1*12c85518Srobert============ 2*12c85518SrobertHLSL Support 3*12c85518Srobert============ 4*12c85518Srobert 5*12c85518Srobert.. contents:: 6*12c85518Srobert :local: 7*12c85518Srobert 8*12c85518SrobertIntroduction 9*12c85518Srobert============ 10*12c85518Srobert 11*12c85518SrobertHLSL Support is under active development in the Clang codebase. This document 12*12c85518Srobertdescribes the high level goals of the project, the guiding principles, as well 13*12c85518Srobertas some idiosyncrasies of the HLSL language and how we intend to support them in 14*12c85518SrobertClang. 15*12c85518Srobert 16*12c85518SrobertProject Goals 17*12c85518Srobert============= 18*12c85518Srobert 19*12c85518SrobertThe long term goal of this project is to enable Clang to function as a 20*12c85518Srobertreplacement for the `DirectXShaderCompiler (DXC) 21*12c85518Srobert<https://github.com/microsoft/DirectXShaderCompiler/>`_ in all its supported 22*12c85518Srobertuse cases. Accomplishing that goal will require Clang to be able to process most 23*12c85518Srobertexisting HLSL programs with a high degree of source compatibility. 24*12c85518Srobert 25*12c85518SrobertNon-Goals 26*12c85518Srobert--------- 27*12c85518Srobert 28*12c85518SrobertHLSL ASTs do not need to be compatible between DXC and Clang. We do not expect 29*12c85518Srobertidentical code generation or that features will resemble DXC's implementation or 30*12c85518Srobertarchitecture. In fact, we explicitly expect to deviate from DXC's implementation 31*12c85518Srobertin key ways. 32*12c85518Srobert 33*12c85518SrobertGuiding Principles 34*12c85518Srobert================== 35*12c85518Srobert 36*12c85518SrobertThis document lacks details for architectural decisions that are not yet 37*12c85518Srobertfinalized. Our top priorities are quality, maintainability, and flexibility. In 38*12c85518Srobertaccordance with community standards we are expecting a high level of test 39*12c85518Srobertcoverage, and we will engineer our solutions with long term maintenance in mind. 40*12c85518SrobertWe are also working to limit modifications to the Clang C++ code paths and 41*12c85518Srobertshare as much functionality as possible. 42*12c85518Srobert 43*12c85518SrobertArchitectural Direction 44*12c85518Srobert======================= 45*12c85518Srobert 46*12c85518SrobertHLSL support in Clang is expressed as C++ minus unsupported C and C++ features. 47*12c85518SrobertThis is different from how other Clang languages are implemented. Most languages 48*12c85518Srobertin Clang are additive on top of C. 49*12c85518Srobert 50*12c85518SrobertHLSL is not a formally or fully specified language, and while our goals require 51*12c85518Sroberta high level of source compatibility, implementations can vary and we have some 52*12c85518Srobertflexibility to be more or less permissive in some cases. For modern HLSL DXC is 53*12c85518Srobertthe reference implementation. 54*12c85518Srobert 55*12c85518SrobertThe HLSL effort prioritizes following similar patterns for other languages, 56*12c85518Srobertdrivers, runtimes and targets. Specifically, We will maintain separation between 57*12c85518SrobertHSLS-specific code and the rest of Clang as much as possible following patterns 58*12c85518Srobertin use in Clang code today (i.e. ParseHLSL.cpp, SemaHLSL.cpp, CGHLSL*.cpp...). 59*12c85518SrobertWe will use inline checks on language options where the code is simple and 60*12c85518Srobertisolated, and prefer HLSL-specific implementation files for any code of 61*12c85518Srobertreasonable complexity. 62*12c85518Srobert 63*12c85518SrobertIn places where the HLSL language is in conflict with C and C++, we will seek to 64*12c85518Srobertmake minimally invasive changes guarded under the HLSL language options. We will 65*12c85518Srobertseek to make HLSL language support as minimal a maintenance burden as possible. 66*12c85518Srobert 67*12c85518SrobertDXC Driver 68*12c85518Srobert---------- 69*12c85518Srobert 70*12c85518SrobertA DXC driver mode will provide command-line compatibility with DXC, supporting 71*12c85518SrobertDXC's options and flags. The DXC driver is HLSL-specific and will create an 72*12c85518SrobertHLSLToolchain which will provide the basis to support targeting both DirectX and 73*12c85518SrobertVulkan. 74*12c85518Srobert 75*12c85518SrobertParser 76*12c85518Srobert------ 77*12c85518Srobert 78*12c85518SrobertFollowing the examples of other parser extensions HLSL will add a ParseHLSL.cpp 79*12c85518Srobertfile to contain the implementations of HLSL-specific extensions to the Clang 80*12c85518Srobertparser. The HLSL grammar shares most of its structure with C and C++, so we will 81*12c85518Srobertuse the existing C/C++ parsing code paths. 82*12c85518Srobert 83*12c85518SrobertSema 84*12c85518Srobert---- 85*12c85518Srobert 86*12c85518SrobertHLSL's Sema implementation will also provide an ``ExternalSemaSource``. In DXC, 87*12c85518Srobertan ``ExternalSemaSource`` is used to provide definitions for HLSL built-in data 88*12c85518Sroberttypes and built-in templates. Clang is already designed to allow an attached 89*12c85518Srobert``ExternalSemaSource`` to lazily complete data types, which is a **huge** 90*12c85518Srobertperformance win for HLSL. 91*12c85518Srobert 92*12c85518SrobertIf precompiled headers are used when compiling HLSL, the ``ExternalSemaSource`` 93*12c85518Srobertwill be a ``MultiplexExternalSemaSource`` which includes both the ``ASTReader`` 94*12c85518Srobertand ``HLSLExternalSemaSource``. For Built-in declarations that are already 95*12c85518Srobertcompleted in the serialized AST, the ``HLSLExternalSemaSource`` will reuse the 96*12c85518Srobertexisting declarations and not introduce new declarations. If the built-in types 97*12c85518Srobertare not completed in the serialized AST, the ``HLSLExternalSemaSource`` will 98*12c85518Srobertcreate new declarations and connect the de-serialized decls as the previous 99*12c85518Srobertdeclaration. 100*12c85518Srobert 101*12c85518SrobertCodeGen 102*12c85518Srobert------- 103*12c85518Srobert 104*12c85518SrobertLike OpenCL, HLSL relies on capturing a lot of information into IR metadata. 105*12c85518Srobert*hand wave* *hand wave* *hand wave* As a design principle here we want our IR to 106*12c85518Srobertbe idiomatic Clang IR as much as possible. We will use IR attributes wherever we 107*12c85518Srobertcan, and use metadata as sparingly as possible. One example of a difference from 108*12c85518SrobertDXC already implemented in Clang is the use of target triples to communicate 109*12c85518Srobertshader model versions and shader stages. 110*12c85518Srobert 111*12c85518SrobertOur HLSL CodeGen implementation should also have an eye toward generating IR 112*12c85518Srobertthat will map directly to targets other than DXIL. While IR itself is generally 113*12c85518Srobertnot re-targetable, we want to share the Clang CodeGen implementation for HLSL 114*12c85518Srobertwith other GPU graphics targets like SPIR-V and possibly other GPU and even CPU 115*12c85518Sroberttargets. 116*12c85518Srobert 117*12c85518SrobertHLSL Language 118*12c85518Srobert============= 119*12c85518Srobert 120*12c85518SrobertThe HLSL language is insufficiently documented, and not formally specified. 121*12c85518SrobertDocumentation is available on `Microsoft's website 122*12c85518Srobert<https://docs.microsoft.com/en-us/windows/win32/direct3dhlsl/dx-graphics-hlsl>`_. 123*12c85518SrobertThe language syntax is similar enough to C and C++ that carefully written C and 124*12c85518SrobertC++ code is valid HLSL. HLSL has some key differences from C & C++ which we will 125*12c85518Srobertneed to handle in Clang. 126*12c85518Srobert 127*12c85518SrobertHLSL is not a conforming or valid extension or superset of C or C++. The 128*12c85518Srobertlanguage has key incompatibilities with C and C++, both syntactically and 129*12c85518Srobertsemantically. 130*12c85518Srobert 131*12c85518SrobertAn Aside on GPU Languages 132*12c85518Srobert------------------------- 133*12c85518Srobert 134*12c85518SrobertDue to HLSL being a GPU targeted language HLSL is a Single Program Multiple Data 135*12c85518Srobert(SPMD) language relying on the implicit parallelism provided by GPU hardware. 136*12c85518SrobertSome language features in HLSL enable programmers to take advantage of the 137*12c85518Srobertparallel nature of GPUs in a hardware abstracted language. 138*12c85518Srobert 139*12c85518SrobertHLSL also prohibits some features of C and C++ which can have catastrophic 140*12c85518Srobertperformance or are not widely supportable on GPU hardware or drivers. As an 141*12c85518Srobertexample, register spilling is often excessively expensive on GPUs, so HLSL 142*12c85518Srobertrequires all functions to be inlined during code generation, and does not 143*12c85518Srobertsupport a runtime calling convention. 144*12c85518Srobert 145*12c85518SrobertPointers & References 146*12c85518Srobert--------------------- 147*12c85518Srobert 148*12c85518SrobertHLSL does not support referring to values by address. Semantically all variables 149*12c85518Srobertare value-types and behave as such. HLSL disallows the pointer dereference 150*12c85518Srobertoperators (unary ``*``, and ``->``), as well as the address of operator (unary 151*12c85518Srobert&). While HLSL disallows pointers and references in the syntax, HLSL does use 152*12c85518Srobertreference types in the AST, and we intend to use pointer decay in the AST in 153*12c85518Srobertthe Clang implementation. 154*12c85518Srobert 155*12c85518SrobertHLSL ``this`` Keyword 156*12c85518Srobert--------------------- 157*12c85518Srobert 158*12c85518SrobertHLSL does support member functions, and (in HLSL 2021) limited operator 159*12c85518Srobertoverloading. With member function support, HLSL also has a ``this`` keyword. The 160*12c85518Srobert``this`` keyword is an example of one of the places where HLSL relies on 161*12c85518Srobertreferences in the AST, because ``this`` is a reference. 162*12c85518Srobert 163*12c85518SrobertBitshifts 164*12c85518Srobert--------- 165*12c85518Srobert 166*12c85518SrobertIn deviation from C, HLSL bitshifts are defined to mask the shift count by the 167*12c85518Srobertsize of the type. In DXC, the semantics of LLVM IR were altered to accommodate 168*12c85518Srobertthis, in Clang we intend to generate the mask explicitly in the IR. In cases 169*12c85518Srobertwhere the shift value is constant, this will be constant folded appropriately, 170*12c85518Srobertin other cases we can clean it up in the DXIL target. 171*12c85518Srobert 172*12c85518SrobertNon-short Circuiting Logical Operators 173*12c85518Srobert-------------------------------------- 174*12c85518Srobert 175*12c85518SrobertIn HLSL 2018 and earlier, HLSL supported logical operators (and the ternary 176*12c85518Srobertoperator) on vector types. This behavior required that operators not short 177*12c85518Srobertcircuit. The non-short circuiting behavior applies to all data types until HLSL 178*12c85518Srobert2021. In HLSL 2021, logical and ternary operators do not support vector types 179*12c85518Srobertinstead builtin functions ``and``, ``or`` and ``select`` are available, and 180*12c85518Srobertoperators short circuit matching C behavior. 181*12c85518Srobert 182*12c85518SrobertPrecise Qualifier 183*12c85518Srobert----------------- 184*12c85518Srobert 185*12c85518SrobertHLSL has a ``precise`` qualifier that behaves unlike anything else in the C 186*12c85518Srobertlanguage. The support for this qualifier in DXC is buggy, so our bar for 187*12c85518Srobertcompatibility is low. 188*12c85518Srobert 189*12c85518SrobertThe ``precise`` qualifier applies in the inverse direction from normal 190*12c85518Srobertqualifiers. Rather than signifying that the declaration containing ``precise`` 191*12c85518Srobertqualifier be precise, it signifies that the operations contributing to the 192*12c85518Srobertdeclaration's value be ``precise``. Additionally, ``precise`` is a misnomer: 193*12c85518Srobertvalues attributed as ``precise`` comply with IEEE-754 floating point semantics, 194*12c85518Srobertand are prevented from optimizations which could decrease *or increase* 195*12c85518Srobertprecision. 196*12c85518Srobert 197*12c85518SrobertDifferences in Templates 198*12c85518Srobert------------------------ 199*12c85518Srobert 200*12c85518SrobertHLSL uses templates to define builtin types and methods, but disallowed 201*12c85518Srobertuser-defined templates until HLSL 2021. HLSL also allows omitting empty template 202*12c85518Srobertparameter lists when all template parameters are defaulted. This is an ambiguous 203*12c85518Srobertsyntax in C++, but Clang detects the case and issues a diagnostic. This makes 204*12c85518Srobertsupporting the case in Clang minimally invasive. 205*12c85518Srobert 206*12c85518SrobertVector Extensions 207*12c85518Srobert----------------- 208*12c85518Srobert 209*12c85518SrobertHLSL uses the OpenCL vector extensions, and also provides C++-style constructors 210*12c85518Srobertfor vectors that are not supported by Clang. 211*12c85518Srobert 212*12c85518SrobertStandard Library 213*12c85518Srobert---------------- 214*12c85518Srobert 215*12c85518SrobertHLSL does not support the C or C++ standard libraries. Like OpenCL, HLSL 216*12c85518Srobertdescribes its own library of built in types, complex data types, and functions. 217*12c85518Srobert 218*12c85518SrobertUnsupported C & C++ Features 219*12c85518Srobert---------------------------- 220*12c85518Srobert 221*12c85518SrobertHLSL does not support all features of C and C++. In implementing HLSL in Clang 222*12c85518Srobertuse of some C and C++ features will produce diagnostics under HLSL, and others 223*12c85518Srobertwill be supported as language extensions. In general, any C or C++ feature that 224*12c85518Srobertcan be supported by the DXIL and SPIR-V code generation targets could be treated 225*12c85518Srobertas a clang HLSL extension. Features that cannot be lowered to DXIL or SPIR-V, 226*12c85518Srobertmust be diagnosed as errors. 227*12c85518Srobert 228*12c85518SrobertHLSL does not support the following C features: 229*12c85518Srobert 230*12c85518Srobert* Pointers 231*12c85518Srobert* References 232*12c85518Srobert* ``goto`` or labels 233*12c85518Srobert* Variable Length Arrays 234*12c85518Srobert* ``_Complex`` and ``_Imaginary`` 235*12c85518Srobert* C Threads or Atomics (or Obj-C blocks) 236*12c85518Srobert* ``union`` types `(in progress for HLSL 202x) <https://github.com/microsoft/DirectXShaderCompiler/pull/4132>`_ 237*12c85518Srobert* Most features C11 and later 238*12c85518Srobert 239*12c85518SrobertHLSL does not support the following C++ features: 240*12c85518Srobert 241*12c85518Srobert* RTTI 242*12c85518Srobert* Exceptions 243*12c85518Srobert* Multiple inheritance 244*12c85518Srobert* Access specifiers 245*12c85518Srobert* Anonymous or inline namespaces 246*12c85518Srobert* ``new`` & ``delete`` operators in all of their forms (array, placement, etc) 247*12c85518Srobert* Constructors and destructors 248*12c85518Srobert* Any use of the ``virtual`` keyword 249*12c85518Srobert* Most features C++11 and later 250