/// Are two symbols equivalent (from the point of view of merging
/// shader instances)? Note that this is not a true ==, it ignores
/// the m_data, m_node, and m_alias pointers!
static bool
equivalent (const Symbol &a, const Symbol &b)
{
// If they aren't used, don't consider them a mismatch
if (! a.everused() && ! b.everused())
return true;
// Different symbol types or data types are a mismatch
if (a.symtype() != b.symtype() || a.typespec() != b.typespec())
return false;
// Don't consider different names to be a mismatch if the symbol
// is a temp or constant.
if (a.symtype() != SymTypeTemp && a.symtype() != SymTypeConst &&
a.name() != b.name())
return false;
// But constants had better match their values!
if (a.symtype() == SymTypeConst &&
memcmp (a.data(), b.data(), a.typespec().simpletype().size()))
return false;
return a.has_derivs() == b.has_derivs() &&
a.lockgeom() == b.lockgeom() &&
a.valuesource() == b.valuesource() &&
a.fieldid() == b.fieldid() &&
a.initbegin() == b.initbegin() &&
a.initend() == b.initend()
;
}
inline std::string
print_vals (const Symbol &s)
{
std::stringstream out;
TypeDesc t = s.typespec().simpletype();
int n = t.aggregate * t.numelements();
if (t.basetype == TypeDesc::FLOAT) {
for (int j = 0; j < n; ++j)
out << (j ? " " : "") << ((float *)s.data())[j];
} else if (t.basetype == TypeDesc::INT) {
for (int j = 0; j < n; ++j)
out << (j ? " " : "") << ((int *)s.data())[j];
} else if (t.basetype == TypeDesc::STRING) {
for (int j = 0; j < n; ++j)
out << (j ? " " : "") << "\"" << ((ustring *)s.data())[j] << "\"";
}
return out.str();
}
virtual Value* eval(Environment& env) {
Symbol *symbol = dynamic_cast<Symbol*>(sym);
if (symbol) {
Value* be = bond->eval(env);
env.bind(symbol->data(), be);
return be->clone();
} else {
return NULL;
}
}
void
ShaderInstance::copy_code_from_master (ShaderGroup &group)
{
ASSERT (m_instops.empty() && m_instargs.empty());
// reserve with enough room for a few insertions
m_instops.reserve (master()->m_ops.size()+10);
m_instargs.reserve (master()->m_args.size()+10);
m_instops = master()->m_ops;
m_instargs = master()->m_args;
// Copy the symbols from the master
ASSERT (m_instsymbols.size() == 0 &&
"should not have copied m_instsymbols yet");
m_instsymbols = m_master->m_symbols;
// Copy the instance override data
// Also set the renderer_output flags where needed.
ASSERT (m_instoverrides.size() == (size_t)std::max(0,lastparam()));
ASSERT (m_instsymbols.size() >= (size_t)std::max(0,lastparam()));
if (m_instoverrides.size()) {
for (size_t i = 0, e = lastparam(); i < e; ++i) {
Symbol *si = &m_instsymbols[i];
if (m_instoverrides[i].valuesource() == Symbol::DefaultVal) {
// Fix the length of any default-value variable length array
// parameters.
if (si->typespec().is_unsized_array())
si->arraylen (si->initializers());
} else {
if (m_instoverrides[i].arraylen())
si->arraylen (m_instoverrides[i].arraylen());
si->valuesource (m_instoverrides[i].valuesource());
si->connected_down (m_instoverrides[i].connected_down());
si->lockgeom (m_instoverrides[i].lockgeom());
si->dataoffset (m_instoverrides[i].dataoffset());
si->data (param_storage(i));
}
if (shadingsys().is_renderer_output (layername(), si->name(), &group)) {
si->renderer_output (true);
renderer_outputs (true);
}
}
}
evaluate_writes_globals_and_userdata_params ();
off_t symmem = vectorbytes(m_instsymbols) - vectorbytes(m_instoverrides);
SymOverrideInfoVec().swap (m_instoverrides); // free it
// adjust stats
{
spin_lock lock (shadingsys().m_stat_mutex);
shadingsys().m_stat_mem_inst_syms += symmem;
shadingsys().m_stat_mem_inst += symmem;
shadingsys().m_stat_memory += symmem;
}
}
void
ShaderInstance::copy_code_from_master ()
{
ASSERT (m_instops.empty() && m_instargs.empty());
// reserve with enough room for a few insertions
m_instops.reserve (master()->m_ops.size()+10);
m_instargs.reserve (master()->m_args.size()+10);
m_instops = master()->m_ops;
m_instargs = master()->m_args;
// Copy the symbols from the master
ASSERT (m_instsymbols.size() == 0 &&
"should not have copied m_instsymbols yet");
m_instsymbols = m_master->m_symbols;
// Copy the instance override data
ASSERT (m_instoverrides.size() == (size_t)std::max(0,lastparam()));
ASSERT (m_instsymbols.size() >= (size_t)std::max(0,lastparam()));
if (m_instoverrides.size()) {
for (size_t i = 0, e = lastparam(); i < e; ++i) {
if (m_instoverrides[i].valuesource() != Symbol::DefaultVal) {
Symbol *si = &m_instsymbols[i];
si->data (param_storage(i));
si->valuesource (m_instoverrides[i].valuesource());
si->connected_down (m_instoverrides[i].connected_down());
si->lockgeom (m_instoverrides[i].lockgeom());
}
}
}
off_t symmem = vectorbytes(m_instsymbols) - vectorbytes(m_instoverrides);
SymOverrideInfoVec().swap (m_instoverrides); // free it
// adjust stats
{
spin_lock lock (shadingsys().m_stat_mutex);
shadingsys().m_stat_mem_inst_syms += symmem;
shadingsys().m_stat_mem_inst += symmem;
shadingsys().m_stat_memory += symmem;
}
}
void
BackendLLVM::llvm_assign_initial_value (const Symbol& sym)
{
// Don't write over connections! Connection values are written into
// our layer when the earlier layer is run, as part of its code. So
// we just don't need to initialize it here at all.
if (sym.valuesource() == Symbol::ConnectedVal &&
!sym.typespec().is_closure_based())
return;
if (sym.typespec().is_closure_based() && sym.symtype() == SymTypeGlobal)
return;
int arraylen = std::max (1, sym.typespec().arraylength());
// Closures need to get their storage before anything can be
// assigned to them. Unless they are params, in which case we took
// care of it in the group entry point.
if (sym.typespec().is_closure_based() &&
sym.symtype() != SymTypeParam && sym.symtype() != SymTypeOutputParam) {
llvm_assign_zero (sym);
return;
}
if ((sym.symtype() == SymTypeLocal || sym.symtype() == SymTypeTemp)
&& shadingsys().debug_uninit()) {
// Handle the "debug uninitialized values" case
bool isarray = sym.typespec().is_array();
int alen = isarray ? sym.typespec().arraylength() : 1;
llvm::Value *u = NULL;
if (sym.typespec().is_closure_based()) {
// skip closures
}
else if (sym.typespec().is_floatbased())
u = ll.constant (std::numeric_limits<float>::quiet_NaN());
else if (sym.typespec().is_int_based())
u = ll.constant (std::numeric_limits<int>::min());
else if (sym.typespec().is_string_based())
u = ll.constant (Strings::uninitialized_string);
if (u) {
for (int a = 0; a < alen; ++a) {
llvm::Value *aval = isarray ? ll.constant(a) : NULL;
for (int c = 0; c < (int)sym.typespec().aggregate(); ++c)
llvm_store_value (u, sym, 0, aval, c);
}
}
return;
}
if ((sym.symtype() == SymTypeLocal || sym.symtype() == SymTypeTemp) &&
sym.typespec().is_string_based()) {
// Strings are pointers. Can't take any chance on leaving
// local/tmp syms uninitialized.
llvm_assign_zero (sym);
return; // we're done, the parts below are just for params
}
ASSERT_MSG (sym.symtype() == SymTypeParam || sym.symtype() == SymTypeOutputParam,
"symtype was %d, data type was %s", (int)sym.symtype(), sym.typespec().c_str());
if (sym.has_init_ops() && sym.valuesource() == Symbol::DefaultVal) {
// Handle init ops.
build_llvm_code (sym.initbegin(), sym.initend());
} else if (! sym.lockgeom() && ! sym.typespec().is_closure()) {
// geometrically-varying param; memcpy its default value
TypeDesc t = sym.typespec().simpletype();
ll.op_memcpy (llvm_void_ptr (sym), ll.constant_ptr (sym.data()),
t.size(), t.basesize() /*align*/);
if (sym.has_derivs())
llvm_zero_derivs (sym);
} else {
// Use default value
int num_components = sym.typespec().simpletype().aggregate;
TypeSpec elemtype = sym.typespec().elementtype();
for (int a = 0, c = 0; a < arraylen; ++a) {
llvm::Value *arrind = sym.typespec().is_array() ? ll.constant(a) : NULL;
if (sym.typespec().is_closure_based())
continue;
for (int i = 0; i < num_components; ++i, ++c) {
// Fill in the constant val
llvm::Value* init_val = 0;
if (elemtype.is_floatbased())
init_val = ll.constant (((float*)sym.data())[c]);
else if (elemtype.is_string())
init_val = ll.constant (((ustring*)sym.data())[c]);
else if (elemtype.is_int())
init_val = ll.constant (((int*)sym.data())[c]);
ASSERT (init_val);
llvm_store_value (init_val, sym, 0, arrind, i);
}
}
if (sym.has_derivs())
llvm_zero_derivs (sym);
}
// Handle interpolated params.
// FIXME -- really, we shouldn't assign defaults or run init ops if
// the values are interpolated. The perf hit is probably small, since
// there are so few interpolated params, but we should come back and
// fix this later.
if ((sym.symtype() == SymTypeParam || sym.symtype() == SymTypeOutputParam)
&& ! sym.lockgeom()) {
std::vector<llvm::Value*> args;
args.push_back (sg_void_ptr());
args.push_back (ll.constant (sym.name()));
args.push_back (ll.constant (sym.typespec().simpletype()));
args.push_back (ll.constant ((int) sym.has_derivs()));
args.push_back (llvm_void_ptr (sym));
ll.call_function ("osl_bind_interpolated_param",
&args[0], args.size());
}
}
void
RuntimeOptimizer::llvm_assign_initial_value (const Symbol& sym)
{
// Don't write over connections! Connection values are written into
// our layer when the earlier layer is run, as part of its code. So
// we just don't need to initialize it here at all.
if (sym.valuesource() == Symbol::ConnectedVal &&
!sym.typespec().is_closure_based())
return;
if (sym.typespec().is_closure_based() && sym.symtype() == SymTypeGlobal)
return;
int arraylen = std::max (1, sym.typespec().arraylength());
// Closures need to get their storage before anything can be
// assigned to them. Unless they are params, in which case we took
// care of it in the group entry point.
if (sym.typespec().is_closure_based() &&
sym.symtype() != SymTypeParam && sym.symtype() != SymTypeOutputParam) {
llvm_assign_zero (sym);
}
if (sym.symtype() != SymTypeParam && sym.symtype() != SymTypeOutputParam &&
sym.symtype() != SymTypeConst && sym.typespec().is_string_based()) {
// Strings are pointers. Can't take any chance on leaving local/tmp
// syms uninitialized.
DASSERT (sym.symtype() == SymTypeLocal || sym.symtype() == SymTypeTemp);
llvm_assign_zero (sym);
return; // we're done, the parts below are just for params
}
if (sym.has_init_ops() && sym.valuesource() == Symbol::DefaultVal) {
// Handle init ops.
build_llvm_code (sym.initbegin(), sym.initend());
} else {
// Use default value
int num_components = sym.typespec().simpletype().aggregate;
for (int a = 0, c = 0; a < arraylen; ++a) {
llvm::Value *arrind = sym.typespec().is_array() ? llvm_constant(a) : NULL;
if (sym.typespec().is_closure_based())
continue;
for (int i = 0; i < num_components; ++i, ++c) {
// Fill in the constant val
llvm::Value* init_val = 0;
TypeSpec elemtype = sym.typespec().elementtype();
if (elemtype.is_floatbased())
init_val = llvm_constant (((float*)sym.data())[c]);
else if (elemtype.is_string())
init_val = llvm_constant (((ustring*)sym.data())[c]);
else if (elemtype.is_int())
init_val = llvm_constant (((int*)sym.data())[c]);
ASSERT (init_val);
llvm_store_value (init_val, sym, 0, arrind, i);
}
}
if (sym.has_derivs())
llvm_zero_derivs (sym);
}
// Handle interpolated params.
// FIXME -- really, we shouldn't assign defaults or run init ops if
// the values are interpolated. The perf hit is probably small, since
// there are so few interpolated params, but we should come back and
// fix this later.
if ((sym.symtype() == SymTypeParam || sym.symtype() == SymTypeOutputParam)
&& ! sym.lockgeom()) {
std::vector<llvm::Value*> args;
args.push_back (sg_void_ptr());
args.push_back (llvm_constant (sym.name()));
args.push_back (llvm_constant (sym.typespec().simpletype()));
args.push_back (llvm_constant ((int) sym.has_derivs()));
args.push_back (llvm_void_ptr (sym));
llvm_call_function ("osl_bind_interpolated_param",
&args[0], args.size());
}
}
