcode::UndefU64:
        return UndefU64To32(block, inst);
    case IR::Opcode::SelectU64:
        return SelectU64To32(block, inst);
    case IR::Opcode::IAbs64:
        return IAbs64To32(block, inst);
    case IR::Opcode::ConvertU64U32:
        return ConvertU64U32To32(block, inst);
    case IR::Opcode::ConvertU32U64:
        return ConvertU32U64To32(block, inst);
    case IR::Opcode::ConvertS64F16:
        return FloatToInt64To32(block, inst, true, 16);
    case IR::Opcode::ConvertS64F32:
        return FloatToInt64To32(block, inst, true, 32);
    case IR::Opcode::ConvertS64F64:
        return FloatToInt64To32(block, inst, true, 64);
    case IR::Opcode::ConvertU64F16:
        return FloatToInt64To32(block, inst, false, 16);
    case IR::Opcode::ConvertU64F32:
        return FloatToInt64To32(block, inst, false, 32);
    case IR::Opcode::ConvertU64F64:
        return FloatToInt64To32(block, inst, false, 64);
    case IR::Opcode::ConvertF16S64:
        return IntToFloat64To32(block, inst, true, 16);
    case IR::Opcode::ConvertF32S64:
        return IntToFloat64To32(block, inst, true, 32);
    case IR::Opcode::ConvertF64S64:
        return IntToFloat64To32(block, inst, true, 64);
    case IR::Opcode::ConvertF16U64:
        return IntToFloat64To32(block, inst, false, 16);
    case IR::Opcode::ConvertF32U64:
        return IntToFloat64To32(block, inst, false, 32);
    case IR::Opcode::ConvertF64U64:
        return IntToFloat64To32(block, inst, false, 64);
    default:
        break;
    }
}
} // Anonymous namespace

void LowerInt64ToInt32(IR::Program& program) {
    const auto end{program.post_order_blocks.rend()};
    for (auto it = program.post_order_blocks.rbegin(); it != end; ++it) {
        IR::Block* const block{*it};
        for (IR::Inst& inst : block->Instructions()) {
            Lower(*block, inst);
        }
    }
}

} // namespace Shader::Optimization