Remove direct access to ShaderGlobals members from llvm_ops

[lgritz]

This is a stepping stone to "shader globals placement".

Turns out that there were only two functions in llvm_ops.cpp that
directly needed to know the layout of the ShaderGlobals struct: the
implementations of osl_calculatenormal, and osl_raytype_bit.

For osl_calculatenormal, just pass the flipHandedness flag directly,
instead of the ShaderGlobals ptr, from which the field was extracted.

For osl_raytype_bit, I changed the function to only retrieve the bit
corresponding to a raytype name (and it's not in llvm_ops any more),
and the "bitwise and" code is just directly generated as bitcode.

I managed to complete the batched shading implementation of
calculatenormal! But for batched raytype_bit, I handled the case with
a constant name, but for the unknown name I got a little confused and
I'm hoping Alex can bail me out with some hand-holding or tell me
what needs to be done exactly.

Signed-off-by: Larry Gritz lg@larrygritz.com

[fpsunflower]

LGTM for the scalar path.

[AlexMWells]

As a matter of sanity checking, might consider pulling the global_symbol_ptr out and passing the optional is_uniform parameter and OSL_ASSERT(!is_uniform) as the function being called will interpret the pointer as varying.

[AlexMWells]

Wondering if less is more, meaning was op_load_int needed, in that LLVM_Util::ptr_to_cast or LLVM_Util::ptr_cast and LLVM_Util::op_load which could be used to provide the same functionality.

[AlexMWells]

You should be able to remove
OSL_BATCHOP int _OSL_OP(raytype_bit)(void* bsg, int bit)
from wide_shadingsys.cpp, and
DECL(__OSL_OP(raytype_bit), "iXi")
from builtindecl_wide_xmacro.h

[AlexMWells]

For the non-const case suggest following the same pattern you did for llvm_gen, go into wide_shading.cpp and rename the uniform and varying versions of raytype_name to raytype_bit.

For the uniform verson of raytype_name just return the bit by removing the bitwise AND return (bsg->uniform.raytype & bit), and in the batched_llvm_gen add a bitwise AND to the raytype from batchedShaderGlobals (very similar to llvm_gen version).

For the varying version, leave the foreach_unique loop, but again drop the bitwise AND int ray_is_named_type = ((bsg->uniform.raytype & bit) != 0) and just pass the bit variable into assign_all. But back in the batched_llvm_gen you will now need to allocate a varying temporary int to pass to varying version of raytype_bit. IE:

BatchedBackendLLVM::TempScope temp_scope(rop);

// We need wide place to hold the results of raytype_bit 
llvm::Value * loc_raytype_bit = rop.getOrAllocateTemp (TypeSpec(TypeDesc::INT), false /*derivs*/, false /*is_uniform*/, false /*forceBool*/, std::string("raytype_bit"));
            
llvm::Value *args[] = {
                sg,
                rop.llvm_void_ptr (loc_raytype_bit ),
                rop.llvm_void_ptr (Name),
                rop.ll.mask_as_int(rop.ll.current_mask())};
rop.ll.call_function (rop.build_name(func_spec), args);

llvm::Value* raytype= rop.ll.op_load(loc_raytype_bit );
llvm::Value* wide_bit = rop.ll.widen_value(bit);
llvm::Value* bitanded = rop.ll.op_and(raytype, wide_bit );
llvm::Value* ret = rop.ll.op_int_to_bool_to_int(bitanded);
OSL_DASSERT(!result_is_uniform);
rop.llvm_store_value (ret, Result);

To simplify, you can can probably combine some of the code between the uniform and non-uniform branches

[AlexMWells]

Does op_int_to_bool_to_int generate better code than opt_bool_to_int(op_ne(a, constant(0)))?
If so might be useful elsewhere.

[AlexMWells]

I don't see osl_raytype_bit_from_name referenced anywhere?
Although I do support using that name instead of just osl_raytype_bit.

[AlexMWells]

Why a FIXME, any technical barrier to passing context? For that matter, I've been pondering a ReadOnlyContext or DeviceContext that might be a stripped down version of the full context to answer just these types of questions.