runtimeoptimize.cpp

// Copyright Contributors to the Open Shading Language project.
// SPDX-License-Identifier: BSD-3-Clause
// https://github.com/AcademySoftwareFoundation/OpenShadingLanguage

#include <cmath>
#include <cstdio>
#include <vector>

#include <OpenImageIO/sysutil.h>
#include <OpenImageIO/thread.h>
#include <OpenImageIO/timer.h>

#include "batched_analysis.h"
#include "oslexec_pvt.h"
#include "runtimeoptimize.h"
using namespace OSL;
using namespace OSL::pvt;


// names of ops we'll be using frequently
static ustring u_nop("nop");
static ustring u_exit("exit");
static ustring u_assign("assign");
static ustring u_add("add");
static ustring u_sub("sub");
static ustring u_mul("mul");
static ustring u_if("if");
static ustring u_for("for");
static ustring u_while("while");
static ustring u_dowhile("dowhile");
static ustring u_functioncall("functioncall");
static ustring u_functioncall_nr("functioncall_nr");
static ustring u_break("break");
static ustring u_continue("continue");
static ustring u_return("return");
static ustring u_useparam("useparam");
static ustring u_closure("closure");
static ustring u_pointcloud_write("pointcloud_write");
static ustring u_isconnected("isconnected");
static ustring u_setmessage("setmessage");
static ustring u_getmessage("getmessage");
static ustring u_getattribute("getattribute");
static ustring u_backfacing("backfacing");
static ustring u_calculatenormal("calculatenormal");
static ustring u_flipHandedness("flipHandedness");
static ustring u_N("N");
static ustring u_I("I");
static ustring main_method_name("___main___");


OSL_NAMESPACE_BEGIN

namespace pvt {  // OSL::pvt

using OIIO::spin_lock;
using OIIO::Timer;

DECLFOLDER(constfold_assign);  // forward decl



/// Wrapper that erases elements of c for which predicate p is true.
/// (Unlike std::remove_if, it resizes the container so that it contains
/// ONLY elements for which the predicate is true.)
template<class Container, class Predicate>
void
erase_if(Container& c, const Predicate& p)
{
    c.erase(std::remove_if(c.begin(), c.end(), p), c.end());
}



OSOProcessorBase::OSOProcessorBase(ShadingSystemImpl& shadingsys,
                                   ShaderGroup& group, ShadingContext* ctx)
    : m_shadingsys(shadingsys)
    , m_group(group)
    , m_context(ctx)
    , m_debug(shadingsys.debug())
    , m_inst(NULL)
{
    set_debug();
}



OSOProcessorBase::~OSOProcessorBase() {}



RuntimeOptimizer::RuntimeOptimizer(ShadingSystemImpl& shadingsys,
                                   ShaderGroup& group, ShadingContext* ctx)
    : OSOProcessorBase(shadingsys, group, ctx)
    , m_optimize(shadingsys.optimize())
    , m_opt_simplify_param(shadingsys.m_opt_simplify_param)
    , m_opt_constant_fold(shadingsys.m_opt_constant_fold)
    , m_opt_stale_assign(shadingsys.m_opt_stale_assign)
    , m_opt_elide_useless_ops(shadingsys.m_opt_elide_useless_ops)
    , m_opt_elide_unconnected_outputs(
          shadingsys.m_opt_elide_unconnected_outputs)
    , m_opt_peephole(shadingsys.m_opt_peephole)
    , m_opt_coalesce_temps(shadingsys.m_opt_coalesce_temps)
    , m_opt_assign(shadingsys.m_opt_assign)
    , m_opt_mix(shadingsys.m_opt_mix)
    , m_opt_middleman(shadingsys.m_opt_middleman)
    , m_opt_batched_analysis(shadingsys.m_opt_batched_analysis)
    , m_keep_no_return_function_calls(shadingsys.m_llvm_debugging_symbols)
    , m_pass(0)
    , m_next_newconst(0)
    , m_next_newtemp(0)
    , m_stat_opt_locking_time(0)
    , m_stat_specialization_time(0)
    , m_stop_optimizing(false)
    , m_raytypes_on(group.raytypes_on())
    , m_raytypes_off(group.raytypes_off())
{
    memset((char*)&m_shaderglobals, 0, sizeof(ShaderGlobals));
    m_shaderglobals.context = shadingcontext();
    // To handle error/warning/print that might be reported through
    // the renderer services, we will need to provide the default one
    // which will just report through the ShadingContext
    m_shaderglobals.renderer = &m_rendererservices;

    // Disable no_function_return_calls for OptiX renderers, because we
    // aren't yet set up to support use of debugging symbols for PTX.
    // FIXME: some day, we are going to want debugging symbols for PTX, and
    // will need some refactoring of the debugging symbol code.
    if (shadingsys.renderer()->supports("OptiX"))
        m_keep_no_return_function_calls = false;
}



RuntimeOptimizer::~RuntimeOptimizer() {}



void
OSOProcessorBase::set_inst(int newlayer)
{
    m_layer = newlayer;
    m_inst  = group()[m_layer];
    OSL_DASSERT(m_inst != NULL);
    set_debug();
}



void
RuntimeOptimizer::set_inst(int newlayer)
{
    OSOProcessorBase::set_inst(newlayer);
    m_all_consts.clear();
    m_symbol_aliases.clear();
    m_block_aliases.clear();
    m_param_aliases.clear();
    m_bblockids.clear();
}



void
OSOProcessorBase::set_debug()
{
    // start with the shading system's idea of debugging level
    m_debug = shadingsys().debug();

    // If either group or layer was specified for debug, surely they want
    // debugging turned on.
    if (!shadingsys().debug_groupname().empty()
        || !shadingsys().debug_layername().empty())
        m_debug = std::max(m_debug, 1);

    // Force debugging off if a specific group was selected for debug
    // and we're not it, or a specific layer was selected for debug and
    // we're not it.
    bool wronggroup = (!shadingsys().debug_groupname().empty()
                       && shadingsys().debug_groupname() != group().name());
    bool wronglayer = (!shadingsys().debug_layername().empty() && inst()
                       && shadingsys().debug_layername()
                              != inst()->layername());
    if (wronggroup || wronglayer)
        m_debug = 0;
}



void
RuntimeOptimizer::set_debug()
{
    OSOProcessorBase::set_debug();

    // If a specific group is isolated for debugging and  the
    // 'optimize_dondebug' flag is on, fully optimize all other groups.
    if (!shadingsys().debug_groupname().empty()
        && shadingsys().debug_groupname() != group().name()) {
        if (shadingsys().m_optimize_nondebug) {
            // Debugging trick: if user said to only debug one group, turn
            // on full optimization for all others!  This prevents
            // everything from running 10x slower just because you want to
            // debug one shader.
            m_optimize                      = 3;
            m_opt_simplify_param            = true;
            m_opt_constant_fold             = true;
            m_opt_stale_assign              = true;
            m_opt_elide_useless_ops         = true;
            m_opt_elide_unconnected_outputs = true;
            m_opt_peephole                  = true;
            m_opt_coalesce_temps            = true;
            m_opt_assign                    = true;
            m_opt_mix                       = true;
            m_opt_middleman                 = true;
        }
    }
}



int
RuntimeOptimizer::find_constant(const TypeSpec& type, const void* data)
{
    for (int c : m_all_consts) {
        const Symbol& s(*inst()->symbol(c));
        OSL_DASSERT(s.symtype() == SymTypeConst);
        if (equivalent(s.typespec(), type)
            && !memcmp(s.data(), data, s.typespec().simpletype().size())) {
            return c;
        }
    }
    return -1;
}



int
RuntimeOptimizer::add_constant(const TypeSpec& type, const void* data,
                               TypeDesc datatype)
{
    int ind = find_constant(type, data);
    if (ind < 0) {
        // support varlen arrays
        TypeSpec newtype = type;
        if (type.is_unsized_array())
            newtype.make_array(datatype.numelements());

        ustring symname;
        if (newtype.is_int()) {
            // If it's a constant integer, make the name $newconst_int42 and
            // if it's negative, $newconst_int_42 (because we can't put a
            // minus sign in a symbol name). This is easier to read than
            // naming them sequentially (you know the value in the constant)
            // and it means the name will be the same regardless of the order
            // in which we made the constants (helpful for PTX cache
            // behavior).
            int val = *(const int*)data;
            symname = ustring::fmtformat("$newconst_int{}{}",
                                         val < 0 ? "_" : "", abs(val));
        } else if (newtype.is_string()) {
            // If it's a constant string, make the name $newconst_str_HASH
            // where HASH is the hex value of the hash of that ustring. This
            // ensures that the name will be the same regardless of the order
            // in which we made the constants (helpful for PTX cache
            // behavior).
            ustring val(*(const char**)data);
            symname = ustring::fmtformat("$newconst_str{:08x}", val.hash());
        } else {
            // All other cases, just name it sequentially as $newconst_1,
            // $newconst_2, etc.
            symname = ustring::fmtformat("$newconst{}", m_next_newconst++);
        }
        Symbol newconst(symname, newtype, SymTypeConst);
        void* newdata = nullptr;
        TypeDesc t(newtype.simpletype());
        size_t n = t.aggregate * t.numelements();
        if (datatype == TypeDesc::UNKNOWN)
            datatype = t;
        size_t datan = datatype.aggregate * datatype.numelements();
        if (t.basetype == TypeDesc::INT && datatype.basetype == TypeDesc::INT
            && n == datan) {
            newdata = inst()->shadingsys().alloc_int_constants(n);
            memcpy(newdata, data, t.size());
        } else if (t.basetype == TypeDesc::FLOAT
                   && datatype.basetype == TypeDesc::FLOAT) {
            newdata = inst()->shadingsys().alloc_float_constants(n);
            if (n == datan)
                for (size_t i = 0; i < n; ++i)
                    ((float*)newdata)[i] = ((const float*)data)[i];
            else if (datan == 1)
                for (size_t i = 0; i < n; ++i)
                    ((float*)newdata)[i] = ((const float*)data)[0];
            else {
                OSL_ASSERT(0 && "unsupported type for add_constant");
            }
        } else if (t.basetype == TypeDesc::FLOAT
                   && datatype.basetype == TypeDesc::INT) {
            newdata = inst()->shadingsys().alloc_float_constants(n);
            if (n == datan)
                for (size_t i = 0; i < n; ++i)
                    ((float*)newdata)[i] = ((const int*)data)[i];
            else if (datan == 1)
                for (size_t i = 0; i < n; ++i)
                    ((float*)newdata)[i] = ((const int*)data)[0];
            else {
                OSL_ASSERT(0 && "unsupported type for add_constant");
            }
        } else if (t.basetype == TypeDesc::STRING
                   && datatype.basetype == TypeDesc::STRING && n == datan) {
            newdata = inst()->shadingsys().alloc_string_constants(n);
            memcpy(newdata, data, t.size());
        } else {
            OSL_ASSERT(0 && "unsupported type for add_constant");
        }
        newconst.set_dataptr(SymArena::Absolute, newdata);
        ind = add_symbol(newconst);
        m_all_consts.push_back(ind);
    }
    return ind;
}



int
RuntimeOptimizer::add_temp(const TypeSpec& type)
{
    return add_symbol(Symbol(ustring::fmtformat("$opttemp{}", m_next_newtemp++),
                             type, SymTypeTemp));
}



int
RuntimeOptimizer::add_global(ustring name, const TypeSpec& type)
{
    int index = inst()->findsymbol(name);
    if (index < 0)
        index = add_symbol(Symbol(name, type, SymTypeGlobal));
    return index;
}



int
RuntimeOptimizer::add_symbol(const Symbol& sym)
{
    size_t index = inst()->symbols().size();
    OSL_ASSERT(inst()->symbols().capacity() > index
               && "we shouldn't have to realloc here");
    inst()->symbols().push_back(sym);
    // Mark the symbol as always read.  Next time we recompute symbol
    // lifetimes, it'll get the correct range for when it's read and
    // written.  But for now, just make sure it doesn't accidentally
    // look entirely unused.
    inst()->symbols().back().mark_always_used();
    return (int)index;
}



void
RuntimeOptimizer::debug_opt_impl(string_view message) const
{
    static OIIO::spin_mutex mutex;
    OIIO::spin_lock lock(mutex);
    std::cout << message;
}



void
RuntimeOptimizer::debug_opt_ops(int opbegin, int opend,
                                string_view message) const
{
    const Opcode& op(inst()->ops()[opbegin]);
    std::string oprange;
    if (opbegin >= 0 && opend - opbegin > 1)
        oprange = fmtformat("ops {}-{} ", opbegin, opend);
    else if (opbegin >= 0)
        oprange = fmtformat("op {} ", opbegin);
    debug_optfmt("  {}{} (@ {}:{})\n", oprange, message, op.sourcefile(),
                 op.sourceline());
}



void
RuntimeOptimizer::debug_turn_into(const Opcode& op, int numops,
                                  string_view newop, int newarg0, int newarg1,
                                  int newarg2, string_view why)
{
    int opnum = &op - &(inst()->ops()[0]);
    std::string msg;
    if (numops == 1)
        msg = fmtformat("turned '{}' to '{}", op_string(op), newop);
    else
        msg = fmtformat("turned to '{}", newop);
    if (newarg0 >= 0)
        msg += fmtformat(" {}", inst()->symbol(newarg0)->name());
    if (newarg1 >= 0)
        msg += fmtformat(" {}", inst()->symbol(newarg1)->name());
    if (newarg2 >= 0)
        msg += fmtformat(" {}", inst()->symbol(newarg2)->name());
    msg += "'";
    if (why.size())
        msg += fmtformat(" : {}", why);
    debug_opt_ops(opnum, opnum + numops, msg);
}



void
RuntimeOptimizer::turn_into_new_op(Opcode& op, ustring newop, int newarg0,
                                   int newarg1, int newarg2, string_view why)
{
    int opnum = &op - &(inst()->ops()[0]);
    OSL_DASSERT(opnum >= 0 && opnum < (int)inst()->ops().size());
    if (debug() > 1)
        debug_turn_into(op, 1, newop, newarg0, newarg1, newarg2, why);
    op.reset(newop, newarg2 < 0 ? 2 : 3);
    inst()->args()[op.firstarg() + 0] = newarg0;
    op.argwriteonly(0);
    opargsym(op, 0)->mark_rw(opnum, false, true);
    inst()->args()[op.firstarg() + 1] = newarg1;
    op.argreadonly(1);
    opargsym(op, 1)->mark_rw(opnum, true, false);
    if (newarg2 >= 0) {
        inst()->args()[op.firstarg() + 2] = newarg2;
        op.argreadonly(2);
        opargsym(op, 2)->mark_rw(opnum, true, false);
    }
}



void
RuntimeOptimizer::turn_into_assign(Opcode& op, int newarg, string_view why)
{
    // We don't know the op num here, so we subtract the pointers
    int opnum = &op - &(inst()->ops()[0]);
    if (debug() > 1)
        debug_turn_into(op, 1, "assign", oparg(op, 0), newarg, -1, why);
    op.reset(u_assign, 2);
    inst()->args()[op.firstarg() + 1] = newarg;
    op.argwriteonly(0);
    op.argread(1, true);
    op.argwrite(1, false);
    // Need to make sure the symbol we're assigning is marked as read
    // for this op.
    OSL_DASSERT(opnum >= 0 && opnum < (int)inst()->ops().size());
    Symbol* arg = opargsym(op, 1);
    arg->mark_rw(opnum, true, false);
}



// Turn the current op into a simple assignment to zero (of the first arg).
void
RuntimeOptimizer::turn_into_assign_zero(Opcode& op, string_view why)
{
    static float zero[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
    Symbol& R(*(inst()->argsymbol(op.firstarg() + 0)));
    int cind = add_constant(R.typespec(), &zero);
    turn_into_assign(op, cind, why);
}



// Turn the current op into a simple assignment to one (of the first arg).
void
RuntimeOptimizer::turn_into_assign_one(Opcode& op, string_view why)
{
    Symbol& R(*(inst()->argsymbol(op.firstarg() + 0)));
    if (R.typespec().is_int()) {
        int one  = 1;
        int cind = add_constant(R.typespec(), &one);
        turn_into_assign(op, cind, why);
    } else {
        OSL_DASSERT(R.typespec().is_triple() || R.typespec().is_float());
        static float one[3] = { 1, 1, 1 };
        int cind            = add_constant(R.typespec(), &one);
        turn_into_assign(op, cind, why);
    }
}



// Turn the op into a no-op
int
RuntimeOptimizer::turn_into_nop(Opcode& op, string_view why)
{
    if (op.opname() != u_nop) {
        if (debug() > 1)
            debug_turn_into(op, 1, "nop", -1, -1, -1, why);
        op.reset(u_nop, 0);
        return 1;
    }
    return 0;
}



int
RuntimeOptimizer::turn_into_nop(int begin, int end, string_view why)
{
    int changed = 0;
    for (int i = begin; i < end; ++i) {
        Opcode& op(inst()->ops()[i]);
        if (op.opname() != u_nop) {
            op.reset(u_nop, 0);
            ++changed;
        }
    }
    if (debug() > 1 && changed)
        debug_turn_into(inst()->ops()[begin], end - begin, "nop", -1, -1, -1,
                        why);
    return changed;
}

// Turn the op into a no-op functioncall
// We keep want to keep the jumps indices so we can correctly
// model an inlined function call for the debugger
int
RuntimeOptimizer::turn_into_functioncall_nr(Opcode& op, string_view why)
{
    if (op.opname() == u_functioncall) {
        if (debug() > 1)
            debug_turn_into(op, 1, "functioncall_nr", -1, -1, -1, why);
        op.transmute_opname(u_functioncall_nr);
        return 1;
    }
    return 0;
}


void
RuntimeOptimizer::insert_code(int opnum, ustring opname,
                              const cspan<int> args_to_add,
                              RecomputeRWRangesOption recompute_rw_ranges,
                              InsertRelation relation)
{
    OpcodeVec& code(inst()->ops());
    std::vector<int>& opargs(inst()->args());
    ustring method = (opnum < (int)code.size()) ? code[opnum].method()
                                                : main_method_name;
    int nargs      = args_to_add.size();
    Opcode op(opname, method, opargs.size(), nargs);
    code.insert(code.begin() + opnum, op);
    opargs.insert(opargs.end(), args_to_add.begin(), args_to_add.end());
    if (opnum < inst()->m_maincodebegin)
        ++inst()->m_maincodebegin;
    ++inst()->m_maincodeend;
    if ((relation == -1 && opnum > 0)
        || (relation == 1 && opnum < (int)code.size() - 1)) {
        code[opnum].method(code[opnum + relation].method());
        code[opnum].source(code[opnum + relation].sourcefile(),
                           code[opnum + relation].sourceline());
    }

    // Unless we were inserting at the end, we may need to adjust
    // the jump addresses of other ops and the param init ranges.
    if (opnum < (int)code.size() - 1) {
        // Adjust jump offsets
        for (auto& c : code) {
            for (int j = 0; j < (int)Opcode::max_jumps && c.jump(j) >= 0; ++j) {
                if (c.jump(j) > opnum) {
                    c.jump(j) = c.jump(j) + 1;
                    // std::cerr << "Adjusting jump target at op " << n << "\n";
                }
            }
        }
        // Adjust param init ranges
        FOREACH_PARAM(auto&& s, inst())
        {
            if (s.initbegin() > opnum)
                s.initbegin(s.initbegin() + 1);
            if (s.initend() > opnum)
                s.initend(s.initend() + 1);
        }
    }

    // Inserting the instruction may change the read/write ranges of
    // symbols.  Not adjusting this can throw off other optimizations.
    if (recompute_rw_ranges) {
        for (auto&& s : inst()->symbols()) {
            if (s.everread()) {
                int first = s.firstread(), last = s.lastread();
                if (first >= opnum)
                    ++first;
                if (last >= opnum)
                    ++last;
                s.set_read(first, last);
            }
            if (s.everwritten()) {
                int first = s.firstwrite(), last = s.lastwrite();
                if (first >= opnum)
                    ++first;
                if (last >= opnum)
                    ++last;
                s.set_write(first, last);
            }
        }
    }

    // Adjust the basic block IDs and which instructions are inside
    // conditionals.
    if (m_bblockids.size()) {
        OSL_DASSERT(m_bblockids.size() == code.size() - 1);
        m_bblockids.insert(m_bblockids.begin() + opnum, 1, m_bblockids[opnum]);
    }
    if (m_in_conditional.size()) {
        OSL_DASSERT(m_in_conditional.size() == code.size() - 1);
        m_in_conditional.insert(m_in_conditional.begin() + opnum, 1,
                                m_in_conditional[opnum]);
    }
    if (m_in_loop.size()) {
        OSL_DASSERT(m_in_loop.size() == code.size() - 1);
        m_in_loop.insert(m_in_loop.begin() + opnum, 1, m_in_loop[opnum]);
    }
    // If the first return happened after this, bump it up
    if (m_first_return >= opnum)
        ++m_first_return;

    if (opname == u_if) {
        // special case for 'if' -- the arg is read, not written
        inst()->symbol(args_to_add[0])->mark_rw(opnum, true, false);
    } else if (opname != u_useparam) {
        // Mark the args as being used for this op (assume that the
        // first is written, the others are read).
        for (int a = 0; a < nargs; ++a)
            inst()->symbol(args_to_add[a])->mark_rw(opnum, a > 0, a == 0);
    }
}

void
RuntimeOptimizer::insert_code(int opnum, ustring opname,
                              InsertRelation relation, int arg0, int arg1,
                              int arg2, int arg3)
{
    int args[4];
    int nargs = 0;
    if (arg0 >= 0)
        args[nargs++] = arg0;
    if (arg1 >= 0)
        args[nargs++] = arg1;
    if (arg2 >= 0)
        args[nargs++] = arg2;
    if (arg3 >= 0)
        args[nargs++] = arg3;
    insert_code(opnum, opname, cspan<int>(args, args + nargs),
                RecomputeRWRanges, relation);
}



/// Insert a 'useparam' instruction in front of instruction 'opnum', to
/// reference the symbols in 'params'.
void
RuntimeOptimizer::insert_useparam(size_t opnum,
                                  const std::vector<int>& params_to_use)
{
    OSL_DASSERT(params_to_use.size() > 0);
    OpcodeVec& code(inst()->ops());
    insert_code(opnum, u_useparam, params_to_use, RecomputeRWRanges,
                GroupWithNext);

    // All ops are "read"
    code[opnum].argwrite(0, false);
    code[opnum].argread(0, true);
    if (opnum < code.size() - 1) {
        // We have no parse node, but we set the new instruction's
        // "source" to the one of the statement right after.
        code[opnum].source(code[opnum + 1].sourcefile(),
                           code[opnum + 1].sourceline());
        // Set the method id to the same as the statement right after
        code[opnum].method(code[opnum + 1].method());
    } else {
        // If there IS no "next" instruction, just call it main
        code[opnum].method(main_method_name);
    }
}



void
RuntimeOptimizer::add_useparam(SymbolPtrVec& allsyms)
{
    OpcodeVec& code(inst()->ops());
    std::vector<int>& opargs(inst()->args());

    // Mark all symbols as un-initialized
    for (auto&& s : inst()->symbols())
        s.initialized(false);

    if (inst()->m_maincodebegin < 0)
        inst()->m_maincodebegin = (int)code.size();

    // Take care of the output params right off the bat -- as soon as the
    // shader starts running 'main'.
    std::vector<int> outputparams;
    for (int i = 0; i < (int)inst()->symbols().size(); ++i) {
        Symbol* s = inst()->symbol(i);
        if (s->symtype() == SymTypeOutputParam
            && (s->connected() || s->connected_down() || s->renderer_output()
                || (s->valuesource() == Symbol::DefaultVal
                    && s->has_init_ops()))) {
            outputparams.push_back(i);
            s->initialized(true);
        }
    }
    if (outputparams.size())
        insert_useparam(inst()->m_maincodebegin, outputparams);

    // Figure out which statements are inside conditional states
    find_conditionals();

    // Loop over all ops...
    for (int opnum = 0; opnum < (int)code.size(); ++opnum) {
        Opcode& op(code[opnum]);  // handy ref to the op
        if (op.opname() == u_useparam)
            continue;  // skip useparam ops themselves, if we hit one
        bool simple_assign = is_simple_assign(op);
        bool in_main_code  = (opnum >= inst()->m_maincodebegin);
        std::vector<int> params;  // list of params referenced by this op
        // For each argument...
        for (int a = 0; a < op.nargs(); ++a) {
            int argind  = op.firstarg() + a;
            SymbolPtr s = inst()->argsymbol(argind);
            OSL_DASSERT(s->dealias() == s);
            // If this arg is a param and is read, remember it
            if (s->symtype() != SymTypeParam
                && s->symtype() != SymTypeOutputParam)
                continue;  // skip non-params
            // skip if we've already 'usedparam'ed it unconditionally
            if (s->initialized() && in_main_code)
                continue;

            bool inside_init = (opnum >= s->initbegin()
                                && opnum < s->initend());
            if (op.argread(a) || (op.argwrite(a) && !inside_init)) {
                // Don't add it more than once
                if (std::find(params.begin(), params.end(), opargs[argind])
                    == params.end()) {
                    // If this arg is the one being written to by a
                    // "simple" assignment, it doesn't need a useparam here.
                    if (!(simple_assign && a == 0))
                        params.push_back(opargs[argind]);
                    // mark as already initialized unconditionally, if we do
                    if (op_is_unconditionally_executed(opnum)
                        && op.method() == main_method_name)
                        s->initialized(true);
                }
            }
        }

        // If the arg we are examining read any params, insert a "useparam"
        // op whose arguments are the list of params we are about to use.
        if (params.size()) {
            insert_useparam(opnum, params);
            // Skip the op we just added
            ++opnum;
        }
    }

    // Mark all symbols as un-initialized
    for (auto&& s : inst()->symbols())
        s.initialized(false);

    // Re-track variable lifetimes, since the inserted useparam
    // instructions will have change the instruction numbers.
    find_basic_blocks();
    track_variable_lifetimes(allsyms);
}



bool
OSOProcessorBase::is_zero(const Symbol& A)
{
    if (!A.is_constant())
        return false;
    const TypeSpec& Atype(A.typespec());
    static Vec3 Vzero(0, 0, 0);
    return (Atype.is_float() && A.get_float() == 0)
           || (Atype.is_int() && A.get_int() == 0)
           || (Atype.is_triple() && A.get_vec3() == Vzero);
}



bool
OSOProcessorBase::is_one(const Symbol& A)
{
    if (!A.is_constant())
        return false;
    const TypeSpec& Atype(A.typespec());
    static Vec3 Vone(1, 1, 1);
    static Matrix44 Mone(1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1);
    return (Atype.is_float() && A.get_float() == 1)
           || (Atype.is_int() && A.get_int() == 1)
           || (Atype.is_triple() && A.get_vec3() == Vone)
           || (Atype.is_matrix() && *(const Matrix44*)A.dataptr() == Mone);
}



bool
OSOProcessorBase::is_nonzero(const Symbol& A)
{
    if (!A.is_constant())
        return false;
    const TypeSpec& Atype(A.typespec());
    int ncomponents = Atype.numelements() * Atype.aggregate();
    if (Atype.is_float_based()) {
        const float* val = (const float*)A.data();
        for (int i = 0; i < ncomponents; ++i)
            if (val[i] == 0.0f)
                return false;
        return true;
    }
    if (Atype.is_int_based()) {
        const int* val = (const int*)A.data();
        for (int i = 0; i < ncomponents; ++i)
            if (val[i] == 0)
                return false;
        return true;
    }
    return false;
}



std::string
OSOProcessorBase::const_value_as_string(const Symbol& A)
{
    if (!A.is_constant())
        return std::string();
    TypeDesc type(A.typespec().simpletype());
    int n = type.numelements() * type.aggregate;
    std::ostringstream s;
    s.imbue(std::locale::classic());  // force C locale
    if (type.basetype == TypeDesc::FLOAT) {
        for (int i = 0; i < n; ++i)
            s << (i ? "," : "") << A.get_float(i);
    } else if (type.basetype == TypeDesc::INT) {
        for (int i = 0; i < n; ++i)
            s << (i ? "," : "") << A.get_int(i);
    } else if (type.basetype == TypeDesc::STRING) {
        for (int i = 0; i < n; ++i)
            s << (i ? "," : "") << '\"' << A.get_string(i) << '\"';
    }
    return s.str();
}



void
RuntimeOptimizer::register_message(ustring name)
{
    m_local_messages_sent.push_back(name);
}



void
RuntimeOptimizer::register_unknown_message()
{
    m_local_unknown_message_sent = true;
}



bool
RuntimeOptimizer::message_possibly_set(ustring name) const
{
    return m_local_unknown_message_sent || m_unknown_message_sent
           || std::find(m_messages_sent.begin(), m_messages_sent.end(), name)
                  != m_messages_sent.end()
           || std::find(m_local_messages_sent.begin(),
                        m_local_messages_sent.end(), name)
                  != m_local_messages_sent.end();
}



/// For all the instance's parameters (that can't be overridden by the
/// geometry), if they can be found to be effectively constants or
/// globals, make constants for them and alias them to the constant. If
/// they are connected to an earlier layer's output, if it can determine
/// that the output will be a constant or global, then sever the
/// connection and just alias our parameter to that value.
void
RuntimeOptimizer::simplify_params()
{
    for (int i = inst()->firstparam(); i < inst()->lastparam(); ++i) {
        Symbol* s(inst()->symbol(i));
        if (s->symtype() != SymTypeParam)
            continue;  // Skip non-params
        // Don't simplify params that are interpolated or interactively
        // editable
        if (s->interpolated() || s->interactive())
            continue;
        if (s->typespec().is_structure() || s->typespec().is_closure_based())
            continue;  // We don't mess with struct placeholders or closures

        if (s->valuesource() == Symbol::InstanceVal) {
            // Instance value -- turn it into a constant and remove init ops
            make_symbol_room(1);
            s = inst()->symbol(i);  // In case make_symbol_room changed ptrs
            int cind = add_constant(s->typespec(), s->data());
            global_alias(i, cind);  // Alias this symbol to the new const
            turn_into_nop(s->initbegin(), s->initend(),
                          "instance value doesn't need init ops");
        } else if (s->valuesource() == Symbol::DefaultVal
                   && !s->has_init_ops()) {
            // Plain default value without init ops -- turn it into a constant
            make_symbol_room(1);
            s = inst()->symbol(i);  // In case make_symbol_room changed ptrs
            int cind = add_constant(s->typespec(), s->data(),
                                    s->typespec().simpletype());
            global_alias(i, cind);  // Alias this symbol to the new const
        } else if (s->valuesource() == Symbol::DefaultVal
                   && s->has_init_ops()) {
            // Default val comes from init ops -- special cases?  Yes,
            // if it's a simple assignment from a global whose value is
            // not reassigned later, we can just alias it, and if we're
            // lucky that may eliminate all uses of the parameter.

            // First, trim init ops in case nops have accumulated
            while (s->has_init_ops() && op(s->initbegin()).opname() == u_nop)
                s->initbegin(s->initbegin() + 1);
            while (s->has_init_ops() && op(s->initend() - 1).opname() == u_nop)
                s->initend(s->initend() - 1);
            if (s->initbegin() == s->initend() - 1) {  // just one op
                Opcode& op(inst()->ops()[s->initbegin()]);
                if (op.opname() == u_assign) {
                    // The default value has init ops, but they consist of
                    // just a single assignment op...
                    Symbol* src = inst()->argsymbol(op.firstarg() + 1);
                    // Is it assigning a global, or a parameter that's
                    // got a default or instance value and isn't on the geom,
                    // and its value is never changed and the types match?
                    if ((src->symtype() == SymTypeGlobal
                         || src->symtype() == SymTypeConst
                         || (src->symtype() == SymTypeParam && src->lockgeom()
                             && (src->valuesource() == Symbol::DefaultVal
                                 || src->valuesource() == Symbol::InstanceVal)))
                        && !src->everwritten()
                        && equivalent(src->typespec(), s->typespec())) {
                        // Great, so let's remember the alias.  We can't
                        // call global_alias() here, because we're still in
                        // init ops, that'll screw us up.  So we just record
                        // it in m_param_aliases and then we'll establish
                        // the global aliases when we hit the main code.
                        m_param_aliases[i] = inst()->arg(op.firstarg() + 1);
                    }
                }
            }
        } else if (s->valuesource() == Symbol::ConnectedVal) {
            // It's connected to an earlier layer.  If the output var of
            // the upstream shader is effectively constant or a global,
            // then so is this variable.
            for (auto&& c : inst()->connections()) {
                if (c.dst.param != i)
                    continue;
                if (c.dst.is_complete()) {
                    /// All components are being set through either
                    /// float->triple or triple->triple
                    /// Get rid of the un-needed init ops.
                    turn_into_nop(s->initbegin(), s->initend(),
                                  "connected value doesn't need init ops");
                }
                if (c.is_complete()) {
                    // srcsym is the earlier group's output param, which
                    // is fully connected as the input to the param we're
                    // examining.
                    ShaderInstance* uplayer = group()[c.srclayer];
                    Symbol* srcsym          = uplayer->symbol(c.src.param);
                    if (srcsym->interpolated())
                        continue;  // Not if it can be overridden by geometry

                    // Is the source symbol known to be a global, from
                    // earlier analysis by find_params_holding_globals?
                    // If so, make sure the global is in this instance's
                    // symbol table, and alias the parameter to it.
                    ustringmap_t& g(m_params_holding_globals[c.srclayer]);
                    auto f = g.find(srcsym->name());
                    if (f != g.end()) {
                        if (debug() > 1)
                            debug_optfmt(
                                "Remapping {}.{} because it's connected to "
                                "{}.{}, which is known to be {}\n",
                                inst()->layername(), s->name(),
                                uplayer->layername(), srcsym->name(),
                                f->second);
                        make_symbol_room(1);
                        s = inst()->symbol(
                            i);  // In case make_symbol_room changed ptrs
                        int ind = add_global(f->second, srcsym->typespec());
                        global_alias(i, ind);
                        shadingsys().m_stat_global_connections += 1;
                        break;
                    }

                    if (!srcsym->everwritten()
                        && (srcsym->valuesource() == Symbol::DefaultVal
                            || srcsym->valuesource() == Symbol::InstanceVal)
                        && !srcsym->has_init_ops()) {
                        make_symbol_room(1);
                        s = inst()->symbol(
                            i);  // In case make_symbol_room changed ptrs
                        int cind = add_constant(s->typespec(), srcsym->data(),
                                                srcsym->typespec().simpletype());
                        // Alias this symbol to the new const
                        global_alias(i, cind);
                        make_param_use_instanceval(
                            s, "- upstream layer sets it to a constant");
                        replace_param_value(s, srcsym->data(),
                                            srcsym->typespec());
                        shadingsys().m_stat_const_connections += 1;
                        break;
                    }
                }
            }
            // FIXME / N.B.: We only optimize "fully complete" connections,
            // not those involving individual components or array elements
            // of the connected parameters, because we sure don't track the
            // constness or aliasing of individual components/element, only
            // whole variables. But there are two cases where the logic
            // above fails to fully exploit the connection propagating a
            // constant value. (a) Partial-to-whole connections, for example
            // connecting one component of an upstream triple output to a
            // downstream float input, should propagate the constant, but we
            // currently neglect this case. (b) If *multiple* connections
            // combine to fully propagate values, for example if someone was
            // foolish enough to connect R, G, and B components of color
            // parameters *separately*, we sure don't notice that and treat
            // it as a full connection of the color.
        }
    }
}



/// For all the instance's parameters, if they are simply assigned globals,
/// record that in m_params_holding_globals.
void
RuntimeOptimizer::find_params_holding_globals()
{
    FOREACH_PARAM(auto&& s, inst())
    {
        // Skip if this isn't a shader output parameter that's connected
        // to a later layer.
        if (s.symtype() != SymTypeParam && s.symtype() != SymTypeOutputParam)
            continue;  // Skip non-params
        if (!s.connected_down())
            continue;  // Skip unconnected params -- who cares
        if (s.valuesource() != Symbol::DefaultVal)
            continue;  // Skip -- must be connected or an instance value
        if (s.firstwrite() < 0 || s.firstwrite() != s.lastwrite())
            continue;  // Skip -- written more than once

        int opnum = s.firstwrite();
        Opcode& op(inst()->ops()[opnum]);
        if (op.opname() != u_assign || !op_is_unconditionally_executed(opnum))
            continue;  // Not a simple assignment unconditionally performed

        // what s is assigned from (fully dealiased)
        Symbol* src = inst()->symbol(dealias_symbol(oparg(op, 1), opnum));

        if (src->symtype() != SymTypeGlobal)
            continue;  // only interested in global assignments

        if (debug() > 1)
            debug_optfmt("I think that {}.{} will always be {}\n",
                         inst()->layername(), s.name(), src->name());
        m_params_holding_globals[layer()][s.name()] = src->name();
    }
}



void
OSOProcessorBase::find_conditionals()
{
    OpcodeVec& code(inst()->ops());

    m_in_conditional.clear();
    m_in_conditional.resize(code.size(), false);
    m_in_loop.clear();
    m_in_loop.resize(code.size(), false);
    m_first_return = (int)code.size();
    for (int i = 0; i < (int)code.size(); ++i) {
        if (code[i].jump(0) >= 0) {
            std::fill(m_in_conditional.begin() + i,
                      m_in_conditional.begin() + code[i].farthest_jump(), true);
            if (code[i].opname() == Strings::op_dowhile
                || code[i].opname() == Strings::op_for
                || code[i].opname() == Strings::op_while) {
                std::fill(m_in_loop.begin() + i,
                          m_in_loop.begin() + code[i].farthest_jump(), true);
            }
        }
        if (code[i].opname() == Strings::op_exit)
            m_first_return = std::min(m_first_return, i);
    }
}



void
OSOProcessorBase::find_basic_blocks()
{
    OpcodeVec& code(inst()->ops());

    // Start by setting all basic block IDs to 0
    m_bblockids.clear();
    m_bblockids.resize(code.size(), 0);

    // First, keep track of all the spots where blocks begin
    std::vector<bool> block_begin(code.size(), false);

    // Init ops start basic blocks
    FOREACH_PARAM(const Symbol& s, inst())
    {
        if (s.has_init_ops())
            block_begin[s.initbegin()] = true;
    }

    // Main code starts a basic block
    block_begin[inst()->maincodebegin()] = true;

    for (size_t opnum = 0; opnum < code.size(); ++opnum) {
        Opcode& op(code[opnum]);
        if (op.opname()
            == u_functioncall_nr) {  // Treat the 'no return' function call as if it were a nop.
            // we use later to generate correct inline debug information.
            continue;
        }
        // Anyplace that's the target of a jump instruction starts a basic block
        for (int j = 0; j < (int)Opcode::max_jumps; ++j) {
            if (op.jump(j) >= 0)
                block_begin[op.jump(j)] = true;
            else
                break;
        }
        // The first instruction in a conditional or loop (which is not
        // itself a jump target) also begins a basic block.  If the op has
        // any jump targets at all, it must be a conditional or loop.
        if (op.jump(0) >= 0)
            block_begin[opnum + 1] = true;
        // 'break', 'continue', 'return', and 'exit' also cause the next
        // statement to begin a new basic block.
        if (op.opname() == u_break || op.opname() == u_continue
            || op.opname() == u_return || op.opname() == u_exit)
            block_begin[opnum + 1] = true;
    }

    // Now color the blocks with unique identifiers
    int bbid = 1;  // next basic block ID to use
    for (size_t opnum = 0; opnum < code.size(); ++opnum) {
        if (block_begin[opnum])
            ++bbid;
        m_bblockids[opnum] = bbid;
    }
}



/// For 'R = A_const' where R and A are different, but coerceable,
/// types, turn it into a constant assignment of the exact type.
/// Return true if a change was made, otherwise return false.
bool
RuntimeOptimizer::coerce_assigned_constant(Opcode& op)
{
    OSL_DASSERT(op.opname() == u_assign);
    Symbol* R(inst()->argsymbol(op.firstarg() + 0));
    Symbol* A(inst()->argsymbol(op.firstarg() + 1));

    if (!A->is_constant() || R->typespec().is_closure_based())
        return false;  // we don't handle those cases

    // turn 'R_float = A_int_const' into a float const assignment
    if (A->typespec().is_int() && R->typespec().is_float()) {
        int cind = add_constant(A->coerce_float());
        turn_into_assign(op, cind, "coerce to correct type");
        return true;
    }

    // turn 'R_int = A_float_const' into an int const assignment
    if (A->typespec().is_float() && R->typespec().is_int()) {
        int cind = add_constant(static_cast<int>(A->get_float()));
        turn_into_assign(op, cind, "coerce to correct type");
        return true;
    }

    // turn 'R_triple = A_int_or_float_const' into a float const assignment
    if (A->typespec().is_int_or_float() && R->typespec().is_triple()) {
        float f = A->coerce_float();
        Vec3 result(f, f, f);
        int cind = add_constant(R->typespec(), &result);
        turn_into_assign(op, cind, "coerce to correct type");
        return true;
    }

    // Turn 'R_triple = A_other_triple_constant' into a triple const assign
    if (A->typespec().is_triple() && R->typespec().is_triple()
        && A->typespec() != R->typespec()) {
        const Vec3* f = &A->get_vec3();
        int cind      = add_constant(R->typespec(), f);
        turn_into_assign(op, cind, "coerce to correct type");
        return true;
    }

    // turn 'R_matrix = A_int_or_float_const' into a matrix const assignment
    if (A->typespec().is_int_or_float() && R->typespec().is_matrix()) {
        float f = A->coerce_float();
        Matrix44 result(f, 0, 0, 0, 0, f, 0, 0, 0, 0, f, 0, 0, 0, 0, f);
        int cind = add_constant(R->typespec(), &result);
        turn_into_assign(op, cind, "coerce to correct type");
        return true;
    }

    return false;
}



void
RuntimeOptimizer::clear_stale_syms()
{
    m_stale_syms.clear();
}



void
RuntimeOptimizer::use_stale_sym(int sym)
{
    FastIntMap::iterator i = m_stale_syms.find(sym);
    if (i != m_stale_syms.end())
        m_stale_syms.erase(i);
}



bool
RuntimeOptimizer::is_simple_assign(Opcode& op, const OpDescriptor* opd)
{
    // Simple only if arg0 is the only write, and is write only.
    if (op.argwrite_bits() != 1 || op.argread(0))
        return false;
    if (!opd)
        opd = shadingsys().op_descriptor(op.opname());
    if (!opd || !opd->simple_assign)
        return false;  // reject all other known non-simple assignments
    // Make sure the result isn't also read
    int result = oparg(op, 0);
    for (int i = 1, e = op.nargs(); i < e; ++i)
        if (oparg(op, i) == result)
            return false;
    return true;
}



void
RuntimeOptimizer::simple_sym_assign(int sym, int opnum)
{
    if (optimize() >= 2 && m_opt_stale_assign) {
        FastIntMap::iterator i = m_stale_syms.find(sym);
        if (i != m_stale_syms.end()) {
            Opcode& uselessop(inst()->ops()[i->second]);
            if (uselessop.opname() != u_nop
                && uselessop.opname() != u_functioncall_nr)
                turn_into_nop(
                    uselessop,
                    debug() > 1
                        ? fmtformat(
                              "remove stale value assignment to {}, reassigned on op {}",
                              opargsym(uselessop, 0)->name(), opnum)
                              .c_str()
                        : "");
        }
    }
    m_stale_syms[sym] = opnum;
}



bool
RuntimeOptimizer::unread_after(const Symbol* A, int opnum)
{
    // Try to figure out if this symbol is completely unused after this
    // op (and thus, any values written to it now will never be needed).

    // Globals may be read by later layers
    if (A->symtype() == SymTypeGlobal)
        return false;

    // Params may be read afterwards if connected to a downstream
    // layer or if "elide_unconnected_outputs" is turned off.
    if (A->symtype() == SymTypeOutputParam || A->symtype() == SymTypeParam) {
        if (!m_opt_elide_unconnected_outputs)
            return false;  // Asked not do do this optimization
        if (A->connected_down())
            return false;  // Connected to something downstream
        if (A->renderer_output())
            return false;  // This is a renderer output -- don't cull it
    }

    // For all else, check if it's either never read at all in this
    // layer or it's only read earlier and we're not part of a loop
    return !A->everread() || (A->lastread() <= opnum && !m_in_loop[opnum]);
}



void
RuntimeOptimizer::replace_param_value(Symbol* R, const void* newdata,
                                      const TypeSpec& newdata_type)
{
    OSL_DASSERT(R->symtype() == SymTypeParam
                || R->symtype() == SymTypeOutputParam);
    TypeDesc Rtype = R->typespec().simpletype();
    OSL_DASSERT(R->dataoffset() >= 0);
    int Rnvals     = int(Rtype.aggregate * Rtype.numelements());
    TypeDesc Ntype = newdata_type.simpletype();
    if (Ntype == TypeDesc::UNKNOWN)
        Ntype = Rtype;
    int Nnvals = int(Ntype.aggregate * Ntype.numelements());
    if (Rtype.basetype == TypeDesc::FLOAT
        && Ntype.basetype == TypeDesc::FLOAT) {
        float* Rdefault = &inst()->m_fparams[R->dataoffset()];
        OSL_DASSERT((R->dataoffset() + Rnvals)
                    <= (int)inst()->m_fparams.size());
        if (Rnvals == Nnvals)  // straight copy
            for (int i = 0; i < Rnvals; ++i)
                Rdefault[i] = ((const float*)newdata)[i];
        else if (Nnvals == 1)  // scalar -> aggregate, by replication
            for (int i = 0; i < Rnvals; ++i)
                Rdefault[i] = ((const float*)newdata)[0];
        else {
            OSL_ASSERT(0 && "replace_param_value: unexpected types");
        }
    } else if (Rtype.basetype == TypeDesc::FLOAT
               && Ntype.basetype == TypeDesc::INT) {
        // Careful, this is an int-to-float conversion
        float* Rdefault = &inst()->m_fparams[R->dataoffset()];
        OSL_DASSERT((R->dataoffset() + Rnvals)
                    <= (int)inst()->m_fparams.size());
        if (Rnvals == Nnvals)  // straight copy
            for (int i = 0; i < Rnvals; ++i)
                Rdefault[i] = ((const int*)newdata)[i];
        else if (Nnvals == 1)  // scalar -> aggregate, by replication
            for (int i = 0; i < Rnvals; ++i)
                Rdefault[i] = ((const int*)newdata)[0];
        else {
            OSL_ASSERT(0 && "replace_param_value: unexpected types");
        }
    } else if (Rtype.basetype == TypeDesc::INT
               && Ntype.basetype == TypeDesc::INT && Rnvals == Nnvals) {
        int* Rdefault = &inst()->m_iparams[R->dataoffset()];
        OSL_DASSERT((R->dataoffset() + Rnvals)
                    <= (int)inst()->m_iparams.size());
        for (int i = 0; i < Rnvals; ++i)
            Rdefault[i] = ((const int*)newdata)[i];
    } else if (Rtype.basetype == TypeDesc::STRING
               && Ntype.basetype == TypeDesc::STRING && Rnvals == Nnvals) {
        ustring* Rdefault = &inst()->m_sparams[R->dataoffset()];
        OSL_DASSERT((R->dataoffset() + Rnvals)
                    <= (int)inst()->m_sparams.size());
        for (int i = 0; i < Rnvals; ++i)
            Rdefault[i] = ((const ustring*)newdata)[i];
    } else {
        OSL_ASSERT(0 && "replace_param_value: unexpected types");
    }
}



// Predicate to test if the connection's destination is never used
struct ConnectionDestIs {
    ConnectionDestIs(const ShaderInstance& inst, const Symbol* sym)
        : m_inst(inst), m_sym(sym)
    {
    }
    bool operator()(const Connection& c)
    {
        return m_inst.symbol(c.dst.param) == m_sym;
    }

private:
    const ShaderInstance& m_inst;
    const Symbol* m_sym;
};



/// Symbol R in the current instance has a connection or init ops we
/// no longer need; turn it into a a plain old instance-value
/// parameter.
void
RuntimeOptimizer::make_param_use_instanceval(Symbol* R, string_view why)
{
    if (debug() > 1)
        std::cout << "Turning " << R->valuesourcename() << ' ' << R->typespec()
                  << ' ' << R->name() << " into an instance value " << why
                  << "\n";

    // Mark its source as the instance value, not connected
    R->valuesource(Symbol::InstanceVal);
    // If it isn't a connection or computed, it doesn't need derivs.
    R->has_derivs(false);

    // Point the symbol's data pointer to its instance value
    // uniform
    void* Rdefault = NULL;
    OSL_DASSERT(R->dataoffset() >= 0);
    TypeDesc Rtype = R->typespec().simpletype();
    if (Rtype.basetype == TypeDesc::FLOAT)
        Rdefault = &inst()->m_fparams[R->dataoffset()];
    else if (Rtype.basetype == TypeDesc::INT)
        Rdefault = &inst()->m_iparams[R->dataoffset()];
    else if (Rtype.basetype == TypeDesc::STRING)
        Rdefault = &inst()->m_sparams[R->dataoffset()];
    OSL_DASSERT(Rdefault != NULL);
    R->set_dataptr(SymArena::Absolute, Rdefault);

    // Get rid of any init ops
    if (R->has_init_ops()) {
        turn_into_nop(R->initbegin(), R->initend(), "init ops not needed");
        R->initbegin(0);
        R->initend(0);
    }
    // Erase R's incoming connections
    erase_if(inst()->connections(), ConnectionDestIs(*inst(), R));
}



/// Check for conditions under which assignments to output parameters
/// can be removed.
///
/// Return true if the assignment is removed entirely.
bool
RuntimeOptimizer::outparam_assign_elision(int opnum, Opcode& op)
{
    OSL_DASSERT(op.opname() == u_assign);
    Symbol* R(inst()->argsymbol(op.firstarg() + 0));
    Symbol* A(inst()->argsymbol(op.firstarg() + 1));

    if (R->symtype() != SymTypeOutputParam)
        return false;  // This logic is only about output params

    // Check for assignment of output params that are written only once
    // in the whole shader -- on this statement -- and assigned a
    // constant, and the assignment is unconditional.  In that case,
    // just alias it to the constant from here on out.
    if (  // R is being assigned a constant of the right type:
        A->is_constant()
        && R->typespec() == A->typespec()
        // FIXME -- can this be equivalent() rather than == ?
        // and it's written only on this op, and unconditionally:
        && R->firstwrite() == opnum && R->lastwrite() == opnum
        && !m_in_conditional[opnum]
        // and this is not a case of an init op for an output param that
        // actually will get an instance value or a connection:
        && !((R->valuesource() == Symbol::InstanceVal || R->connected())
             && R->initbegin() <= opnum && R->initend() > opnum)) {
        // Alias it to the constant it's being assigned
        int cind = inst()->args()[op.firstarg() + 1];
        global_alias(inst()->args()[op.firstarg()], cind);
        // If it's also never read before this assignment and isn't a
        // designated renderer output (which we obviously must write!), just
        // replace its default value entirely and get rid of the assignment.
        if (R->firstread() > opnum && !R->renderer_output()
            && m_opt_elide_unconnected_outputs) {
            make_param_use_instanceval(
                R, fmtformat(
                       "- written once, with a constant ({}), before any reads",
                       const_value_as_string(*A)));
            replace_param_value(R, A->data(), A->typespec());
            turn_into_nop(
                op,
                debug() > 1
                    ? fmtformat("oparam {} never subsequently read or connected",
                                R->name())
                          .c_str()
                    : "");
            return true;
        }
    }

    // If the output param will neither be read later in the shader nor
    // connected to a downstream layer, then we don't really need this
    // assignment at all. Note that unread_after() does take into
    // consideration whether it's a renderer output.
    if (unread_after(R, opnum)) {
        turn_into_nop(
            op,
            debug() > 1
                ? fmtformat("oparam {} never subsequently read or connected",
                            R->name())
                      .c_str()
                : "");
        return true;
    }

    return false;
}



/// If every potentially-written argument to this op is NEVER read, turn
/// it into a nop and return true.  We don't do this to ops that have no
/// written args at all, since they tend to have side effects (e.g.,
/// printf, setmessage).
bool
RuntimeOptimizer::useless_op_elision(Opcode& op, int opnum)
{
    if (op.nargs()) {
        bool writes_something = false;
        for (int a = 0; a < op.nargs(); ++a) {
            if (op.argwrite(a)) {
                writes_something = true;
                Symbol* A        = opargsym(op, a);
                if (!unread_after(A, opnum))
                    return false;
            }
        }
        // If we get this far, nothing written had any effect
        if (writes_something) {
            // Enumerate exceptions -- ops that write something, but have
            // side effects that means they shouldn't be eliminated.
            if (op.opname() == u_pointcloud_write)
                return false;
            // It's a useless op, eliminate it
            turn_into_nop(op, "eliminated op whose writes will never be read");
            return true;
        }
    }
    return false;
}



int
RuntimeOptimizer::dealias_symbol(int symindex, int opnum)
{
    do {
        int i = block_alias(symindex);
        if (i >= 0) {
            // block-specific alias for the sym
            symindex = i;
            continue;
        }
        FastIntMap::const_iterator found;
        found = m_symbol_aliases.find(symindex);
        if (found != m_symbol_aliases.end()) {
            // permanent alias for the sym
            symindex = found->second;
            continue;
        }
        if (inst()->symbol(symindex)->symtype() == SymTypeParam
            && opnum >= inst()->maincodebegin()) {
            // Only check parameter aliases for main code
            found = m_param_aliases.find(symindex);
            if (found != m_param_aliases.end()) {
                symindex = found->second;
                continue;
            }
        }
    } while (0);
    return symindex;
}



void
RuntimeOptimizer::block_unalias(int symindex)
{
    m_block_aliases.erase(symindex);
    // In addition to the current block_aliases, unalias from any
    // saved alias lists.
    for (auto& ba : m_block_aliases_stack)
        ba->erase(symindex);
}



/// Make sure there's room for at least one more symbol, so that we can
/// add a const if we need to, without worrying about the addresses of
/// symbols changing if we add a new one soon.
void
RuntimeOptimizer::make_symbol_room(int howmany)
{
    inst()->make_symbol_room(howmany);
}



// Predicate to test if a symbol (specified by symbol index, symbol
// pointer, or by the inbound Connection record) is never used within
// the shader or passed along.  Subtlety: you can't base the test for
// params on sym->everused(), since of course it may be used within its
// own init ops, but then never subsequently used, and thus be a prime
// candidate for culling.  Instead, for params we test whether it was
// used at any point AFTER its init ops.
class SymNeverUsed {
public:
    SymNeverUsed(const RuntimeOptimizer& rop, const ShaderInstance* inst)
        : m_rop(rop), m_inst(inst)
    {
    }
    bool operator()(const Symbol& sym) const
    {
        if (sym.symtype() == SymTypeParam)
            return (sym.lastuse() < sym.initend()) && !sym.connected_down();
        if (sym.symtype() == SymTypeOutputParam) {
            if (!m_rop.opt_elide_unconnected_outputs())
                return false;  // Asked not to do this optimization
            if (sym.connected_down())
                return false;  // Connected to something downstream
            if (sym.renderer_output())
                return false;  // This is a renderer output
            return (sym.lastuse() < sym.initend());
        }
        return !sym.everused();  // all other symbol types
    }
    bool operator()(int symid) const { return (*this)(*m_inst->symbol(symid)); }
    bool operator()(const Connection& c) const { return (*this)(c.dst.param); }

private:
    const RuntimeOptimizer& m_rop;
    const ShaderInstance* m_inst;
};



int
RuntimeOptimizer::next_block_instruction(int opnum)
{
    int end = (int)inst()->ops().size();
    for (int n = opnum + 1; n < end && m_bblockids[n] == m_bblockids[opnum];
         ++n)
        if (inst()->ops()[n].opname() != u_nop
            && inst()->ops()[n].opname() != u_functioncall_nr)
            return n;  // Found it!
    return 0;          // End of ops or end of basic block
}



int
RuntimeOptimizer::peephole2(int opnum, int op2num)
{
    Opcode& op(inst()->ops()[opnum]);
    Opcode& next(inst()->ops()[op2num]);

    // N.B. Some of these transformations may look strange, you may
    // think "nobody will write code that does that", but (a) they do;
    // and (b) it can end up like that after other optimizations have
    // changed the code around.

    // Ping-pong assignments can eliminate the second one:
    //     assign a b
    //     assign b a    <-- turn into nop
    // But note that if a is an int and b is a float, this transformation
    // is not safe because of the intentional truncation.
    if (op.opname() == u_assign && next.opname() == u_assign) {
        Symbol* a = opargsym(op, 0);
        Symbol* b = opargsym(op, 1);
        Symbol* c = opargsym(next, 0);
        Symbol* d = opargsym(next, 1);
        if (a == d && b == c) {
            // Exclude the integer truncation case
            if (!(a->typespec().is_int() && b->typespec().is_float_based())) {
                // std::cerr << "ping-pong assignment " << opnum << " of "
                //           << opargsym(op,0)->mangled() << " and "
                //           << opargsym(op,1)->mangled() << "\n";
                turn_into_nop(next, "ping-pong assignments");
                return 1;
            }
        }
    }

    // Daisy chain assignments -> use common source
    //     assign a b
    //     assign c a
    // turns into:
    //     assign a b
    //     assign c b
    // This may allow a to be eliminated if it's not used elsewhere.
    // But note that this doesn't work for float = int = float,
    // which intentionally truncates before the assignment to c!
    if (op.opname() == u_assign && next.opname() == u_assign) {
        Symbol* a = opargsym(op, 0);
        Symbol* b = opargsym(op, 1);
        Symbol* c = opargsym(next, 0);
        Symbol* d = opargsym(next, 1);
        if (a == d && assignable(c->typespec(), b->typespec())) {
            // Exclude the float=int=float case
            if (!(a->typespec().is_int() && b->typespec().is_float_based()
                  && c->typespec().is_float_based()
                  && !c->typespec().is_array())) {
                turn_into_assign(next, inst()->arg(op.firstarg() + 1),
                                 "daisy-chain assignments");
                return 1;
            }
        }
    }

    // Look for adjacent add and subtract of the same value:
    //     add a a b
    //     sub a a b
    // (or vice versa)
    if (((op.opname() == u_add && next.opname() == u_sub)
         || (op.opname() == u_sub && next.opname() == u_add))
        && opargsym(op, 0) == opargsym(next, 0)
        && opargsym(op, 1) == opargsym(next, 1)
        && opargsym(op, 2) == opargsym(next, 2)
        && opargsym(op, 0) == opargsym(op, 1)) {
        // std::cerr << "dueling add/sub " << opnum << " & " << op2num << ": "
        //           << opargsym(op,0)->mangled() << "\n";
        turn_into_nop(op, "simplify add/sub pair");
        turn_into_nop(next, "simplify add/sub pair");
        return 2;
    }

    // Look for add of a value then subtract of the same value
    //     add a b c     or:    sub a b c
    //     sub d a c            add d a c
    // the second instruction should be changed to
    //     assign d b
    // and furthermore, if the only use of a is on these two lines or
    // if a == d, then the first instruction can be changed to a 'nop'.
    // Careful, "only used on these two lines" can be tricky if 'a' is a
    // global or output parameter, which are used after the shader finishes!
    if (((op.opname() == u_add && next.opname() == u_sub)
         || (op.opname() == u_sub && next.opname() == u_add))
        && opargsym(op, 0) == opargsym(next, 1)
        && opargsym(op, 2) == opargsym(next, 2)
        && opargsym(op, 0) != opargsym(next, 2) /* a != c */) {
        Symbol* a = opargsym(op, 0);
        Symbol* d = opargsym(next, 0);
        turn_into_assign(next, oparg(op, 1) /*b*/, "simplify add/sub pair");
        if ((a->firstuse() >= opnum && a->lastuse() <= op2num
             && ((a->symtype() != SymTypeGlobal
                  && a->symtype() != SymTypeOutputParam)))
            || a == d) {
            turn_into_nop(op, "simplify add/sub pair");
            return 2;
        } else
            return 1;
    }

    // Look for simple functions followed by an assignment:
    //    OP a b...
    //    assign c a
    // If OP is "simple" (completely overwrites its first argument, only
    // reads the rest), and a and c are the same type, and a is never
    // used again, then we can replace those two instructions with:
    //    OP c b...
    // Careful, "never used again" can be tricky if 'a' is a global or
    // output parameter, which are used after the shader finishes!
    if (next.opname() == u_assign && op.nargs() >= 1
        && opargsym(op, 0) == opargsym(next, 1) && is_simple_assign(op)) {
        Symbol* a = opargsym(op, 0);
        Symbol* c = opargsym(next, 0);
        if (a->firstuse() >= opnum && a->lastuse() <= op2num
            && (a->symtype() != SymTypeGlobal
                && a->symtype() != SymTypeOutputParam)
            && equivalent(a->typespec(), c->typespec())) {
            if (debug() > 1)
                debug_opt_ops(
                    opnum, opnum + 1,
                    fmtformat(
                        "turned '{} {}...' to '{} {}...' as part of daisy-chain",
                        op.opname(), a->name(), op.opname(), c->name()));
            inst()->args()[op.firstarg()] = inst()->args()[next.firstarg()];
            c->mark_rw(opnum, false, true);
            // Any time we write to a variable that wasn't written to at
            // this op previously, we need to block_unalias it, or it
            // can dealias to the wrong thing when examining subsequent
            // instructions.
            block_unalias(oparg(op, 0));  // clear any aliases
            turn_into_nop(next, "daisy-chain op and assignment");
            return 2;
        }
    }

    // Convert this combination
    //     closure A name arg...
    //     mul B A weight
    // into
    //     closure B C name arg...
    // That is, collapse a creation and immediate scale of a closure into
    // a single closure-with-scale constructor. (Valid if A is not used
    // elsewhere.)  Further refinement: if weight = 1, no need to do
    // the scale, and if weight == 0, eliminate the work entirely.
    // We only do this optimization on pass > 1, to give a fair chance
    // for other optimizations to be able to turn the weight into a
    // constant before we do this one (since if it's 1 or 0, we can
    // simplify further).
    if (op.opname() == u_closure && next.opname() == u_mul
        && optimization_pass() > 1) {
        Symbol* a      = opargsym(op, 0);
        Symbol* name   = opargsym(op, 1);
        Symbol* aa     = opargsym(next, 1);
        Symbol* weight = opargsym(next, 2);
        int weightarg  = 2;
        if (weight->typespec().is_closure()) {  // opposite order
            std::swap(aa, weight);
            weightarg = 1;
        }
        if (name->typespec().is_string() && a->firstuse() >= opnum
            && a->lastuse() <= op2num && a == aa
            && weight->typespec().is_triple()) {
            if (is_zero(*weight)) {
                turn_into_nop(op, "zero-weighted closure");
                turn_into_assign(next, add_constant(0.0f),
                                 "zero-weighted closure");
                return 1;
            }
            // FIXME - handle weight being a float as well
            std::vector<int> newargs;
            newargs.push_back(oparg(next, 0));  // B
            if (!is_one(*weight))
                newargs.push_back(oparg(next, weightarg));  // weight
            for (int i = 1; i < op.nargs(); ++i)
                newargs.push_back(oparg(op, i));
            turn_into_nop(op, "combine closure+mul");
            turn_into_nop(next, "combine closure+mul");
            insert_code(opnum, u_closure, newargs, RecomputeRWRanges,
                        GroupWithNext);
            if (debug() > 1)
                std::cout << "op " << opnum << "-" << (op2num)
                          << " combined closure+mul\n";
            return 1;
        }
    }

    // No changes
    return 0;
}



/// Mark our params that feed to later layers, and whether we have any
/// outgoing connections.
void
RuntimeOptimizer::mark_outgoing_connections()
{
    OSL_ASSERT(!inst()->m_instoverrides.size()
               && "don't call this before copy_code_from_master");
    inst()->outgoing_connections(false);
    FOREACH_PARAM(auto&& s, inst())
    s.connected_down(false);
    for (int lay = layer() + 1; lay < group().nlayers(); ++lay) {
        for (auto&& c : group()[lay]->m_connections)
            if (c.srclayer == layer()) {
                inst()->symbol(c.src.param)->connected_down(true);
                inst()->outgoing_connections(true);
            }
    }
}



/// Check all params and output params to find any that are neither used
/// in the shader (aside from their own init ops, which shouldn't count)
/// nor connected to downstream layers, and for those, remove their init
/// ops and connections.
/// Precondition: mark_outgoing_connections should be up to date.
int
RuntimeOptimizer::remove_unused_params()
{
    int alterations = 0;
    SymNeverUsed param_never_used(*this, inst());  // handy predicate

    // Get rid of unused params' init ops and clear their read/write ranges
    FOREACH_PARAM(auto&& s, inst())
    {
        if (param_never_used(s) && s.has_init_ops()) {
            std::string why;
            if (debug() > 1)
                why = fmtformat("remove init ops of unused param {} {}",
                                s.typespec(), s.name());
            turn_into_nop(s.initbegin(), s.initend(), why);
            s.set_initrange(0, 0);
            s.clear_rw();  // mark as totally unused
            ++alterations;
        }
    }

    // Get rid of the Connections themselves
    if (debug() > 1) {
        for (auto&& c : inst()->connections()) {
            if (param_never_used(c)) {
                debug_optfmt("  Connection no longer needed: {} {}\n",
                             group()[c.srclayer]->layername(),
                             c.str(group(), inst()));
            }
        }
    }
    erase_if(inst()->connections(), param_never_used);

    return alterations;
}



void
RuntimeOptimizer::catalog_symbol_writes(int opbegin, int opend,
                                        FastIntSet& syms)
{
    for (int i = opbegin; i < opend; ++i) {
        const Opcode& op(inst()->ops()[i]);
        for (int a = 0, nargs = op.nargs(); a < nargs; ++a) {
            if (op.argwrite(a))
                syms.insert(oparg(op, a));
        }
    }
}



/// Find situations where an output is simply a copy of a connected
/// input, and eliminate the middleman.
int
RuntimeOptimizer::eliminate_middleman()
{
    int changed = 0;
    FOREACH_PARAM(auto&& s, inst())
    {
        // Skip if this isn't a shader output parameter that's connected
        // to a later layer.
        if (s.symtype() != SymTypeOutputParam || !s.connected_down())
            continue;
        // If it's written more than once, or has init ops, don't bother
        if (s.firstwrite() != s.lastwrite() || s.has_init_ops())
            continue;
        // Ok, s is a connected output, written only once, without init ops.

        // If the one time it's written isn't a simple assignment, never mind
        int opnum = s.firstwrite();
        Opcode& op(inst()->ops()[opnum]);
        if (op.opname() != u_assign)
            continue;  // only consider direct assignments
        // Now what's it assigned from?  If it's not a connected
        // parameter, or if it's not an equivalent data type, or if it's
        // a closure, never mind.
        int src_index = oparg(op, 1);
        Symbol* src   = opargsym(op, 1);

        if (!(src->symtype() == SymTypeParam && src->connected())
            || !equivalent(src->typespec(), s.typespec())
            || s.typespec().is_closure())
            continue;

        // Only works if the assignment is unconditional.  Needs to not
        // be in a conditional or loop, and not have any exit or return
        // statement before the assignment.
        if (!op_is_unconditionally_executed(opnum))
            continue;

        // OK, output param 's' is simply and unconditionally assigned
        // the value of the equivalently-typed input parameter 'src'.
        // Doctor downstream shaders that use s to connect directly to
        // src.

        // First, find what src is connected to.
        int upstream_layer = -1, upstream_symbol = -1;
        for (int i = 0, e = inst()->nconnections(); i < e; ++i) {
            const Connection& c = inst()->connection(i);
            if (c.dst.param == src_index &&  // the connection we want
                c.src.is_complete() && c.dst.is_complete()
                && equivalent(c.src.type, c.dst.type)
                && !c.src.type.is_closure() && !c.dst.type.is_closure()) {
                upstream_layer  = c.srclayer;
                upstream_symbol = c.src.param;
                break;
            }
        }
        if (upstream_layer < 0 || upstream_symbol < 0)
            continue;  // not a complete connection, forget it

        ShaderInstance* upinst = group()[upstream_layer];
        if (debug() > 1)
            std::cout << "Noticing that " << inst()->layername() << "."
                      << s.name() << " merely copied from " << src->name()
                      << ", connected from " << upinst->layername() << "."
                      << upinst->symbol(upstream_symbol)->name() << "\n";

        // Find all the downstream connections of s, make them
        // connections to src.
        int s_index = inst()->symbolindex(&s);
        for (int laynum = layer() + 1; laynum < group().nlayers(); ++laynum) {
            ShaderInstance* downinst = group()[laynum];
            for (int i = 0, e = downinst->nconnections(); i < e; ++i) {
                Connection& c = downinst->connections()[i];
                if (c.srclayer == layer() &&   // connected to our layer
                    c.src.param == s_index &&  // connected to s
                    c.src.is_complete() && c.dst.is_complete()
                    && equivalent(c.src.type, c.dst.type)) {
                    // just change the connection's referrant to the
                    // upstream source of s.
                    c.srclayer  = upstream_layer;
                    c.src.param = upstream_symbol;
                    ++changed;
                    shadingsys().m_stat_middlemen_eliminated += 1;
                    if (debug() > 1) {
                        const Symbol* dsym = downinst->symbol(c.dst.param);
                        if (!dsym)
                            dsym = downinst->mastersymbol(c.dst.param);
                        const Symbol* usym = upinst->symbol(upstream_symbol);
                        if (!usym)
                            usym = upinst->mastersymbol(upstream_symbol);
                        OSL_DASSERT(dsym && usym);
                        std::cout << "Removed " << inst()->layername() << "."
                                  << s.name() << " middleman for "
                                  << downinst->layername() << "."
                                  << dsym->name() << ", now connected to "
                                  << upinst->layername() << "." << usym->name()
                                  << "\n";
                    }
                }
            }
        }
    }
    return changed;
}



int
RuntimeOptimizer::optimize_assignment(Opcode& op, int opnum)
{
    // Various optimizations specific to assignment statements
    OSL_DASSERT(op.opname() == u_assign);
    int changed = 0;
    Symbol* R(inst()->argsymbol(op.firstarg() + 0));
    Symbol* A(inst()->argsymbol(op.firstarg() + 1));
    bool R_local_or_tmp = (R->symtype() == SymTypeLocal
                           || R->symtype() == SymTypeTemp);
    if (block_alias(inst()->arg(op.firstarg()))
            == inst()->arg(op.firstarg() + 1)
        || block_alias(inst()->arg(op.firstarg() + 1))
               == inst()->arg(op.firstarg())) {
        // We're re-assigning something already aliased, skip it
        turn_into_nop(op, "reassignment of current value (2)");
        return ++changed;
    }
    if (coerce_assigned_constant(op)) {
        // A may have changed, so we need to reset it
        A = inst()->argsymbol(op.firstarg() + 1);
        ++changed;
    }
    // NOW do assignment constant folding, only after we
    // have performed all the other transformations that may
    // turn this op into an assignment.
    changed += constfold_assign(*this, opnum);
    if (op.opname() != u_assign) {
        // The const fold has changed the assignment to something
        // other than assign (presumably nop), so skip the other
        // assignment transformations below.
        return 0;
    }
    if ((A->is_constant() || A->lastwrite() < opnum)
        && equivalent(R->typespec(), A->typespec())) {
        // Safe to alias R to A for this block, if A is a
        // constant or if it's never written to again.
        block_alias(inst()->arg(op.firstarg()), inst()->arg(op.firstarg() + 1));
        // std::cerr << opnum << " aliasing " << R->mangled() << " to "
        //       << inst()->argsymbol(op.firstarg()+1)->mangled() << "\n";
    }
    if (A->is_constant() && R->typespec() == A->typespec() && R_local_or_tmp
        && R->firstwrite() == opnum && R->lastwrite() == opnum) {
        // This local or temp is written only once in the
        // whole shader -- on this statement -- and it's
        // assigned a constant.  So just alias it to the
        // constant.
        int cind = inst()->args()[op.firstarg() + 1];
        global_alias(inst()->args()[op.firstarg()], cind);
        turn_into_nop(op, "replace symbol with constant");
        return ++changed;
    }
    if (R_local_or_tmp && !R->everread()) {
        // This local is written but NEVER READ.  nop it.
        turn_into_nop(op, "local/tmp never read");
        return ++changed;
    }
    if (outparam_assign_elision(opnum, op)) {
        return ++changed;
    }
    if (R == A) {
        // Just an assignment to itself -- turn into NOP!
        turn_into_nop(op, "self-assignment");
        return ++changed;
    } else if (R_local_or_tmp && R->lastread() < opnum && !m_in_loop[opnum]) {
        // Don't bother assigning if we never read it again
        turn_into_nop(op, "symbol never read again");
        return ++changed;
    }
    return changed;
}



void
RuntimeOptimizer::copy_block_aliases(const FastIntMap& old_block_aliases,
                                     FastIntMap& new_block_aliases,
                                     const FastIntSet* excluded,
                                     bool copy_temps)
{
    OSL_ASSERT(&old_block_aliases != &new_block_aliases
               && "copy_block_aliases does not work in-place");
    // Find all symbols written anywhere in the instruction range
    new_block_aliases.clear();
    new_block_aliases.reserve(old_block_aliases.size());
    for (auto&& oba : old_block_aliases) {
        if (oba.second < 0)
            continue;  // erased alias -- don't copy
        if (!copy_temps
            && (inst()->symbol(oba.first)->is_temp()
                || inst()->symbol(oba.second)->is_temp()))
            continue;  // don't copy temp aliases unless told to
        if (excluded
            && (excluded->find(oba.first) != excluded->end()
                || excluded->find(oba.second) != excluded->end()))
            continue;  // don't copy from excluded list
        new_block_aliases[oba.first] = oba.second;
    }
}



int
RuntimeOptimizer::optimize_ops(int beginop, int endop,
                               FastIntMap* seed_block_aliases)
{
    if (beginop >= endop)
        return 0;

    // Constant aliases valid for just this basic block
    clear_block_aliases();

    // Provide a place where, if we recurse, we can save prior block
    // aliases. Register them on the block_aliases_stack so that calls to
    // block_unalias() will unalias from there, too.
    FastIntMap saved_block_aliases;
    m_block_aliases_stack.push_back(&saved_block_aliases);

    int lastblock      = -1;
    int skipops        = 0;  // extra inserted ops to skip over
    int changed        = 0;
    size_t num_ops     = inst()->ops().size();
    size_t old_num_ops = num_ops;  // track when it changes
    for (int opnum = beginop; opnum < endop; opnum += 1) {
        OSL_DASSERT(old_num_ops == num_ops);  // better not happen unknowingly
        OSL_DASSERT(num_ops == inst()->ops().size());
        OSL_DASSERT(size_t(opnum) < inst()->ops().size());
        if (m_stop_optimizing)
            break;
        Opcode* op = &inst()->ops()[opnum];
        if (skipops) {
            // If a previous optimization inserted ops and told us
            // to skip over the new ones, we still need to unalias
            // any symbols written by this op, but otherwise skip
            // all subsequent optimizations until we run down the
            // skipops counter.
            block_unalias_written_args(*op);
            OSL_ASSERT(lastblock == m_bblockids[opnum]
                       && "this should not be a new basic block");
            --skipops;
            continue;  // Move along to the next op, no optimization here
        }
        // Things to do if we've just moved to a new basic block
        if (lastblock != m_bblockids[opnum]) {
            clear_block_aliases(seed_block_aliases);
            seed_block_aliases = NULL;  // only the first time
            clear_stale_syms();
            lastblock = m_bblockids[opnum];
        }
        // Things to do at the start of main code:
        // * Alias output params to their initial values, if known.
        if (opnum == inst()->m_maincodebegin) {
            for (int i = inst()->firstparam(); i < inst()->lastparam(); ++i) {
                Symbol* s(inst()->symbol(i));
                if (s->symtype() == SymTypeOutputParam && !s->interpolated()
                    && (s->valuesource() == Symbol::DefaultVal
                        || s->valuesource() == Symbol::InstanceVal)
                    && !s->has_init_ops() && !s->typespec().is_closure_based()
                    && !s->typespec().is_structure_based()) {
                    make_symbol_room(1);  // Make sure add_constant is ok
                    s        = inst()->symbol(i);
                    int cind = add_constant(s->typespec(), s->data());
                    block_alias(i, cind);  // Alias this symbol to the new const
                }
            }
        }
        // Nothing below here to do for no-ops, take early out.
        if (op->opname() == u_nop || op->opname() == u_functioncall_nr)
            continue;
        // De-alias the readable args to the op and figure out if
        // there are any constants involved.
        for (int i = 0, e = op->nargs(); i < e; ++i) {
            if (!op->argwrite(i)) {  // Don't de-alias args that are written
                int argindex    = op->firstarg() + i;
                int argsymindex = dealias_symbol(inst()->arg(argindex), opnum);
                inst()->args()[argindex] = argsymindex;
            }
            if (op->argread(i))
                use_stale_sym(oparg(*op, i));
        }

        const OpDescriptor* opd = shadingsys().op_descriptor(op->opname());
        // If it's a simple assignment and the lvalue is "stale", go
        // back and eliminate its last assignment.
        if (is_simple_assign(*op, opd))
            simple_sym_assign(oparg(*op, 0), opnum);
        // Make sure there's room for several more symbols, so that we
        // can add a few consts if we need to, without worrying about
        // the addresses of symbols changing when we add a new one below.
        make_symbol_room(max_new_consts_per_fold);
        // For various ops that we know how to effectively
        // constant-fold, dispatch to the appropriate routine.
        if (optimize() >= 2 && m_opt_constant_fold) {
            if (opd && opd->folder) {
                int c = (*opd->folder)(*this, opnum);
                if (c) {
                    changed += c;
                    // Re-check num_ops in case the folder inserted something
                    num_ops = inst()->ops().size();
                    skipops = num_ops - old_num_ops;
                    endop += num_ops - old_num_ops;  // adjust how far we loop
                    old_num_ops = num_ops;
                    op          = &inst()->ops()[opnum];  // in case ops resized
                }
            }
        }
        // Clear local block aliases for any args that were written
        // by this op
        block_unalias_written_args(*op);

        // Now we handle assignments.
        if (optimize() >= 2 && op->opname() == u_assign && m_opt_assign)
            changed += optimize_assignment(*op, opnum);
        if (optimize() >= 2 && m_opt_elide_useless_ops && opd
            && !(opd->flags & OpDescriptor::SideEffects))
            changed += useless_op_elision(*op, opnum);
        if (m_stop_optimizing)
            break;
        // Peephole optimization involving pair of instructions (the second
        // instruction will be in the same basic block.
        if (optimize() >= 2 && m_opt_peephole && op->opname() != u_nop
            && op->opname() != u_functioncall_nr) {
            // Find the next instruction in the same basic block
            int op2num = next_block_instruction(opnum);
            if (op2num) {
                int c = peephole2(opnum, op2num);
                if (c) {
                    changed += c;
                    // Re-check num_ops in case the folder inserted something
                    num_ops = inst()->ops().size();
                    // skipops = num_ops - old_num_ops;
                    endop += num_ops - old_num_ops;  // adjust how far we loop
                    old_num_ops = num_ops;
                    op          = &inst()->ops()[opnum];  // in case ops resized
                }
            }
        }

        // Special cases for "if", "functioncall", and loops: Optimize the
        // sequences of instructions in the bodies recursively in a way that
        // allows us to be clever about the basic block alias tracking.
        ustring opname = op->opname();
        if ((opname == u_if || opname == u_functioncall || opname == u_for
             || opname == u_while || opname == u_dowhile)
            && shadingsys().m_opt_seed_bblock_aliases) {
            // Find all symbols written anywhere in the instruction range
            // of the bodies.
            FastIntSet symwrites;
            catalog_symbol_writes(opnum + 1, op->farthest_jump(), symwrites);
            // Save the aliases from the basic block we are exiting.
            // If & function call: save all prior aliases.
            // Loops: dont save aliases involving syms written in the loop.
            // Note that for both cases, we don't copy aliases involving
            // temps, because that breaks our later assumptions (for temp
            // coalescing) that temp uses never cross basic block boundaries.
            if (opname == u_if || opname == u_functioncall)
                copy_block_aliases(m_block_aliases, saved_block_aliases);
            else
                copy_block_aliases(m_block_aliases, saved_block_aliases,
                                   &symwrites);
            // 'if' has 2 blocks (then, else), function call has just
            // one (the body), loops have 4 (init, cond, body, incr),
            int njumps = (opname == u_if) ? 2
                                          : (opname == u_functioncall ? 1 : 4);
            // Recursively optimize each body block.
            for (int j = 0; j < njumps; ++j) {
                changed += optimize_ops(j == 0 ? opnum + 1 : op->jump(j - 1),
                                        op->jump(j), &saved_block_aliases);
                op = &inst()->ops()[opnum];  // in case ops resized
            }
            // Adjust optimization loop end if any instructions were added
            num_ops = inst()->ops().size();
            endop += num_ops - old_num_ops;
            old_num_ops = num_ops;
            // Now we can restore the original aliases to seed the basic
            // block that follows. For if/function, we need to remove all
            // aliases referencing syms written within the conditional or
            // function body. For loops, recall that we already excluded
            // the written syms from the saved_block_aliases.
            if (opname == u_if || opname == u_functioncall) {
                FastIntMap restored_aliases;
                restored_aliases.swap(saved_block_aliases);
                // catalog again, in case optimizations in those blocks
                // caused writes that weren't apparent before.
                catalog_symbol_writes(opnum + 1, op->farthest_jump(),
                                      symwrites);
                copy_block_aliases(restored_aliases, saved_block_aliases,
                                   &symwrites);
            }
            seed_block_aliases = &saved_block_aliases;
            // Get ready to increment to the next instruction
            opnum = op->farthest_jump() - 1;
        }
    }
    m_block_aliases_stack.pop_back();  // Done with saved_block_aliases
    return changed;
}



void
RuntimeOptimizer::optimize_instance()
{
    // If "opt_layername" attribute is set, only optimize the named layer
    if (!shadingsys().m_opt_layername.empty()
        && shadingsys().m_opt_layername != inst()->layername())
        return;

    // Make a list of the indices of all constants.
    for (int i = 0, e = (int)inst()->symbols().size(); i < e; ++i)
        if (inst()->symbol(i)->symtype() == SymTypeConst)
            m_all_consts.push_back(i);

    // Turn all parameters with instance or default values, and which
    // cannot be overridden by geometry values, into constants or
    // aliases for globals.  Also turn connections from earlier layers'
    // outputs that are known to be constants or globals into constants
    // or global aliases without any connection.
    if (optimize() >= 2 && m_opt_simplify_param) {
        simplify_params();
    }

#ifndef NDEBUG
    // Confirm that the symbols between [firstparam,lastparam] are all
    // input or output params.
    FOREACH_PARAM(const Symbol& s, inst())
    {
        OSL_DASSERT(s.symtype() == SymTypeParam
                    || s.symtype() == SymTypeOutputParam);
    }
#endif

    // Recompute which of our params have downstream connections.
    mark_outgoing_connections();

    // Try to fold constants.  We take several passes, until we get to
    // the point that not much is improving.  It rarely goes beyond 3-4
    // passes, but we have a hard cutoff just to be sure we don't
    // ever get into an infinite loop from an unforeseen cycle where we
    // end up inadvertently transforming A => B => A => etc.
    int reallydone = 0;  // Force a few passes after we think we're done
    int npasses    = shadingsys().opt_passes();
    for (m_pass = 0; m_pass < npasses; ++m_pass) {
        // Once we've made one pass (and therefore called
        // mark_outgoing_connections), we may notice that the layer is
        // unused, and therefore can stop doing work to optimize it.
        if (m_pass != 0 && inst()->unused())
            break;

        if (m_stop_optimizing)
            break;

        if (debug() > 1)
            debug_optfmt("layer {} \"{}\", pass {}:\n", layer(),
                         inst()->layername(), m_pass);

        // Track basic blocks and conditional states
        find_conditionals();
        find_basic_blocks();

        // Clear local messages for this instance
        m_local_unknown_message_sent = false;
        m_local_messages_sent.clear();

        // Figure out which params are just aliases for globals (only
        // necessary to do once, on the first pass).
        if (m_pass == 0 && optimize() >= 2)
            find_params_holding_globals();

        // Here is the meat of the optimization, where we pass over the
        // code for this instance and make various transformations.
        int changed = optimize_ops(0, (int)inst()->ops().size());

        // Now that we've rewritten the code, we need to re-track the
        // variable lifetimes.
        track_variable_lifetimes();

        // Recompute which of our params have downstream connections.
        mark_outgoing_connections();

        // Find situations where an output is simply a copy of a connected
        // input, and eliminate the middleman.
        if (optimize() >= 2 && m_opt_middleman) {
            int c = eliminate_middleman();
            if (c)
                mark_outgoing_connections();
            changed += c;
        }

        // Elide unconnected parameters that are never read.
        if (optimize() >= 1)
            changed += remove_unused_params();

        // FIXME -- we should re-evaluate whether writes_globals() is still
        // true for this layer.

        // If nothing changed, we're done optimizing.  But wait, it may be
        // that after re-tracking variable lifetimes, we can notice new
        // optimizations!  So force another pass, then we're really done.
        if (changed < 1) {
            if (++reallydone > 3)
                break;
        } else {
            reallydone = 0;
        }
    }

    // A layer that was allowed to run lazily originally, if it no
    // longer (post-optimized) has any outgoing connections, is no
    // longer needed at all.
    if (inst()->unused()) {
        // Not needed.  Remove all its connections and ops.
        inst()->connections().clear();
        turn_into_nop(
            0, (int)inst()->ops().size() - 1,
            debug() > 1
                ? fmtformat("eliminate layer {} with no outward connections",
                            inst()->layername())
                      .c_str()
                : "");
        for (auto&& s : inst()->symbols())
            s.clear_rw();
    }

    // Now that we've optimized this layer, walk through the ops and
    // note which messages may have been sent, so subsequent layers will
    // know.
    for (auto& op : inst()->ops()) {
        if (op.opname() == u_setmessage) {
            Symbol& Name(*inst()->argsymbol(op.firstarg() + 0));
            if (Name.is_constant())
                m_messages_sent.push_back(Name.get_string());
            else
                m_unknown_message_sent = true;
        }
    }
}



void
RuntimeOptimizer::resolve_isconnected()
{
    for (auto& op : inst()->ops()) {
        if (op.opname() == u_isconnected) {
            inst()->make_symbol_room(1);
            SymbolPtr s = inst()->argsymbol(op.firstarg() + 1);
            while (const StructSpec* structspec = s->typespec().structspec()) {
                // How to deal with structures -- just change the reference
                // to the first field in the struct.
                // FIXME -- if we ever allow separate layer connection of
                // individual struct members, this will need something more
                // sophisticated.
                OSL_DASSERT(structspec && structspec->numfields() >= 1);
                std::string fieldname = (s->name().string() + "."
                                         + structspec->field(0).name.string());
                int fieldsymid        = inst()->findparam(ustring(fieldname));
                OSL_DASSERT(fieldsymid >= 0);
                s = inst()->symbol(fieldsymid);
            }
            bool upconnected = s->connected();
            if (s->interpolated() && shadingsys().userdata_isconnected())
                upconnected = true;
            bool downconnected = s->connected_down() || s->renderer_output();
            int val = (upconnected ? 1 : 0) + (downconnected ? 2 : 0);
            turn_into_assign(op, add_constant(TypeInt, &val),
                             "resolve isconnected()");
        }
    }
}



void
RuntimeOptimizer::track_variable_lifetimes(const SymbolPtrVec& allsymptrs)
{
    SymbolPtrVec oparg_ptrs;
    oparg_ptrs.reserve(inst()->args().size());
    for (auto&& a : inst()->args())
        oparg_ptrs.push_back(inst()->symbol(a));

    if (m_bblockids.size() != inst()->ops().size())
        find_basic_blocks();

    track_variable_lifetimes_main(inst()->ops(), oparg_ptrs, allsymptrs,
                                  &m_bblockids);
}



void
RuntimeOptimizer::track_variable_lifetimes()
{
    SymbolPtrVec allsymptrs;
    allsymptrs.reserve(inst()->symbols().size());
    for (auto&& s : inst()->symbols())
        allsymptrs.push_back(&s);

    track_variable_lifetimes(allsymptrs);
}


// This has O(n^2) memory usage, so only for debugging
//#define DEBUG_SYMBOL_DEPENDENCIES

// Add to the dependency map that "symbol A depends on symbol B".
void
RuntimeOptimizer::add_dependency(SymDependency& dmap, int A, int B)
{
    OSL_DASSERT(A < (int)inst()->symbols().size());
    OSL_DASSERT(B < (int)inst()->symbols().size());
    dmap[A].insert(B);

#ifdef DEBUG_SYMBOL_DEPENDENCIES
    // Unification -- make all of B's dependencies be dependencies of A.
    for (auto&& r : dmap[B])
        dmap[A].insert(r);
#endif
}


void
RuntimeOptimizer::syms_used_in_op(Opcode& op, std::vector<int>& rsyms,
                                  std::vector<int>& wsyms)
{
    rsyms.clear();
    wsyms.clear();
    for (int i = 0; i < op.nargs(); ++i) {
        int arg = inst()->arg(i + op.firstarg());
        if (op.argread(i))
            if (std::find(rsyms.begin(), rsyms.end(), arg) == rsyms.end())
                rsyms.push_back(arg);
        if (op.argwrite(i))
            if (std::find(wsyms.begin(), wsyms.end(), arg) == wsyms.end())
                wsyms.push_back(arg);
    }
}



// Fake symbol index for "derivatives" entry in dependency map.
static const int DerivSym = -1;


// Recursively mark symbols that have derivatives from dependency map
void
RuntimeOptimizer::mark_symbol_derivatives(SymDependency& symdeps,
                                          SymIntSet& visited, int d)
{
    for (auto&& r : symdeps[d]) {
        if (visited.find(r) == visited.end()) {
            visited.insert(r);

            Symbol* s = inst()->symbol(r);

            if (s->typespec().elementtype().is_float_based())
                s->has_derivs(true);

            mark_symbol_derivatives(symdeps, visited, r);
        }
    }
}


/// Run through all the ops, for each one marking its 'written'
/// arguments as dependent upon its 'read' arguments (and performing
/// unification as we go), yielding a dependency map that lets us look
/// up any symbol and see the set of other symbols on which it ever
/// depends on during execution of the shader.
void
RuntimeOptimizer::track_variable_dependencies()
{
    SymDependency symdeps;

    // It's important to note that this is simplistically conservative
    // in that it overestimates dependencies.  To see why this is the
    // case, consider the following code:
    //       // inputs a,b; outputs x,y; local variable t
    //       t = a;
    //       x = t;
    //       t = b;
    //       y = t;
    // We can see that x depends on a and y depends on b.  But the
    // dependency analysis we do below thinks that y also depends on a
    // (because t depended on both a and b, but at different times).
    //
    // This naivite will never miss a dependency, but it may
    // overestimate dependencies.  (Hence we call this "conservative"
    // rather than "wrong.")  We deem this acceptable for now, since
    // it's so much easier to implement the conservative dependency
    // analysis, and it's not yet clear that getting it closer to
    // optimal will have any performance impact on final shaders. Also
    // because this is probably no worse than the "dependency slop" that
    // would happen with loops and conditionals.  But we certainly may
    // revisit with a more sophisticated algorithm if this crops up
    // a legitimate issue.
    //
    // Because of this conservative approach, it is critical that this
    // analysis is done BEFORE temporaries are coalesced (which would
    // cause them to be reassigned in exactly the way that confuses this
    // analysis).

    symdeps.clear();

    std::vector<int> read, written;
    bool forcederivs = shadingsys().force_derivs();
    // Loop over all ops...
    for (auto&& op : inst()->ops()) {
        // Gather the list of syms read and written by the op.  Reuse the
        // vectors defined outside the loop to cut down on malloc/free.
        read.clear();
        written.clear();
        syms_used_in_op(op, read, written);

        // FIXME -- special cases here!  like if any ops implicitly read
        // or write to globals without them needing to be arguments.

        // For each symbol w written by the op...
        for (auto&& w : written) {
            // For each symbol r read by the op, make w depend on r.
            // (Unless r is a constant , in which case it's not necessary.)
            for (auto&& r : read)
                if (inst()->symbol(r)->symtype() != SymTypeConst)
                    add_dependency(symdeps, w, r);
            // If the op takes derivs, make the pseudo-symbol DerivSym
            // depend on those arguments.
            if (op.argtakesderivs_all() || forcederivs) {
                for (int a = 0; a < op.nargs(); ++a)
                    if (op.argtakesderivs(a) || forcederivs) {
                        Symbol& s(*opargsym(op, a));
                        // Constants can't take derivs
                        if (s.symtype() == SymTypeConst)
                            continue;
                        // Non-float types can't take derivs
                        if (s.typespec().is_closure()
                            || s.typespec().simpletype().basetype
                                   != TypeDesc::FLOAT)
                            continue;
                        // Careful -- not all globals can take derivs
                        if (s.symtype() == SymTypeGlobal
                            && !(s.mangled() == Strings::P
                                 || s.mangled() == Strings::I
                                 || s.mangled() == Strings::u
                                 || s.mangled() == Strings::v
                                 || s.mangled() == Strings::Ps))
                            continue;
                        add_dependency(symdeps, DerivSym,
                                       inst()->arg(a + op.firstarg()));
                    }
            }
        }
    }

    // Propagate derivative dependencies for any syms already known to
    // need derivs.  It's probably marked that way because another layer
    // downstream connects to it and needs derivatives of that
    // connection.
    int snum = 0;
    for (auto&& s : inst()->symbols()) {
        // Globals that get written should always provide derivs.
        // Exclude N, since its derivs are unreliable anyway, so no point
        // making it cause the whole disp shader to need derivs.
        if (s.symtype() == SymTypeGlobal && s.everwritten()
            && !s.typespec().is_closure_based() && s.mangled() != Strings::N)
            s.has_derivs(true);
        // If the renderer requests a symbol to be written as an output with
        // derivs, mark them to be generated.
        if (s.renderer_output()) {
            auto symloc = group().find_symloc(s.name(), inst()->layername(),
                                              SymArena::Outputs);
            if (symloc && symloc->derivs && s.everwritten()
                && !s.typespec().is_closure_based())
                s.has_derivs(true);
        }
        if (s.has_derivs())
            add_dependency(symdeps, DerivSym, snum);
        ++snum;
    }

    // Mark all symbols needing derivatives as such
    SymIntSet visited;
    mark_symbol_derivatives(symdeps, visited, DerivSym);

    // Only some globals are allowed to have derivatives
    for (auto&& s : inst()->symbols()) {
        if (s.symtype() == SymTypeGlobal
            && !(s.mangled() == Strings::P || s.mangled() == Strings::I
                 || s.mangled() == Strings::u || s.mangled() == Strings::v
                 || s.mangled() == Strings::Ps))
            s.has_derivs(false);
    }

#ifdef DEBUG_SYMBOL_DEPENDENCIES
    // Helpful for debugging

    std::cerr << "track_variable_dependencies\n";
    std::cerr << "\nDependencies:\n";
    for (auto&& m : symdeps) {
        if (m.first == DerivSym)
            std::cerr << "$derivs depends on ";
        else
            std::cerr << inst->symbol(m.first)->mangled() << " depends on ";
        for (auto&& d : m.second) {
            if (d == DerivSym)
                std::cerr << "$derivs ";
            else
                std::cerr << inst->symbol(d)->mangled() << ' ';
        }
        std::cerr << "\n";
    }
    std::cerr << "\n\n";

    // Invert the dependency
    SymDependency influences;
    for (auto&& m : symdeps)
        for (auto&& d : m.second)
            influences[d].insert(m.first);

    std::cerr << "\nReverse dependencies:\n";
    for (auto&& m : influences) {
        if (m.first == DerivSym)
            std::cerr << "$derivs contrbutes to ";
        else
            std::cerr << inst->symbol(m.first)->mangled() << " contributes to ";
        for (auto&& d : m.second) {
            if (d == DerivSym)
                std::cerr << "$derivs ";
            else
                std::cerr << inst->symbol(d)->mangled() << ' ';
        }
        std::cerr << "\n";
    }
    std::cerr << "\n\n";
#endif
}



// Is the symbol coalescable?
inline bool
coalescable(const Symbol& s)
{
    return (s.symtype() == SymTypeTemp &&    // only coalesce temporaries
            s.everused() &&                  // only if they're used
            s.dealias() == &s &&             // only if not already aliased
            !s.typespec().is_structure() &&  // only if not a struct
            s.fieldid() < 0);                //    or a struct field
}



/// Coalesce temporaries.  During code generation, we make a new
/// temporary EVERY time we need one.  Now we examine them all and merge
/// ones of identical type and non-overlapping lifetimes.
void
RuntimeOptimizer::coalesce_temporaries()
{
    // We keep looping until we can't coalesce any more.
    int ncoalesced = 1;
    while (ncoalesced) {
        ncoalesced = 0;  // assume we're done, unless we coalesce something

        // We use a greedy algorithm that loops over each symbol, and
        // then examines all higher-numbered symbols (in order) and
        // tries to merge the first one it can find that doesn't overlap
        // lifetimes.  The temps were created as we generated code, so
        // they are already sorted by their "first use".  Thus, for any
        // pair t1 and t2 that are merged, it is guaranteed that t2 is
        // the symbol whose first use the earliest of all symbols whose
        // lifetimes do not overlap t1.

        SymbolVec::iterator s;
        for (s = inst()->symbols().begin(); s != inst()->symbols().end(); ++s) {
            // Skip syms that can't be (or don't need to be) coalesced
            if (!coalescable(*s))
                continue;

            int sfirst = s->firstuse();
            int slast  = s->lastuse();

            // Loop through every other symbol
            for (SymbolVec::iterator t = s + 1; t != inst()->symbols().end();
                 ++t) {
                // Coalesce s and t if both syms are coalescable,
                // equivalent types, have nonoverlapping lifetimes,
                // and either both do or both do not need derivatives.
                if (coalescable(*t) && equivalent(s->typespec(), t->typespec())
                    && s->has_derivs() == t->has_derivs()
                    && (slast < t->firstuse() || sfirst > t->lastuse())
                    && (s->is_uniform() == t->is_uniform())
                    && (s->forced_llvm_bool() == t->forced_llvm_bool())) {
                    // Make all future t references alias to s
                    t->alias(&(*s));
                    // s gets union of the lifetimes
                    s->union_rw(t->firstread(), t->lastread(), t->firstwrite(),
                                t->lastwrite());
                    sfirst = s->firstuse();
                    slast  = s->lastuse();
                    // t gets marked as unused
                    t->clear_rw();
                    ++ncoalesced;
                }
            }
        }
        // std::cerr << "Coalesced " << ncoalesced << "\n";
    }

    // Since we may have aliased temps, now we need to make sure all
    // symbol refs are dealiased.
    for (auto&& arg : inst()->args()) {
        Symbol* s = inst()->symbol(arg);
        s         = s->dealias();
        arg       = s - inst()->symbol(0);
    }
}



void
RuntimeOptimizer::post_optimize_instance()
{
    inst()->evaluate_writes_globals_and_userdata_params();

    if (inst()->unused())
        return;  // skip the expensive stuff if we're not used anyway

    SymbolPtrVec allsymptrs;
    allsymptrs.reserve(inst()->symbols().size());
    for (auto&& s : inst()->symbols())
        allsymptrs.push_back(&s);

    m_bblockids.clear();  // Keep insert_code from getting confused
    m_in_conditional.clear();
    m_in_loop.clear();

    add_useparam(allsymptrs);

#if OSL_USE_BATCHED
    // We must identify which symbols are uniform before
    // trying to coalesce to avoid merging a varying with
    // a uniform symbol or forced_llvm_bool with an integer
    if (m_opt_batched_analysis) {
        BatchedAnalysis batched_analysis(shadingsys(), group());
        batched_analysis.analyze_layer(inst());
    }
#endif

    if (optimize() >= 1 && m_opt_coalesce_temps)
        coalesce_temporaries();
}



void
RuntimeOptimizer::collapse_syms()
{
    //
    // Make a new symbol table that removes all the unused symbols.
    //

    // Mark our params that feed to later layers, so that unused params
    // that aren't needed downstream can be removed.
    mark_outgoing_connections();

    SymbolVec new_symbols;                   // buffer for new symbol table
    std::vector<int> symbol_remap;           // mapping of old sym index to new
    int total_syms = 0;                      // number of new symbols we'll need
    SymNeverUsed never_used(*this, inst());  // handy predicate

    // First, just count how many we need and set up the mapping
    for (auto&& s : inst()->symbols()) {
        symbol_remap.push_back(total_syms);
        if (!never_used(s))
            ++total_syms;
    }

    // Now make a new table of the right (new) size, and copy the used syms
    new_symbols.reserve(total_syms);
    for (auto&& s : inst()->symbols()) {
        if (!never_used(s))
            new_symbols.push_back(s);
    }

    // Remap all the function arguments to the new indices
    for (auto&& arg : inst()->m_instargs)
        arg = symbol_remap[arg];

    // Fix our connections from upstream shaders
    for (auto&& c : inst()->m_connections)
        c.dst.param = symbol_remap[c.dst.param];

    // Fix downstream connections that reference us
    for (int lay = layer() + 1; lay < group().nlayers(); ++lay) {
        for (auto&& c : group()[lay]->m_connections)
            if (c.srclayer == layer())
                c.src.param = symbol_remap[c.src.param];
    }

    // Swap the new symbol list for the old.
    std::swap(inst()->m_instsymbols, new_symbols);
    {
        // adjust memory stats
        // Remember that they're already swapped
        off_t mem = vectorbytes(new_symbols)
                    - vectorbytes(inst()->m_instsymbols);
        ShadingSystemImpl& ss(shadingsys());
        spin_lock lock(ss.m_stat_mutex);
        ss.m_stat_mem_inst_syms -= mem;
        ss.m_stat_mem_inst -= mem;
        ss.m_stat_memory -= mem;
    }

    // Miscellaneous cleanup of other things that used symbol indices
    inst()->m_Psym       = -1;
    inst()->m_Nsym       = -1;
    inst()->m_firstparam = -1;
    inst()->m_lastparam  = -1;
    int i                = 0;
    for (auto&& s : inst()->symbols()) {
        if (s.symtype() == SymTypeParam || s.symtype() == SymTypeOutputParam) {
            if (inst()->m_firstparam < 0)
                inst()->m_firstparam = i;
            inst()->m_lastparam = i + 1;
        }
        if (s.name() == Strings::P)
            inst()->m_Psym = i;
        else if (s.name() == Strings::N)
            inst()->m_Nsym = i;
        ++i;
    }
#ifndef NDEBUG
    // Confirm that the symbols between [firstparam,lastparam] are all
    // input or output params.
    FOREACH_PARAM(const Symbol& s, inst())
    {
        OSL_DASSERT(s.symtype() == SymTypeParam
                    || s.symtype() == SymTypeOutputParam);
    }
#endif
}



void
RuntimeOptimizer::collapse_ops()
{
    //
    // Make new code that removes all the nops
    //
    OpcodeVec new_ops;          // buffer for new code
    std::vector<int> op_remap;  // mapping of old opcode indices to new
    int total_ops = 0;          // number of new ops we'll need

    // First, just count how many we need and set up the mapping
    for (auto&& op : inst()->ops()) {
        op_remap.push_back(total_ops);
        if (op.opname() != u_nop)
            ++total_ops;
    }

    // Now make a new table of the right (new) size, copy the used ops, and
    // reset the jump addresses.
    new_ops.reserve(total_ops);
    for (auto&& op : inst()->ops()) {
        if (op.opname() != u_nop) {
            new_ops.push_back(op);
            Opcode& newop(new_ops.back());
            for (int i = 0; i < (int)Opcode::max_jumps; ++i)
                if (newop.jump(i) >= 0)
                    newop.jump(i) = op_remap[newop.jump(i)];
        }
    }

    // Adjust 'main' code range and init op ranges
    inst()->m_maincodebegin = op_remap[inst()->m_maincodebegin];
    inst()->m_maincodeend   = (int)new_ops.size();
    FOREACH_PARAM(auto&& s, inst())
    {
        if (s.has_init_ops()) {
            s.initbegin(op_remap[s.initbegin()]);
            if (s.initend() < (int)op_remap.size())
                s.initend(op_remap[s.initend()]);
            else
                s.initend((int)new_ops.size());
        }
    }

    // Swap the new code for the old.
    std::swap(inst()->m_instops, new_ops);

    // These are no longer valid
    m_bblockids.clear();
    m_in_conditional.clear();
    m_in_loop.clear();
}



std::ostream&
RuntimeOptimizer::printinst(std::ostream& out) const
{
    out << "Shader " << inst()->shadername() << "\n";
    out << (inst()->unused() ? " UNUSED" : "");
    out << " connections in=" << inst()->nconnections();
    out << " out=" << inst()->outgoing_connections();
    out << (inst()->writes_globals() ? " writes_globals" : "");
    out << (inst()->userdata_params() ? " userdata_params" : "");
    out << (inst()->run_lazily() ? " run_lazily" : " run_unconditionally");
    out << (inst()->outgoing_connections() ? " outgoing_connections" : "");
    out << (inst()->renderer_outputs() ? " renderer_outputs" : "");
    out << (inst()->writes_globals() ? " writes_globals" : "");
    out << (inst()->entry_layer() ? " entry_layer" : "");
    out << (inst()->last_layer() ? " last_layer" : "");
    out << "\n";
    out << "  symbols:\n";
    for (size_t i = 0, e = inst()->symbols().size(); i < e; ++i)
        inst()->symbol(i)->print(out, 256);
#if 0
    out << "  int consts:\n    ";
    for (size_t i = 0;  i < inst()->m_iconsts.size();  ++i)
        out << inst()->m_iconsts[i] << ' ';
    out << "\n";
    out << "  float consts:\n    ";
    for (size_t i = 0;  i < inst()->m_fconsts.size();  ++i)
        out << inst()->m_fconsts[i] << ' ';
    out << "\n";
    out << "  string consts:\n    ";
    for (size_t i = 0;  i < inst()->m_sconsts.size();  ++i)
        out << "\"" << Strutil::escape_chars(inst()->m_sconsts[i]) << "\" ";
    out << "\n";
#endif
    out << "  code:\n";
    for (size_t i = 0, e = inst()->ops().size(); i < e; ++i) {
        const Opcode& op(inst()->ops()[i]);
        if (i == (size_t)inst()->maincodebegin())
            out << "(main)\n";
        out << "    " << i << ": " << op.opname();
        bool allconst = true;
        for (int a = 0; a < op.nargs(); ++a) {
            const Symbol* s(inst()->argsymbol(op.firstarg() + a));
            out << " " << s->name();
            if (s->symtype() == SymTypeConst) {
                out << " (";
                s->print_vals(out, 16);
                out << ")";
            }
            if (op.argread(a))
                allconst &= s->is_constant();
        }
        for (size_t j = 0; j < Opcode::max_jumps; ++j)
            if (op.jump(j) >= 0)
                out << " " << op.jump(j);
        out << "\t# ";
        //        out << "    rw " << fmtformat("{:x}", op.argread_bits())
        //            << ' ' << op.argwrite_bits();
        if (op.argtakesderivs_all())
            out << " %derivs(" << op.argtakesderivs_all() << ") ";
        if (allconst)
            out << "  CONST";
        if (i == 0 || bblockid(i) != bblockid(i - 1))
            out << "  BBLOCK-START";
        std::string filename = op.sourcefile().string();
        size_t slash         = filename.find_last_of("/");
        if (slash != std::string::npos)
            filename.erase(0, slash + 1);
        if (filename.length())
            out << "  (" << filename << ":" << op.sourceline() << ")";
        out << "\n";
    }
    if (inst()->nconnections()) {
        out << "  connections upstream:\n";
        for (int i = 0, e = inst()->nconnections(); i < e; ++i) {
            const Connection& c(inst()->connection(i));
            out << "    " << c.dst.type.c_str() << ' '
                << inst()->symbol(c.dst.param)->name();
            if (c.dst.arrayindex >= 0)
                out << '[' << c.dst.arrayindex << ']';
            out << " upconnected from layer " << c.srclayer << ' ';
            const ShaderInstance* up = group()[c.srclayer];
            out << "(" << up->layername() << ") ";
            out << "    " << c.src.type.c_str() << ' '
                << up->symbol(c.src.param)->name();
            if (c.src.arrayindex >= 0)
                out << '[' << c.src.arrayindex << ']';
            out << "\n";
        }
    }
    return out;
}



void
RuntimeOptimizer::run()
{
    Timer rop_timer;
    int nlayers = (int)group().nlayers();
    if (debug())
        shadingcontext()->infofmt(
            "About to optimize shader group {} ({} layers):", group().name(),
            nlayers);
    if (debug())
        std::cout << "About to optimize shader group " << group().name()
                  << "\n";

    for (int layer = 0; layer < nlayers; ++layer) {
        set_inst(layer);
        // These need to happen before merge_instances
        inst()->copy_code_from_master(group());
        mark_outgoing_connections();
    }

    // Inventory the network and print pre-optimized debug info
    size_t old_nsyms = 0, old_nops = 0;
    for (int layer = 0; layer < nlayers; ++layer) {
        set_inst(layer);
        if (debug() /* && optimize() >= 1*/) {
            find_basic_blocks();
            std::cout.flush();
            std::cout << "Before optimizing layer " << layer << " \""
                      << inst()->layername() << "\" (ID " << inst()->id()
                      << ") :\n";
            printinst(std::cout);
            std::cout << "\n--------------------------------\n" << std::endl;
        }
        old_nsyms += inst()->symbols().size();
        old_nops += inst()->ops().size();
    }

    // Clear messages sent for the group, they will be filled in by
    // optimize_instance().
    m_messages_sent.clear();
    m_unknown_message_sent = false;

    if (shadingsys().m_opt_merge_instances == 1)
        shadingsys().merge_instances(group());

    m_params_holding_globals.resize(nlayers);

    // Inventory for error calls so that if lazyerror=0 we don't incorrectly
    // assume the layer is unused.
    check_for_error_calls(false);

    // Optimize each layer, from first to last
    for (int layer = 0; layer < nlayers; ++layer) {
        set_inst(layer);
        if (inst()->unused())
            continue;
        // N.B. we need to resolve isconnected() calls before the instance
        // is otherwise optimized, or else isconnected() may not reflect
        // the original connectivity after substitutions are made.
        resolve_isconnected();
        optimize_instance();
    }
    check_for_error_calls(false);  // re-check

    // Optimize each layer again, from last to first (because some
    // optimizations are only apparent when the subsequent shaders have
    // been simplified).
    for (int layer = nlayers - 1; layer >= 0; --layer) {
        set_inst(layer);
        if (!inst()->unused())
            optimize_instance();
    }

    // Try merging instances again, now that we've optimized
    shadingsys().merge_instances(group(), true);

    for (int layer = nlayers - 1; layer >= 0; --layer) {
        set_inst(layer);
        if (inst()->unused())
            continue;
        find_basic_blocks();
        track_variable_dependencies();

        // For our parameters that require derivatives, mark their
        // upstream connections as also needing derivatives.
        for (auto&& c : inst()->m_connections) {
            if (inst()->symbol(c.dst.param)->has_derivs()) {
                Symbol* source = group()[c.srclayer]->symbol(c.src.param);
                if (source->typespec().elementtype().is_float_based())
                    source->has_derivs(true);
            }
        }
    }

    // Post-opt cleanup: add useparam, coalesce temporaries, etc.
    for (int layer = 0; layer < nlayers; ++layer) {
        set_inst(layer);
        post_optimize_instance();
    }

    // Last chance to eliminate duplicate instances
    shadingsys().merge_instances(group(), true);

    // Last inventory of error() calls, issue warnings if needed.
    check_for_error_calls(true);

    // Get rid of nop instructions and unused symbols.
    size_t new_nsyms = 0, new_nops = 0, new_deriv_syms = 0;
    for (int layer = 0; layer < nlayers; ++layer) {
        set_inst(layer);
        if (inst()->unused())
            continue;  // no need to print or gather stats for unused layers
        if (optimize() >= 1) {
            collapse_syms();
            collapse_ops();
        }
        if (debug() && !inst()->unused()) {
            track_variable_lifetimes();
            std::cout << "After optimizing layer " << layer << " \""
                      << inst()->layername() << "\" (ID " << inst()->id()
                      << ") :\n";
            printinst(std::cout);
            std::cout << "\n--------------------------------\n" << std::endl;
        }
        new_nsyms += inst()->symbols().size();
        new_nops += inst()->ops().size();
    }

    m_unknown_textures_needed   = false;
    m_unknown_closures_needed   = false;
    m_unknown_attributes_needed = false;
    m_textures_needed.clear();
    m_closures_needed.clear();
    m_globals_read  = 0;
    m_globals_write = 0;
    m_globals_needed.clear();
    m_userdata_needed.clear();
    m_attributes_needed.clear();
    bool does_nothing = true;
    std::vector<uint8_t> interactive_data;
    for (int layer = 0; layer < nlayers; ++layer) {
        set_inst(layer);
        if (inst()->unused())
            continue;  // no need to print or gather stats for unused layers
        FOREACH_SYM(Symbol & s, inst())
        {
            // set the layer numbers
            s.layer(layer);
            // Find interpolated parameters
            if ((s.symtype() == SymTypeParam
                 || s.symtype() == SymTypeOutputParam)
                && s.interpolated()) {
                UserDataNeeded udn(s.name(), layer, s.typespec().simpletype(),
                                   s.data(), s.has_derivs());
                std::set<UserDataNeeded>::iterator found;
                found = m_userdata_needed.find(udn);
                if (found == m_userdata_needed.end())
                    m_userdata_needed.insert(udn);
                else if (udn.derivs && !found->derivs) {
                    m_userdata_needed.erase(found);
                    m_userdata_needed.insert(udn);
                }
            }
            // Find interactive parameters
            if (s.symtype() == SymTypeParam && s.interactive()) {
                // Make enough room in interactive_data to accommodate this
                // interactive parameter, including correct alignment.
                size_t offset    = interactive_data.size();
                size_t typesize  = s.typespec().simpletype().size();
                size_t totalsize = typesize * (s.has_derivs() ? 3 : 1);
                size_t alignment = typesize > 4 ? 8 : 4;
                offset = OIIO::round_to_multiple_of_pow2(offset, alignment);
                interactive_data.resize(offset + totalsize);
                // Copy from the instance value to the interactive block
                // If the value is a string, copy its hash.
                if (s.typespec().is_string()) {
                    ustring string_data = *reinterpret_cast<ustring*>(s.data());
                    ustringhash string_hash(string_data);
                    memcpy(&interactive_data[offset], &string_hash, typesize);
                } else
                    memcpy(&interactive_data[offset], s.data(), typesize);
                if (totalsize > typesize)
                    memset(&interactive_data[offset] + typesize, 0,
                           totalsize - typesize);
                // Make sure the symbol remembers it's stored in the interactive
                // arena with the right offset.
                group().add_interactive_param(layer, s.name(), offset);
            }
            // Track which globals the group needs
            if (s.symtype() == SymTypeGlobal) {
                m_globals_needed.insert(s.name());
                int bit = int(ShadingSystem::globals_bit(s.name()));
                if (s.everread())
                    m_globals_read |= bit;
                if (s.everwritten())
                    m_globals_write |= bit;
            }
            if (s.has_derivs())
                ++new_deriv_syms;
        }
        for (auto&& op : inst()->ops()) {
            const OpDescriptor* opd = shadingsys().op_descriptor(op.opname());
            if (!opd)
                continue;
            // a non-unused layer with a nontrivial op does something
            if (op.opname() != Strings::end && op.opname() != Strings::useparam)
                does_nothing = false;
            // Useparam of a down-connected or renderer output does something
            if (op.opname() == Strings::useparam) {
                for (int i = 0, e = op.nargs(); i < e; ++i) {
                    Symbol* sym = opargsym(op, i);
                    if (sym->connected_down() || sym->renderer_output())
                        does_nothing = false;
                }
            }
            if (opd->flags & OpDescriptor::Tex) {
                // for all the texture ops, arg 1 is the texture name
                Symbol* sym = opargsym(op, 1);
                OSL_DASSERT(sym && sym->typespec().is_string());
                if (sym->is_constant()) {
                    m_textures_needed.insert(sym->get_string());
                } else {
                    m_unknown_textures_needed = true;
                }
            }
            if (op.opname() == u_closure) {
                // It's either 'closure result weight name' or 'closure result name'
                Symbol* sym = opargsym(op, 1);  // arg 1 is the closure name
                if (sym && !sym->typespec().is_string())
                    sym = opargsym(op, 2);
                OSL_DASSERT(sym && sym->typespec().is_string());
                if (sym->is_constant()) {
                    ustring closurename = sym->get_string();
                    m_closures_needed.insert(closurename);
                } else {
                    m_unknown_closures_needed = true;
                }
            } else if (op.opname() == u_backfacing) {
                m_globals_needed.insert(u_backfacing);
            } else if (op.opname() == u_calculatenormal) {
                m_globals_needed.insert(u_flipHandedness);
            } else if (op.opname() == u_getattribute) {
                Symbol* sym1 = opargsym(op, 1);
                OSL_DASSERT(sym1 && sym1->typespec().is_string());
                if (sym1->is_constant()) {
                    const auto insert = [&](const AttributeNeeded& attr) {
                        auto found = m_attributes_needed.find(attr);
                        if (found == m_attributes_needed.end())
                            m_attributes_needed.insert(attr);
                        else if (attr.derivs && !found->derivs) {
                            m_attributes_needed.erase(found);
                            m_attributes_needed.insert(attr);
                        }
                    };
                    if (op.nargs() == 3) {
                        Symbol* sym2 = opargsym(op, 2);
                        // getattribute( attributename, result )
                        insert(AttributeNeeded(sym1->get_string(), ustring(),
                                               sym2->typespec().simpletype(),
                                               sym2->has_derivs()));
                    } else if (op.nargs() == 4) {
                        Symbol* sym2 = opargsym(op, 2);
                        Symbol* sym3 = opargsym(op, 3);
                        if (sym2->typespec().is_string()) {
                            // getattribute( scopename, attributename, result )
                            if (sym2->is_constant()) {
                                insert(AttributeNeeded(
                                    sym2->get_string(), sym1->get_string(),
                                    sym3->typespec().simpletype(),
                                    sym3->has_derivs()));
                            } else {
                                m_unknown_attributes_needed = true;
                            }
                        } else {
                            // getattribute( attributename, arrayindex, result )
                            insert(
                                AttributeNeeded(sym1->get_string(), ustring(),
                                                sym3->typespec().simpletype(),
                                                sym3->has_derivs()));
                        }
                    } else if (op.nargs() == 5) {
                        Symbol* sym2 = opargsym(op, 2);
                        Symbol* sym4 = opargsym(op, 4);
                        if (sym2->typespec().is_string()) {
                            // getattribute( scopename, attributename, arrayindex, result )
                            if (sym2->is_constant()) {
                                insert(AttributeNeeded(
                                    sym2->get_string(), sym1->get_string(),
                                    sym4->typespec().simpletype(),
                                    sym4->has_derivs()));
                            } else {
                                m_unknown_attributes_needed = true;
                            }
                        }
                    } else {
                        OSL_DASSERT(false);
                    }
                } else {  // sym1 not constant
                    m_unknown_attributes_needed = true;
                }
            }
        }
    }
    group().does_nothing(does_nothing);
    group().setup_interactive_arena(interactive_data);

    m_stat_specialization_time = rop_timer();
    {
        // adjust memory stats
        ShadingSystemImpl& ss(shadingsys());
        spin_lock lock(ss.m_stat_mutex);
        ss.m_stat_preopt_syms += old_nsyms;
        ss.m_stat_preopt_ops += old_nops;
        ss.m_stat_postopt_syms += new_nsyms;
        ss.m_stat_postopt_ops += new_nops;
        ss.m_stat_syms_with_derivs += new_deriv_syms;
        if (does_nothing)
            ss.m_stat_empty_groups += 1;
    }
    if (shadingsys().m_compile_report) {
        shadingcontext()->infofmt("Optimized shader group {}:", group().name());
        shadingcontext()->infofmt(
            " spec {:1.2f}s, New syms {}/{} ({:5.1f}%), ops {}/{} ({:5.1f}%)",
            m_stat_specialization_time, new_nsyms, old_nsyms,
            100.0 * double((long long)new_nsyms - (long long)old_nsyms)
                / double(old_nsyms),
            new_nops, old_nops,
            100.0 * double((long long)new_nops - (long long)old_nops)
                / double(old_nops));
    }
    if (shadingsys().m_compile_report > 1) {
        if (does_nothing)
            shadingcontext()->infofmt("Group does nothing");
        if (m_textures_needed.size()) {
            shadingcontext()->infofmt("Group needs textures:");
            for (auto&& f : m_textures_needed)
                shadingcontext()->infofmt("    {}", f);
            if (m_unknown_textures_needed)
                shadingcontext()->infofmt(
                    "    Also may construct texture names on the fly.");
        }
        if (m_userdata_needed.size()) {
            shadingcontext()->infofmt("Group potentially needs userdata:");
            for (auto&& f : m_userdata_needed)
                shadingcontext()->infofmt("    {} {} {}", f.name, f.type,
                                          f.derivs ? "(derivs)" : "");
        }
        if (m_attributes_needed.size()) {
            shadingcontext()->infofmt("Group needs attributes:");
            for (auto&& f : m_attributes_needed)
                shadingcontext()->infofmt("    {} {} {}", f.name, f.scope,
                                          f.type);
            if (m_unknown_attributes_needed)
                shadingcontext()->infofmt(
                    "    Also may construct attribute names on the fly.");
        }
    }
}



bool
RuntimeOptimizer::police_(int type, const Opcode& op, string_view msg)
{
    if ((type & police_gpu_err) && shadingsys().m_gpu_opt_error) {
        shadingcontext()->errorfmt("Optimization error for GPUs:\n"
                                   "  group:  {}\n"
                                   "  layer:  {}\n"
                                   "  source: {}:{}\n"
                                   "  issue:  {}",
                                   group().name(), inst()->layername(),
                                   op.sourcefile(), op.sourceline(), msg);
        return true;
    } else if ((type & police_opt_warn) && shadingsys().m_opt_warnings) {
        shadingcontext()->warningfmt("Optimization warning:\n"
                                     "  group:  {}\n"
                                     "  layer:  {}\n"
                                     "  source: {}:{}\n"
                                     "  issue:  {}",
                                     group().name(), inst()->layername(),
                                     op.sourcefile(), op.sourceline(), msg);
    }
    return false;
}



bool
RuntimeOptimizer::check_for_error_calls(bool warn)
{
    // If the "lazyerror" option is set, there's nothing to do.
    if (shadingsys().m_lazyerror)
        return false;

    // Check all the layers (even ones we think are unused) for `error()`
    // calls that still remain after runtime optimization. If found, warn
    // and mark the layer as having error calls.
    bool err    = false;
    int nlayers = (int)group().nlayers();
    for (int layer = 0; layer < nlayers; ++layer) {
        set_inst(layer);
        inst()->has_error_op(false);
        for (auto&& op : inst()->ops()) {
            if (op.opname() == Strings::error) {
                inst()->has_error_op(true);
                if (warn)
                    err |= police(op,
                                  "error() call present in optimized shader.");
            }
        }
    }
    return err;
}



bool
RuntimeOptimizer::police_failed_optimizations()
{
    bool err        = false;
    int do_warn     = shadingsys().m_opt_warnings;
    bool do_gpu_err = shadingsys().m_gpu_opt_error;
    if (!do_warn && !do_gpu_err)
        return false;  // no need for any of this expense

    int nlayers = (int)group().nlayers();
    for (int layer = 0; layer < nlayers; ++layer) {
        set_inst(layer);
        if (inst()->unused())
            continue;  // no need to print or gather stats for unused layers
        for (auto&& op : inst()->ops()) {
            const OpDescriptor* opd = shadingsys().op_descriptor(op.opname());
            if (!opd)
                continue;
            if (!(opd->flags & OpDescriptor::Police))
                continue;
            if (opd->flags & OpDescriptor::Tex) {
                Symbol* sym = opargsym(op, 1);  // arg 1 is texture name
                OSL_DASSERT(sym && sym->typespec().is_string());
                if (!sym->is_constant()) {
                    err |= police(
                        police_gpu_err, op,
                        "{}(): texture name cannot be converted to a handle.",
                        op.opname());
                }
            }
            if (opd->flags & OpDescriptor::StrChars) {
                err |= police(
                    op,
                    "{}(): need for string characters couldn't be optimized away.",
                    op.opname());
            }
            if (opd->flags & OpDescriptor::StrCreate) {
                err |= police(
                    op,
                    "{}(): new string creation couldn't be optimized to a constant.",
                    op.opname());
            }
            // FIXME: Will add more tests and warnings as we go
        }
    }
    return err;
}

};  // namespace pvt
OSL_NAMESPACE_END