summaryrefslogtreecommitdiff
path: root/source/slang/slang-emit-spirv.cpp
diff options
context:
space:
mode:
authorTim Foley <tfoleyNV@users.noreply.github.com>2019-11-14 13:11:07 -0800
committerGitHub <noreply@github.com>2019-11-14 13:11:07 -0800
commitce4829b03622c7c23096253b0ee80b0fc923321e (patch)
treee1232ee908b2e3fa604d1c68c08ca4df4f4f8652 /source/slang/slang-emit-spirv.cpp
parentd631233f4fcc2e41a9e7d7e0d3e277c90c81582b (diff)
Initial work on direct emission of SPIR-V (#1118)
* Initial work on direct emission of SPIR-V This change adds a first vertical slice of support for emitting SPIR-V code directly from the Slang IR, instead of generating it indirectly via GLSL. This work isn't usable for anything valuable right now; the goal is just to get something checked in that we can incrementally extend over time. When invoking `slangc`, the `-emit-spirv-directly` option can be used to turn on the new code path. I have not bothered to add an equivalent API option, because this flag is only intended to be used for testing in the immediate future. The existing `emitEntryPoint()` function has become `emitEntryPointSource()` to more accurately reflect its role in a world where we can also emit entry points to a binary format. Much of the logic that was inside `emitEntryPoint()` had to do with linking and then optimizing/transforming Slang IR code to get it ready for emission on a particular target. This logic has been factored into a new `linkAndOptimizeIR()` function that can be shared between the path that emits source and the new one that emits SPIR-V. The meat of the change is then the `emitSPIRVFromIR()` function in `slang-emit-spirv.cpp`, which is called *after* all the optimizations and transformations have been applied to the Slang IR to get it ready. Rather than repeat myself here, I will try to make the comments in `slang-emit-spirv.cpp` usable as documentation of the approach being taken. Smaller notes: * I've included a test case that compares `slangc` output directly to expected SPIR-V. This is perhaps not an ideal plan for how to test SPIR-V emission going forward, but it suffices for now. * The `external/` directory needed to be added to the include dirs for the `slang` project so that the new code can depend on the SPIR-V header. * In `slang-ir-link`, the direct SPIR-V generation path means that we now link with a target of SPIR-V instead of GLSL. In principle this can be used to ensure that appropriate variants of intrinsics are selected based on the knowledge that we are emitting SPIR-V. In practice, that isn't being used at all. * Fixup: path for SPIR-V headers While working on this PR I used a copy of `spirv.h` that I placed into the repository tree manually, but since I started the work we ended up with SPIR-V headers in our tree anyway, albeit at a different path. This change tries to fix things up so that my code uses the headers that were already placed in the repository. * fixup; 64-bit build issue * fixup: typo fixes based on review
Diffstat (limited to 'source/slang/slang-emit-spirv.cpp')
-rw-r--r--source/slang/slang-emit-spirv.cpp1141
1 files changed, 1141 insertions, 0 deletions
diff --git a/source/slang/slang-emit-spirv.cpp b/source/slang/slang-emit-spirv.cpp
new file mode 100644
index 000000000..c6f2f7468
--- /dev/null
+++ b/source/slang/slang-emit-spirv.cpp
@@ -0,0 +1,1141 @@
+// slang-emit-spirv.cpp
+#include "slang-emit.h"
+
+#include "slang-compiler.h"
+#include "slang-ir.h"
+#include "slang-ir-insts.h"
+
+#include "spirv/unified1/spirv.h"
+
+namespace Slang
+{
+
+// Our goal in this file is to convert a module in the Slang IR over to an
+// equivalent module in the SPIR-V intermediate language.
+//
+// The Slang IR is (intentionally) similar to SPIR-V in many ways, and both
+// can represent shaders at similar levels of abstraction, so much of the
+// translation involves one-to-one translation of Slang IR instructions
+// to their SPIR-V equivalents.
+//
+// SPIR-V differs from Slang IR in some key ways, and the SPIR-V
+// specification places many restrictions on how the IR can be encoded.
+// In some cases we will rely on earlier IR passes to convert Slang IR
+// into a form closer to what SPIR-V expects (e.g., by moving all
+// varying entry point parameters to global scope), but other differences
+// will be handled during the translation process.
+//
+// The logic in this file relies on the formal [SPIR-V Specification].
+// When we are making use of or enforcing some property from the spec,
+// we will try to refer to the relevant section in comments.
+//
+// [SPIR-V Specification]: https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html
+
+// [2.3: Physical Layout of a SPIR-V Module and Instruction]
+//
+// > A SPIR-V module is a single linear stream of words.
+//
+// [2.2: Terms]
+//
+// > Word: 32 bits.
+//
+// Despite the importance to SPIR-V, the `spirv.h` header doesn't
+// define a type for words, so we'll do it here.
+
+ /// A SPIR-V word.
+typedef uint32_t SpvWord;
+
+// [2.3: Physical Layout of a SPIR-V Module and Instruction]
+//
+// > All remaining words are a linear sequence of instructions.
+// > Each instruction is a stream of words
+//
+// After a fixed-size header, the contents of a SPIR-V module
+// is just a flat sequence of instructions, each of which is
+// just a sequence of words.
+//
+// In principle we could try to emit instructions directly
+// in one pass as a stream of words, but there are additional
+// constraints placed by the SPIR-V encoding that would make
+// a single-pass strategy very hard, so we don't attempt it.
+//
+// [2.4 Logical Layout of a Module]
+//
+// SPIR-V imposes some global ordering constraints on instructions,
+// such that certain instructions must come before or after others.
+// For example, all `OpCapability` instructions must come before any
+// `OpEntryPoint` instructions.
+//
+// While the SPIR-V spec doesn't use such a term, we will take
+// the enumeration of the ordering in Section 2.4 and use it to
+// define a list of *logical sections* that make up a SPIR-V module.
+
+ /// Logical sections of a SPIR-V module.
+enum class SpvLogicalSectionID
+{
+ Capabilities,
+ Extensions,
+ ExtIntInstImports,
+ MemoryModel,
+ EntryPoints,
+ ExecutionModes,
+ DebugStringsAndSource,
+ DebugNames,
+ Annotations,
+ Types,
+ Constants,
+ GlobalVariables,
+ FunctionDeclarations,
+ FunctionDefinitions,
+
+ Count,
+};
+
+// While the SPIR-V module is nominally (according to the spec) just
+// a flat sequence of instructions, in practice some of the instructions
+// are logically in a parent/child relationship.
+//
+// In particular, functions "own" the instructions between an `OpFunction`
+// and the matching `OpFunctionEnd`. We can also think of basic
+// blocks within a function as owning the instructions between
+// an `OpLabel` (which represents the bloc) and the next label
+// or the end of the function.
+//
+// Furthermore, the common case is SPIR-V is that an instruction
+// that defines some value must appear before any instruction
+// that uses that value as an operand. This property is often true
+// in a Slang IR module, but isn't strictly enforced for things at
+// the global scope.
+//
+// To deal with the above issues, our strategy will be to emit
+// SPIR-V instructions into a lightweight intermediate structure
+// that simplifies dealing with ordering constraiints on
+// instructions.
+//
+// We will start by forward-declaring the type we will
+// use to represent instructions:
+//
+struct SpvInst;
+
+// Next, we will define a base type that can serve as a parent
+// to SPIR-V instructions. Both the logical sections defined
+// earlier and instructions such as functions will be used
+// as parents.
+
+ /// Base type for SPIR-V instructions and logical sections of a module
+ ///
+ /// Holds and supports appending to a list of child instructions.
+struct SpvInstParent
+{
+public:
+ /// Add an instruction to the end of the list of children
+ void addInst(SpvInst* inst);
+
+ /// Dump all children, recursively, to a flattened list of SPIR-V words
+ void dumpTo(List<SpvWord>& ioWords);
+
+private:
+ /// The first child, if any.
+ SpvInst* m_firstChild = nullptr;
+
+ /// A pointer to the null pointer at the end of the linked list.
+ ///
+ /// If the list of children is empty this points to `m_firstChild`,
+ /// while if it is non-empty it points to the `nextSibling` field
+ /// of the last instruction.
+ ///
+ SpvInst** m_link = &m_firstChild;
+};
+
+// A SPIR-V instruction is then (in the general case) a potential
+// parent to other instructions.
+
+ /// A type to represent a SPIR-V instruction to be emitted.
+ ///
+ /// This type alows the instruction to be built up across
+ /// multiple steps in a mutable fashion.
+ ///
+struct SpvInst : SpvInstParent
+{
+ // [2.3: Physical Layout of a SPIR-V Module and Instruction]
+ //
+ // > Each instruction is a stream of words
+ //
+ // > Opcode: The 16 high-order bits are the WordCount of the instruction.
+ // > The 16 low-order bits are the opcode enumerant.
+ //
+ // We will store the "opcode enumerant" directly in our
+ // intermediate structure, and compute the word count on
+ // the fly when writing an instruction to an output buffer.
+
+ /// The SPIR-V opcode for the instruction
+ SpvOp opcode;
+
+ // [2.3: Physical Layout of a SPIR-V Module and Instruction]
+ //
+ // > Optional instruction type <id> (presence determined by opcode)
+ // > Optional instruction Result <id> (presence determined by opcode)
+ // > Operand 1 (if needed)
+ // > Operand 2 (if needed)
+ // > ...
+ //
+ // We represent the remaining words of the instruction (after
+ // the opcode word) as an undifferentiated array. Any code
+ // that encodes an instruction is responsible for knowing the
+ // opcode-specific data that is required.
+ //
+ // Our code does not need to process instruction operands after
+ // they have been written into a `SpvInst`. If we ever had
+ // cases where we needed to do post-processing, then we would
+ // need to store a more refined representation here.
+
+ /// The additional words of the instruction after the opcode
+ List<SpvWord> operandWords;
+
+ // We will store the instructions in a given `SpvInstParent`
+ // using an intrusive linked list.
+
+ /// The next instruction in the same `SpvInstParent`
+ SpvInst* nextSibling = nullptr;
+
+ /// The result <id> produced by this instruction, or zero if it has no result.
+ SpvWord id = 0;
+
+ /// Dump the instruction (and any children, recursively) into the flat array of SPIR-V words.
+ void dumpTo(List<SpvWord>& ioWords)
+ {
+ // [2.2: Terms]
+ //
+ // > Word Count: The complete number of words taken by an instruction,
+ // > including the word holding the word count and opcode, and any optional
+ // > operands. An instruction’s word count is the total space taken by the instruction.
+ //
+ SpvWord wordCount = 1 + SpvWord(operandWords.getCount());
+
+ // [2.3: Physical Layout of a SPIR-V Module and Instruction]
+ //
+ // > Opcode: The 16 high-order bits are the WordCount of the instruction.
+ // > The 16 low-order bits are the opcode enumerant.
+ //
+ ioWords.add(wordCount << 16 | opcode);
+
+ // The operand words simply follow the opcode word.
+ //
+ for( auto word : operandWords )
+ {
+ ioWords.add(word);
+ }
+
+ // In our representation choice, the children of a
+ // parent instruction will always follow the encoded
+ // words of a parent:
+ //
+ // * The instructions inside a function always follow the `OpFunction`
+ // * The instructions inside a block always follow the `OpLabel`
+ //
+ SpvInstParent::dumpTo(ioWords);
+ }
+};
+
+ /// A logical section of a SPIR-V module
+struct SpvLogicalSection : SpvInstParent
+{
+};
+
+// Now that we've filled in the definition of `SpvInst`, we can
+// go back and define the key operations on `SpvInstParent`.
+
+void SpvInstParent::addInst(SpvInst* inst)
+{
+ SLANG_ASSERT(inst);
+
+ // The user shouldn't be trying to add multiple instructions at once.
+ // If they really want that then they probably wanted to give `inst`
+ // some children.
+ //
+ SLANG_ASSERT(!inst->nextSibling);
+
+ *m_link = inst;
+ m_link = &inst->nextSibling;
+}
+
+void SpvInstParent::dumpTo(List<SpvWord>& ioWords)
+{
+ for( auto child = m_firstChild; child; child = child->nextSibling )
+ {
+ child->dumpTo(ioWords);
+ }
+}
+
+// Now that we've defined the intermediate data structures we will
+// use to represent SPIR-V code during emission, we will move on
+// to defining the main context type that will drive SPIR-V
+// code generation.
+
+ /// Context used for translating a Slang IR module to SPIR-V
+struct SPIRVEmitContext
+{
+ /// The Slang IR module being translated
+ IRModule* m_irModule;
+
+ // [2.2: Terms]
+ //
+ // > <id>: A numerical name; the name used to refer to an object, a type,
+ // > a function, a label, etc. An <id> always consumes one word.
+ // > The <id>s defined by a module obey SSA.
+ //
+ // [2.3: Physical Layout of a SPIR-V Module and Instruction]
+ //
+ // > Bound; where all <id>s in this module are guaranteed to satisfy
+ // > 0 < id < Bound
+ // > Bound should be small, smaller is better, with all <id> in a module being densely packed and near 0.
+ //
+ // Instructions will be referred to by their <id>s.
+ // We need to generate <id>s for instructions, and also
+ // compute the "bound" value that will be stored in
+ // the module header.
+ //
+ // We will use a single counter and allocate <id>s
+ // on demand. There may be some slop where we allocate
+ // an <id> for something that never gets referenced,
+ // but we expect the amount of slop to be small (and
+ // it can be cleaned up by other tools/passes).
+
+ /// The next destination `<id>` to allocate.
+ SpvWord m_nextID = 1;
+
+ // We will store the logical sections of the SPIR-V module
+ // in a single array so that we can easily look up a
+ // section by its `SpvLogicalSectionID`.
+
+ /// The logical sections of the SPIR-V module
+ SpvLogicalSection m_sections[int(SpvLogicalSectionID::Count)];
+
+ /// Get a logical section based on its `SpvLogicalSectionID`
+ SpvLogicalSection* getSection(SpvLogicalSectionID id)
+ {
+ return &m_sections[int(id)];
+ }
+
+ // At the end of emission we need a single linear stream of words,
+ // so we will eventually flatten `m_sections` into a single array.
+
+ /// The final array of SPIR-V words that defines the encoded module
+ List<SpvWord> m_words;
+
+ /// Emit the concrete words that make up the binary SPIR-V module.
+ ///
+ /// This function fills in `m_words` based on the data in `m_sections`.
+ /// This function should only be called once.
+ ///
+ void emitPhysicalLayout()
+ {
+ // [2.3: Physical Layout of a SPIR-V Module and Instruction]
+ //
+ // > Magic Number
+ //
+ m_words.add(SpvMagicNumber);
+
+ // > Version nuumber
+ //
+ m_words.add(SpvVersion);
+
+ // > Generator's magic number.
+ // > Its value does not affect any semantics, and is allowed to be 0.
+ //
+ // TODO: We should eventually register a non-zero
+ // magic number to represent Slang/slangc.
+ //
+ m_words.add(0);
+
+ // > Bound
+ //
+ // As described above, we use `m_nextID` to allocate
+ // <id>s, so its value when we are done emitting code
+ // can serve as the bound.
+ //
+ m_words.add(m_nextID);
+
+ // > 0 (Reserved for instruction schema, if needed.)
+ //
+ m_words.add(0);
+
+ // > First word of instruction stream
+ // > All remaining words are a linear sequence of instructions.
+ //
+ // Once we are done emitting the header, we emit all
+ // the instructions in our logical sections.
+ //
+ for( int ii = 0; ii < int(SpvLogicalSectionID::Count); ++ii )
+ {
+ m_sections[ii].dumpTo(m_words);
+ }
+ }
+
+ // We will often need to refer to an instrcition by its
+ // <id>, given only the Slang IR instruction that represents
+ // it (e.g., when it is used as an operand of another
+ // instruction).
+ //
+ // To that end we will keep a map of instructions that
+ // have been emitted, where a Slang IR instruction maps
+ // to the corresponding SPIR-V instruction.
+
+ /// Map a Slang IR instruction to the corresponding SPIR-V instruction
+ Dictionary<IRInst*, SpvInst*> m_mapIRInstToSpvInst;
+
+ /// Register that `irInst` maps to `spvInst`
+ void registerInst(IRInst* irInst, SpvInst* spvInst)
+ {
+ m_mapIRInstToSpvInst.Add(irInst, spvInst);
+ }
+
+ // When we are emitting an instruction that can produce
+ // a result, we will allocate an <id> to it so that other
+ // instructions can refer to it.
+ //
+ // We will allocate <id>s on emand as they are needed.
+
+ /// Get the <id> for `inst`, or assign one if it doesn't have one yet
+ SpvWord getID(SpvInst* inst)
+ {
+ auto id = inst->id;
+ if( !id )
+ {
+ id = m_nextID++;
+ inst->id = id;
+ }
+ return id;
+ }
+
+ // We will build up `SpvInst`s in a stateful fashion,
+ // mostly for convenience. We could in theory compute
+ // the number of words each instruction needs, then allocate
+ // the words, then fill them in, but that would make the
+ // emit logic more complicated and we'd like to keep it simple
+ // until we are sure performance is an issue.
+ //
+ // Emitting an instruction starts with picking the opcode
+ // and allocating the `SpvInst`.
+
+ /// Begin emitting an instruction with the given SPIR-V `opcode`.
+ ///
+ /// If `irInst` is non-null, then the resulting SPIR-V instruction
+ /// will be registered as corresponding to `irInst`.
+ ///
+ SpvInst* beginInst(SpvOp opcode, IRInst* irInst = nullptr)
+ {
+ // TODO: We are currently just leaking the `SpvInst`s we allocate.
+ // We should set up a pool allocator that this pass can use for
+ // both the `SpvInst`s and for their constituent words.
+ //
+ auto spvInst = new SpvInst();
+ spvInst->opcode = opcode;
+
+ if(irInst)
+ {
+ registerInst(irInst, spvInst);
+ }
+
+ return spvInst;
+ }
+
+ // Once an instruction has been created, we append the operand
+ // words to it with `emitOperand`. There are a few different
+ // case of operands that we handle.
+ //
+ // The simplest case is when an instruction takes an operand
+ // that is just a literal SPIR-V word.
+
+ /// Emit a literal `word` as an operand to `dst`.
+ void emitOperand(SpvInst* dst, SpvWord word)
+ {
+ dst->operandWords.add(word);
+ }
+
+ // The most common case of operand is an <id> that represents
+ // some other instruction. In cases where we already have
+ // an <id> we can emit it as a literal and the meaning is
+ // the same. If we have a `SpvInst` we can look up or
+ // generate an <id> for it.
+
+ /// Emit an operand to the `dst` instruction, which references `src` by its <id>
+ void emitOperand(SpvInst* dst, SpvInst* src)
+ {
+ emitOperand(dst, getID(src));
+ }
+
+ // Commonly, we will have an operand in the form of an `IRInst`
+ // which might either represent an instruction we've already
+ // emitted (e.g., because it came earlier in a function body)
+ // or which we have yet to emit (because it is a global-scope
+ // instruction that has not been referenced before).
+
+ /// Emit an operand to the `dst` instruction, which references `src` by its <id>
+ void emitOperand(SpvInst* dst, IRInst* src)
+ {
+ // We first ensure that the `src` instruction has been emitted,
+ // and then handle it as for any other <id> operand.
+ //
+ SpvInst* spvSrc = ensureInst(src);
+ emitOperand(dst, getID(spvSrc));
+ }
+
+ /// Ensure that an instruction has been emitted
+ SpvInst* ensureInst(IRInst* irInst)
+ {
+ SpvInst* spvInst = nullptr;
+ if( !m_mapIRInstToSpvInst.TryGetValue(irInst, spvInst) )
+ {
+ // If the `irInst` hasn't already been emitted,
+ // then we will assume that is is a global instruction
+ // (a constant, type, function, etc.) and we should make
+ // sure it gets emitted now.
+ //
+ // Note: this step means that emitting an instruction
+ // can be re-entrant/recursive. Because we emit the SPIR-V
+ // words for an instruction into an intermediate structure
+ // we don't have to worry about the re-entrancy causing
+ // the ordering of instruction words to be interleaved.
+ //
+ spvInst = emitGlobalInst(irInst);
+ }
+ return spvInst;
+ }
+
+ // Some instructions take a string as a literal operand,
+ // which requires us to follow the SPIR-V rules to
+ // encode the string into multiple operand words.
+
+ /// Emit an operand that is encoded as a literal string
+ void emitOperand(SpvInst* dst, UnownedStringSlice const& text)
+ {
+ // [Section 2.2.1 : Instructions]
+ //
+ // > Literal String: A nul-terminated stream of characters consuming
+ // > an integral number of words. The character set is Unicode in the
+ // > UTF-8 encoding scheme. The UTF-8 octets (8-bit bytes) are packed
+ // > four per word, following the little-endian convention (i.e., the
+ // > first octet is in the lowest-order 8 bits of the word).
+ //
+ // We start by emitting the contents of `text` in
+ // 4-byte chunks.
+ //
+ char const* cursor = text.begin();
+ char const* end = text.end();
+ while( (end - cursor) >= 4 )
+ {
+ SpvWord word;
+ memcpy(&word, cursor, 4);
+ emitOperand(dst, word);
+ cursor += 4;
+ }
+ //
+ // > The final word contains the string’s nul-termination character (0), and
+ // > all contents past the end of the string in the final word are padded with 0.
+ //
+ // For the last word, the low-order bytes will
+ // come from the remainder of the string (if
+ // there is anything left), and the rest will
+ // be left as zeros.
+ //
+ // TODO: This code should probably assert that `text`
+ // doesn't contain any embedded nul bytes, since they
+ // could lead to invalid encoded results.
+ //
+ SLANG_ASSERT((end - cursor) <= 3);
+ SpvWord lastWord = 0;
+ memcpy(&lastWord, cursor, (end - cursor));
+ emitOperand(dst, lastWord);
+ }
+
+ // Sometimes we will want to pass down an argument that
+ // represents a result <id> operand, but we won't yet
+ // have access to the `SpvInst` that will get the <id>.
+ // We will use a dummy `enum` type to support this case.
+
+ enum ResultIDToken { kResultID };
+
+ void emitOperand(SpvInst* dst, ResultIDToken)
+ {
+ // A result <id> operand uses the <id> of the instruction itself.
+ emitOperand(dst, getID(dst));
+ }
+
+ // As another convenience, there are often cases where
+ // we will want to emit all of the operands of some
+ // IR instruction as <id> operands of a SPIR-V
+ // instruction. This is handy in cases where the
+ // Slang IR and SPIR-V instructions agree on the
+ // number, order, and meaning of their operands.
+
+ /// Helper type for emitting all the operands of some IR instruction
+ struct OperandsOf
+ {
+ OperandsOf(IRInst* irInst)
+ : irInst(irInst)
+ {}
+
+ IRInst* irInst = nullptr;
+ };
+
+ /// Emit operand words for all the operands of a given IR instruction
+ void emitOperand(SpvInst* dst, OperandsOf const& other)
+ {
+ auto irInst = other.irInst;
+ auto operandCount = irInst->getOperandCount();
+ for( UInt ii = 0; ii < operandCount; ++ii )
+ {
+ emitOperand(dst, irInst->getOperand(ii));
+ }
+ }
+
+ // With the above routines, code can easily construct a SPIR-V
+ // instruction with arbitrary operands over multiple lines of code.
+ //
+ // In many cases, however, it is desirable to be able to emit
+ // an instruction more compactly, and for that we will introduce
+ // a number of `emitInst()` helpers that handle creating an
+ // instruction, filling in its operands, and adding it to a parent.
+ //
+ // These routines are overloaded on the number of operands, and
+ // also templates to work with any of the types for which
+ // `emitOperand()` works.
+ //
+ // In all of these cases, the caller takes responsibility for
+ // correctly matching the SPIR-V encoding rules for the chosen
+ // opcode, including whether a type <id> or result <id> is
+ // required.
+
+ SpvInst* emitInst(SpvInstParent* parent, IRInst* irInst, SpvOp opcode)
+ {
+ auto spvInst = beginInst(opcode, irInst);
+ parent->addInst(spvInst);
+ return spvInst;
+ }
+
+ template<typename A>
+ SpvInst* emitInst(SpvInstParent* parent, IRInst* irInst, SpvOp opcode, A const& a)
+ {
+ auto spvInst = beginInst(opcode, irInst);
+ emitOperand(spvInst, a);
+ parent->addInst(spvInst);
+ return spvInst;
+ }
+
+ template<typename A, typename B>
+ SpvInst* emitInst(SpvInstParent* parent, IRInst* irInst, SpvOp opcode, A const& a, B const& b)
+ {
+ auto spvInst = beginInst(opcode, irInst);
+ emitOperand(spvInst, a);
+ emitOperand(spvInst, b);
+ parent->addInst(spvInst);
+ return spvInst;
+ }
+
+ template<typename A, typename B, typename C>
+ SpvInst* emitInst(SpvInstParent* parent, IRInst* irInst, SpvOp opcode, A const& a, B const& b, C const& c)
+ {
+ auto spvInst = beginInst(opcode, irInst);
+ emitOperand(spvInst, a);
+ emitOperand(spvInst, b);
+ emitOperand(spvInst, c);
+ parent->addInst(spvInst);
+ return spvInst;
+ }
+
+ template<typename A, typename B, typename C, typename D>
+ SpvInst* emitInst(SpvInstParent* parent, IRInst* irInst, SpvOp opcode, A const& a, B const& b, C const& c, D const& d)
+ {
+ auto spvInst = beginInst(opcode, irInst);
+ emitOperand(spvInst, a);
+ emitOperand(spvInst, b);
+ emitOperand(spvInst, c);
+ emitOperand(spvInst, d);
+ parent->addInst(spvInst);
+ return spvInst;
+ }
+
+ template<typename A, typename B, typename C, typename D, typename E>
+ SpvInst* emitInst(SpvInstParent* parent, IRInst* irInst, SpvOp opcode, A const& a, B const& b, C const& c, D const& d, E const& e)
+ {
+ auto spvInst = beginInst(opcode, irInst);
+ emitOperand(spvInst, a);
+ emitOperand(spvInst, b);
+ emitOperand(spvInst, c);
+ emitOperand(spvInst, d);
+ emitOperand(spvInst, e);
+ parent->addInst(spvInst);
+ return spvInst;
+ }
+
+ // Now that we've gotten the core infrastructure out of the way,
+ // let's start looking at emitting some instructions that make
+ // up a SPIR-V module.
+ //
+ // We will start with certain instructions that are required
+ // to appear in a well-formed SPIR-V module for Vulkan, but
+ // which do not directly relate to any instruction in the
+ // Slang IR.
+
+ /// Emit the mandatory "front-matter" instructions that
+ /// the SPIR-V module must include to make it usable.
+ void emitFrontMatter()
+ {
+ // TODO: We should ideally add SPIR-V capabilities to
+ // the module as we emit instructions that require them.
+ // For now we will always emit the `Shader` capability,
+ // since every Vulkan shader module will use it.
+ //
+ emitInst(getSection(SpvLogicalSectionID::Capabilities), nullptr, SpvOpCapability, SpvCapabilityShader);
+
+ // [2.4: Logical Layout of a Module]
+ //
+ // > The single required OpMemoryModel instruction.
+ //
+ // A memory model is always required in SPIR-V module.
+ //
+ // The Vulkan spec further says:
+ //
+ // > The `Logical` addressing model must be selected
+ //
+ // It isn't clear if the GLSL450 memory model is also
+ // a requirement, but it is what glslang produces,
+ // so we will use it for now.
+ //
+ emitInst(getSection(SpvLogicalSectionID::MemoryModel), nullptr, SpvOpMemoryModel, SpvAddressingModelLogical, SpvMemoryModelGLSL450);
+ }
+
+ // Next, let's look at emitting some of the instructions
+ // that can occur at global scope.
+
+ /// Emit an instruction that is expected to appear at the global scope of the SPIR-V module.
+ ///
+ /// Returns the corresponding SPIR-V instruction.
+ ///
+ SpvInst* emitGlobalInst(IRInst* inst)
+ {
+ switch( inst->op )
+ {
+ // [3.32.6: Type-Declaration Instructions]
+ //
+
+#define CASE(IROP, SPVOP) \
+ case IROP: return emitInst(getSection(SpvLogicalSectionID::Types), inst, SPVOP, kResultID)
+
+ // > OpTypeVoid
+ CASE(kIROp_VoidType, SpvOpTypeVoid);
+
+ // > OpTypeBool
+ CASE(kIROp_BoolType, SpvOpTypeBool);
+
+#undef CASE
+
+ // > OpTypeInt
+
+#define CASE(IROP, BITS, SIGNED) \
+ case IROP: return emitInst(getSection(SpvLogicalSectionID::Types), inst, SpvOpTypeInt, kResultID, BITS, SIGNED)
+
+ CASE(kIROp_IntType, 32, 1);
+ CASE(kIROp_UIntType, 32, 0);
+ CASE(kIROp_Int64Type, 64, 1);
+ CASE(kIROp_UInt64Type, 64, 0);
+
+#undef CASE
+
+ // > OpTypeFloat
+
+#define CASE(IROP, BITS) \
+ case IROP: return emitInst(getSection(SpvLogicalSectionID::Types), inst, SpvOpTypeFloat, kResultID, BITS)
+
+ CASE(kIROp_HalfType, 16);
+ CASE(kIROp_FloatType, 32);
+ CASE(kIROp_DoubleType, 64);
+
+#undef CASE
+
+ // > OpTypeVector
+ // > OpTypeMatrix
+ // > OpTypeImage
+ // > OpTypeSampler
+ // > OpTypeArray
+ // > OpTypeRuntimeArray
+ // > OpTypeStruct
+ // > OpTypeOpaque
+ // > OpTypePointer
+
+ case kIROp_FuncType:
+ // > OpTypeFunction
+ //
+ // Both Slang and SPIR-V encode a function type
+ // with the result-type operand coming first,
+ // followed by operand sfor all the parameter types.
+ //
+ return emitInst(getSection(SpvLogicalSectionID::Types), inst, SpvOpTypeFunction, kResultID, OperandsOf(inst));
+
+ // > OpTypeForwardPointer
+
+ case kIROp_Func:
+ // [3.32.6: Function Instructions]
+ //
+ // > OpFunction
+ //
+ // Functions are complex enough that we'll handle
+ // them in a dedicated subroutine.
+ //
+ return emitFunc(as<IRFunc>(inst));
+
+ // ...
+
+ default:
+ SLANG_UNIMPLEMENTED_X("unhandled instruction opcode");
+ UNREACHABLE_RETURN(nullptr);
+ }
+ }
+
+ /// Emit the given `irFunc` to SPIR-V
+ SpvInst* emitFunc(IRFunc* irFunc)
+ {
+ // [2.4: Logical Layout of a Module]
+ //
+ // > All function declarations ("declarations" are functions
+ // > without a body; there is no forward declaration to a
+ // > function with a body).
+ // > ...
+ // > All function definitions (functions with a body).
+ //
+ // We need to treat functions differently based
+ // on whether they have a body or not, since these
+ // are encoded differently (and to different sections).
+ //
+ if( isDefinition(irFunc) )
+ {
+ return emitFuncDefinition(irFunc);
+ }
+ else
+ {
+ return emitFuncDeclaration(irFunc);
+ }
+ }
+
+ /// Emit a declaration for the given `irFunc`
+ SpvInst* emitFuncDeclaration(IRFunc* irFunc)
+ {
+ // For now we aren't handling function declarations;
+ // we expect to deal only with fully linked modules.
+ //
+ SLANG_UNUSED(irFunc);
+ SLANG_UNEXPECTED("function declaration in SPIR-V emit");
+ UNREACHABLE_RETURN(nullptr);
+ }
+
+ /// Emit a SPIR-V function definition for the Slang IR function `irFunc`.
+ SpvInst* emitFuncDefinition(IRFunc* irFunc)
+ {
+ // [2.4: Logical Layout of a Module]
+ //
+ // > All function definitions (functions with a body).
+ //
+ auto section = getSection(SpvLogicalSectionID::FunctionDefinitions);
+ //
+ // > A function definition is as follows.
+ // > * Function definition, using OpFunction.
+ // > * Function parameter declarations, using OpFunctionParameter.
+ // > * Block
+ // > * Block
+ // > * ...
+ // > * Function end, using OpFunctionEnd.
+ //
+
+ // [3.24. Function Control]
+ //
+ // TODO: We should eventually support emitting the "function control"
+ // mask to include inline and other hint bits based on decorations
+ // set on `irFunc`.
+ //
+ SpvFunctionControlMask spvFunctionControl = SpvFunctionControlMaskNone;
+
+ // [3.32.9. Function Instructions]
+ //
+ // > OpFunction
+ //
+ // Note that the type <id> of a SPIR-V function uses the
+ // *result* type of the function, while the actual function
+ // type is given as a later operand. Slan IR instead uses
+ // the type of a function instruction store, you know, its *type*.
+ //
+ SpvInst* spvFunc = emitInst(section, irFunc, SpvOpFunction,
+ irFunc->getDataType()->getResultType(),
+ kResultID,
+ spvFunctionControl,
+ irFunc->getDataType());
+
+ // > OpFunctionParameter
+ //
+ // Unlike Slang, where parameters always belong to blocks,
+ // the parameters of a SPIR-V function must appear as direct
+ // children of the function instruction, and before any basic blocks.
+ //
+ for( auto irParam : irFunc->getParams() )
+ {
+ emitInst(spvFunc, irParam, SpvOpFunctionParameter,
+ irParam->getFullType(),
+ kResultID);
+ }
+
+ // [3.32.17. Control-Flow Instructions]
+ //
+ // > OpLabel
+ //
+ // A Slang `IRBlock` corresponds to a SPIR-V `OpLabel`:
+ // each represents a basic block in the control flow
+ // graph of a parent function.
+ //
+ // We will allocate SPIR-V instructions to represent
+ // all of the blocks in a function before we emit
+ // body instructions into any of them. We do this
+ // because it is possible for one block to make
+ // forward reference to another (wheras that is
+ // not possible for ordinary instructions within
+ // the blocks in the Slang IR)
+ //
+ for( auto irBlock : irFunc->getBlocks() )
+ {
+ emitInst(spvFunc, irBlock, SpvOpLabel, kResultID);
+ }
+
+ // Once all the basic blocks have had instructions allocated
+ // for them, we go through and fill them in with their bodies.
+ //
+ for( auto irBlock : irFunc->getBlocks() )
+ {
+ // Note: because we already created the block above,
+ // we can be sure that it will have been registred.
+ //
+ SpvInst* spvBlock = nullptr;
+ m_mapIRInstToSpvInst.TryGetValue(irBlock, spvBlock);
+ SLANG_ASSERT(spvBlock);
+
+ // [3.32.17. Control-Flow Instructions]
+ //
+ // > OpPhi
+ //
+ // TODO: We eventually need to emit `OpPhi` instructions corresponding
+ // to the parameters of any non-entry block, with operands representing
+ // the values passed along incoming edges from the predecessor blocks.
+
+ for( auto irInst : irBlock->getOrdinaryInsts() )
+ {
+ // Any instructions local to the block will be emitted as children
+ // of the block.
+ //
+ emitLocalInst(spvBlock, irInst);
+ }
+ }
+
+ // [3.32.9. Function Instructions]
+ //
+ // > OpFunctionEnd
+ //
+ // In the SPIR-V encoding a function is logically the parent of any
+ // instructions up to a matching `OpFunctionEnd`. In our intermediate
+ // structure we will make the `OpFunctionEnd` be the last child of
+ // the `OpFunction`.
+ //
+ emitInst(spvFunc, nullptr, SpvOpFunctionEnd);
+
+ // We will emit any decorations pertinent to the function to the
+ // appropriate section of the module.
+ //
+ emitDecorations(irFunc, getID(spvFunc));
+
+ return spvFunc;
+ }
+
+ // The instructions that appear inside the basic blocks of
+ // functions are what we will call "local" instructions.
+ //
+ // When emititng blobal instructions, we usually have to
+ // pick the right logical section to emit them into, while
+ // for local instructions they will usually emit into
+ // a known parent (the basic block that contains them).
+
+ /// Emit an instruction that is local to the body of the given `parent`.
+ SpvInst* emitLocalInst(SpvInstParent* parent, IRInst* inst)
+ {
+ switch( inst->op )
+ {
+ default:
+ SLANG_UNIMPLEMENTED_X("unhandled instruction opcode");
+ break;
+
+ // [3.32.17. Control-Flow Instructions]
+ //
+ // > OpReturn
+ case kIROp_ReturnVoid: return emitInst(parent, inst, SpvOpReturn);
+ }
+ }
+
+ // Both "local" and "global" instructions can have decorations.
+ // When we decide to emit an instruction, we typically also want
+ // to emit any decoratons that were attached to it that have
+ // a SPIR-V equivalent.
+
+ /// Emit appropriate SPIR-V decorations for the given IR `irInst`.
+ ///
+ /// The given `dstID` should be the `<id>` of the SPIR-V instruction being decorated,
+ /// and should correspond to `irInst`.
+ ///
+ void emitDecorations(IRInst* irInst, SpvWord dstID)
+ {
+ for( auto decoration : irInst->getDecorations() )
+ {
+ emitDecoration(dstID, decoration);
+ }
+ }
+
+ /// Emit an appropriate SPIR-V decoration for the given IR `decoration`, if necessary and possible.
+ ///
+ /// The given `dstID` should be the `<id>` of the SPIR-V instruction being decorated,
+ /// and should correspond to the parent of `decoration` in the Slang IR.
+ ///
+ void emitDecoration(SpvWord dstID, IRDecoration* decoration)
+ {
+ // Unlike in the Slang IR, decorations in SPIR-V are not children
+ // of the instruction they decorate, and instead are free-standing
+ // instructions at global scope, which reference their target
+ // instruction by its `<id>`.
+ //
+ // The `IRDecoration` hierarchy in Slang also maps to several
+ // different categories of instruction in SPIR-V, only a subset
+ // of which are officialy called "decorations."
+ //
+ // We will continue to use the Slang terminology here, since
+ // this code path is a catch-all for stuff that only needs to
+ // be emitted if the owning instruction gets emitted.
+
+ switch( decoration->op )
+ {
+ default:
+ break;
+
+ // [3.32.2. Debug Instructions]
+ //
+ // > OpName
+ //
+ case kIROp_NameHintDecoration:
+ {
+ auto section = getSection(SpvLogicalSectionID::DebugNames);
+ auto nameHint = cast<IRNameHintDecoration>(decoration);
+ emitInst(section, decoration, SpvOpName, dstID, nameHint->getName());
+ }
+ break;
+
+ // [3.32.5. Mode-Setting Instructions]
+ //
+ // > OpEntryPoint
+ // > Declare an entry point, its execution model, and its interface.
+ //
+ case kIROp_EntryPointDecoration:
+ {
+ auto section = getSection(SpvLogicalSectionID::EntryPoints);
+
+ // TODO: The `OpEntryPoint` is required to list an varying
+ // input or output parameters (by `<id>`) used by the entry point,
+ // although these are encoded as global variables in the IR.
+ //
+ // Currently we have a pass that moves entry-point varying
+ // parameters to global scope for the benefit of GLSL output,
+ // but we do not maintain a connection between those parameters
+ // and the original entry point. That pass should be updated
+ // to attach a decoration linking the original entry point
+ // to the new globals, which would be used in the SPIR-V emit case.
+
+ auto entryPointDecor = cast<IREntryPointDecoration>(decoration);
+ auto spvStage = mapStageToExecutionModel(entryPointDecor->getProfile().GetStage());
+ auto name = entryPointDecor->getName()->getStringSlice();
+ emitInst(section, decoration, SpvOpEntryPoint, spvStage, dstID, name);
+ }
+ break;
+
+ // > OpExecutionMode
+
+ // [3.6. Execution Mode]: LocalSize
+ case kIROp_NumThreadsDecoration:
+ {
+ auto section = getSection(SpvLogicalSectionID::ExecutionModes);
+
+ // TODO: The `LocalSize` execution mode option requires
+ // literal values for the X,Y,Z thread-group sizes.
+ // There is a `LocalSizeId` variant that takes `<id>`s
+ // for those sizes, and we should consider using that
+ // and requiring the appropriate capabilities
+ // if any of the operands to the decoration are not
+ // literals (in a future where we support non-literals
+ // in those positions in the Slang IR).
+ //
+ auto numThreads = cast<IRNumThreadsDecoration>(decoration);
+ emitInst(section, decoration, SpvOpExecutionMode, dstID, SpvExecutionModeLocalSize,
+ SpvWord(numThreads->getX()->getValue()),
+ SpvWord(numThreads->getY()->getValue()),
+ SpvWord(numThreads->getZ()->getValue()));
+ }
+ break;
+
+ // ...
+ }
+ }
+
+ /// Map a Slang `Stage` to a corresponding SPIR-V execution model
+ SpvExecutionModel mapStageToExecutionModel(Stage stage)
+ {
+ switch( stage )
+ {
+ default:
+ SLANG_UNEXPECTED("unhandled stage");
+ UNREACHABLE_RETURN((SpvExecutionModel)0);
+
+#define CASE(STAGE, MODEL) \
+ case Stage::STAGE: return SpvExecutionModel##MODEL
+
+ CASE(Vertex, Vertex);
+ CASE(Hull, TessellationControl);
+ CASE(Domain, TessellationEvaluation);
+ CASE(Geometry, Geometry);
+ CASE(Fragment, Fragment);
+ CASE(Compute, GLCompute);
+
+ // TODO: Extended execution models for ray tracing, etc.
+
+#undef CASE
+ }
+ }
+};
+
+SlangResult emitSPIRVFromIR(
+ BackEndCompileRequest* compileRequest,
+ IRModule* irModule,
+ IRFunc* irEntryPoint,
+ List<uint8_t>& spirvOut)
+{
+ SLANG_UNUSED(compileRequest);
+ SLANG_UNUSED(irModule);
+ SLANG_UNUSED(irEntryPoint);
+
+ spirvOut.clear();
+
+ SPIRVEmitContext context;
+ context.m_irModule = irModule;
+ context.emitFrontMatter();
+ context.ensureInst(irEntryPoint);
+ context.emitPhysicalLayout();
+
+ spirvOut.addRange(
+ (uint8_t const*) context.m_words.getBuffer(),
+ context.m_words.getCount() * sizeof(context.m_words[0]));
+
+ return SLANG_OK;
+}
+
+
+} // namespace Slang