| Age | Commit message (Collapse) | Author |
|---|
|
* Cleanup refactoring work around the IR builder
We have some long-term goals for the IR that require a more centralized and disciplined set of rules for how IR instructions get created/emitted. I had been working on trying to set things up so that all IR instruction creation goes through a single bottleneck point, but the non-trivial work in that branch was getting drowned out by the sheer volume of cleanup and refactoring changes. This change tries to pull together several of the more important cleanups.
The big pieces are:
* `IRBuilder` and `SharedIRBuilder` now protect their data members and rely on users to initialize them more directly via constructor of an `init()` method. This change affects a *bunch* of sites where `IRBuilder`s were created. I changed use sites to use the constructors whenever possible, and to use `init()` in cases where we had longer-lived builders that needed to be initialized multiple times.
* The insertion location for the `IRBuilder` now uses an encapsulated type called `IRInsertLoc`. This new type can replace what used to be just two `IRInst*` fields in the builder, and also covers some new functionality (if we ever want to take advantage of it). Very little client code cares about this change, but it is still a nice cleanup in terms of making things more explicit.
* The creation of an `IRModule` has been moded *out* of `IRBuilder`, because in practice we `IRBuilder` always wants to be associated with a pre-existing `IRModule` at creation time (via its `SharedIRBuilder`). There is now an `IRModule::create()` operation instead. This required changing the sequencing at many `IRModule` creation sites, since most had been contriving to make an `IRBuilder` first. There were also several cleanups because code had been carelessly using non-reference-counted pointers for `IRModule`s in ways that broke now that `IRModule::create()` always returns a `RefPtr`.
* The core operations to actually allocate memory for IR instructions were moved into `IRModule` (since they interact with the memory pool that the module owns). These *were* called `createEmptyInst()` but have been renamed into `_allocateInst()`. In principle these seem like they should only be needed to be called by the `IRBuilder`, but in practice they are also needed by the IR deserialization logic.
* A few core operations for emitting IR instructions that were associted with `IRBuilder` were moved to actually be methods on `IRBuilder`. First is `_findOrEmitConstant` which is the primary bottleneck for creating simple scalar constant values. Another is `_createInst` (formerly part of the templated `createInstImpl` along with `createInstWithSizeImpl`) which is the main bottleneck for allocation and initialization of any instruction other than a constant (well, the `IRModuleInst` is the other exception...). Finally, there is also `_maybeSetSourceLoc()`, which is obvious to scope inside the `IRBuilder` once it is protecting the source-location info.
Notes:
* The `minSizeInBytes` parameter to `_createInst()` might not actually be needed at all. At this point any `IRInst` subtypes that need data allocated for things other than their operands already get created manually via `_allocateInst` or `_findOrEmitConstant`, so I *think* we could remove that part. I will handle that in a subsequent cleanup if it turns out to be the case.
* There is one IR pass (`slang-ir-string-hash.cpp`) that is using manual `_allocateInst()` instead of going through an `IRBuilder`. It could be easily cleaned up to not do so (and I will probably make that change down the line), but for now I wanted to avoid doing anything that wasn't close to pure refactoring if I could.
* At this point in our design an `IRBuilder` is a very lightweight thing - it basically just owns the insertion location plus a source location to write into instructions. A lot of our code currently treats `IRBuilder`s like they are expensive and/or need to be re-used (which leads to them being used in more mutable/stateful ways). It is quite likely that as we clean up other aspects of the implementation of IR creation/emission we can make `IRBuilder` use feel more lightweight in ways that can streamline and simplify code.
* The next step for this work is to identify the different paths that eventually lead to `_createInst()` being called, and unify them at a single bottleneck operation that can own the decisions around when to create an instruction vs. when to re-use an existing one (rather than those decisions being baked into the various `IRBuilder` subroutines that create instructions of the various subtypes).
* fixup: gcc/clang C++ spec details
|
|
* #include an absolute path didn't work - because paths were taken to always be relative.
* WIP: First pass in supporting output of line error information.
* Add support for lexing to better be able to indicate SourceLocation information.
* Fix lexer usage in DiagnosticSink in C++ extractor.
* Update diagnostics tests to have line location info.
* Fixed test expected output that now have source location information in them.
* Better handling of tab.
* Fix test expected results for tabbing change.
* DiagnosticLexer -> DiagnosticSink::SourceLocationLexer
Added line continuation tests.
* Fix typo.
* Added String::appendRepeatedChar
* Change to rerun tests.
* Added source locations to IR dumping.
* Output column for IR dump source loc.
* Add support for closing brace location to AST.
Use closing brace location in lowering when adding return void.
* Set the source location through SourceLoc - simplifies identifying if current loc is valid.
* Copy terminator sloc.
* Test for improved #line handling.
* Made writer the last parameter for dumpIR.
Small improvements to comments.
* Disable sloc output on dump IR by default.
* Fix issue with #line and inlining.
* Fix for output with improved #line output.
* Small comment change - mainly to kick off TC build.
Co-authored-by: Tim Foley <tfoleyNV@users.noreply.github.com>
|
|
* Add an accessor for IRInst opcode
This main changing is renaming `IRInst::op` over to `IRInst::m_op` and then adds an accessor `IRInst::getOp()` to read it. The rest of the changes are just changing use sites to `getOp` (or to `m_op` in the limited cases where we write to it).
This work is in anticipation of a future change that might need to store an extra bit in the same field as the opcode. It seemed better to do this massive refactoring as a separate PR.
* fixup
|
|
This change includes a few different fixes for issues that arose in a user shader that made use of DXR 1.1.
The existing solution we had for handling the DXR 1.1 `RayQuery` type relied on the fact that a declaration like:
```hlsl
RayQuery<0> myRayQuery;
```
Looks like an undefined variable to existing Slang, while to dxc it is a variable declaration that runs an implicit default constructor (sneaking a bit of C++ into HLSL, but only in a way the standard library can use).
Slang was getting away with the fact that this maps to an undefined variable because it turns out that our emit logic would output the exact same declaration for an undefined value (since declaring a variable without initializing it is the simplest way to get an undefined value of a given type in a C-like language).
The main bug that arose here was that if the `RayQuery<...>`-typed variable was declared under control flow, then the `undefined` instructions introduced by our SSA pass would actually get inserted into the wrong block. Basically, when a block was trying to read a variable, and there was no preceding `store` to that variable in the block, we'd start looking for incoming values from its predecessor block(s). In the case where the variable *never* gets stored to, this search would eventually reach the first block of the function, where we'd realize the value must be `undefined`. The result was that we might insert an `undefined` instruction of some `T` into the first block of a function, but the type `T` might be the result of a lookup operation performed later in that function. This ends up creating a use of `T` that isn't dominated by the definition, which violates the SSA property. This violation of the SSA properties lead to us generating incorrect code in a later pass that deals with scoping differences between SSA form and our structured output statements; that code would end up creating a local variable to hold a *type* instead of a value.
The main fix is in `slang-ir-ssa.cpp`, where we catch the case of trying to read a variable in the block that declares it, if there we no preceding `store`s. We simply insert an `undefined` instruction before the first such read and write that out as the value of the variable to be used for subsequent instructions (up to the next `store`). This fixes the SSA dominance property for the `undefined` values that get introduced and thus technically fixes the output code for the user shader.
A secondary issue is that it is kind of gross to be relying on the behavior of `undefined` instructions in the IR for the semantics of an important standard library type like `RayQuery<...>`. A preceding change already added basic support for Slang to run default initializers (declared as `__init()`) on variables that are declared without an initial-value expression. This change adds such a default initializer to `RayQuery<...>` and maps it to a dedicated IR instruction that is intended to represent the idea of running a C++-style default constructor to produce a value. It turns out that the code we need to emit in that cse is identical to what we currently emit for `undefined` instructions, so that is helpful.
A tertiary issue is that when trying to run the user shader in debug mode, I ran into an assertion because our type layout logic for reflection had never dealt with the issue of user-defined `enum` types being used in constant buffers or other memory that needs layout. I added a quick fix that lays out any `enum` types as their "tag type" (which defaults to `int`).
Unfortunately, there is no easy way to check in a regression test for the user issue, because official `dxcompiler` versions with support for DXR 1.1 are not yet released (at least as of last time I checked).
|
|
* Synthesize "active mask" for CUDA
The Big Picture
===============
The most important change here is to `hlsl.meta.slang`, where the declaration of `WaveGetActiveMask()` is changed so that instead of mapping to `__activemask()` on CUDA (which is semantically incorrect) it maps to a dedicated IR instruction.
The other `WaveActive*()` intrinsics that make use of the implicit "active mask" concept had already been changed in #1336 so that they explicitly translate to call the equivalent `WaveMask*()` intrinsic with the result of `WaveGetActiveMask()`. As a result, all of the `WaveActive*()` functions are now no different from a user-defined function that uses `WaveGetActiveMask()`.
The bulk of the work in this change goes into an IR pass to replace the new instruction for getting the active mask gets replaced with appropriately computed values before we generate output CUDA code. That work is in `slang-ir-synthesize-active-mask.{h,cpp}`.
Utilities
=========
There are a few pieces of code that were helpful in writing the main pass but that can be explained separately:
* IR instructions were added corresponding to the Slang `WaveMaskBallot()` and `WaveMaskMatch()` functions, which map to the CUDA `__ballot_sync()` and `__match_any_sync()` operations, respectively. These are only implemented for the CUDA target because they are only being generated as part of our CUDA-only pass.
* The `IRDominatorTree` type was updated to make it a bit more robust in the presence of unreachable blocks in the CFG. It is possible that the same ends could be achieved more efficiently by folding the corner cases into the main logic, but I went ahead and made things very explicit for now.
* I added an `IREdge` utility type to better encapsulate the way that certain code operating on the predecessors/successors of an `IRBlock` were using an `IRUse*` to represent a control-flow edge. The `IREdge` type makes the logic of those operations more explicit. A future change should proably change it so that `IRBlock::getPredecessors()` and `getSuccessors()` are instead `getIncomingEdges()` and `getOutgoingEdges()` and work as iterators over `IREdge` values, given the way that the predecessor and successor lists today can contain duplicates.
* Using the above `IREdge` type, the logic for detecting and break critical edges was broken down into something that is a bit more clear (I hope), and that also factors out the breaking of an edge (by inserting a block along it) into a reusable subroutine.
The Main Pass
=============
The implementation of the new pass is in `slang-ir-synthesize-active-mask.cpp`, and that file attempts to include enough comments to make the logic clear. A brief summary for the benefit of the commit history:
* The first order of business is to identify functions that need to have the active mask value piped into them, and to add an additional parameter to them so that the active mask is passed down explicitly. Call sites are adjusted to pass down the active mask which can then result in new functions being identified as needing the active mask.
* The next challenge is for a function that uses the active mask, to compute the active mask value to use in each basic block. The entry block can easily use the active mask value that was passed in, while other blocks need more work.
* When doing a conditional branch, we can compute the new mask for the block we branch to as a function of the existing mask and the branch condition. E.g., the value `WaveMaskBallot(existingMask, condition)` can be used as the mask for the "then" block of an `if` statement.
* When control flow paths need to "reconverge" at a point after a structured control-flow statement, we need to insert logic to synchronize and re-build the mask that will execute after the statement, while also excluding any lanes/threads that exited the statement in other ways (e.g., an early `return` from the function).
The explanation here is fairly hand-wavy, but the actual pass uses much more crisp definitions, so the code itself should be inspected if you care about the details.
Tests
=====
The tests for the new feature are all under `tests/hlsl-intrinsic/active-mask/`. Most of them stress a single control-flow construct (`if`, `switch`, or loop) and write out the value of `WaveGetActiveMask()` at various points in the code.
In practice, our definition of the active mask doesn't always agree with what D3D/Vulkan implementations seem to produce in practice, and as a result a certain amount of effort has gone into adding tweaks to the tests that force them to produce the expected output on existing graphics APIs. These tweaks usually amount to introducing conditional branches that aren't actually conditional in practice (the branch condition is always `true` or always `false` at runtime), in order to trick some simplistic analysis approaches that downstream compilers seem to employ.
One test case currently fails on our CUDA target (`switch-trivial-fallthrough.slang`) and has been disabled. This is an expected failure, because making it produce the expected value requires a bit of detailed/careful coding that would add a lot of additional complexity to this change. It seemed better to leave that as future work.
Future Work
===========
* As discussed under "Tests" above, the handling of simple `switch` statements in the current pass is incomplete.
* There's an entire can of worms to be dealt with around the handling of fall-through for `switch`.
* The current work also doesn't handle `discard` statements, which is unimportant right now (CUDA doesn't have fragment shaders), but might matter if we decide to synthesize masks for other targets. Similar work would probably be needed if we ever have `throw` or other non-local control flow that crosses function boundaries.
* An important optimization opportunity is being left on the floor in this change. When block that comes "after" a structured control-flow region (which is encoded explicitly in Slang IR and SPIR-V) post-dominates the entry block of the region, then we know that the active mask when exiting the region must be the same as the mask when entering the region, and there is no need to insert explicit code to cause "re-convergence." This should be addressed in a follow-on change once we add code to Slang for computing a post-dominator tree from a function CFG.
* Related to the above, the decision-making around whether a basic block "needs" the active mask is perhaps too conservative, since it decides that any block that precedes one needing the active mask also needs it. This isn't true in cases where the active mask for a merge block can be inferred by post-dominance (as described above), so that the blocks that branch to it don't need to compute an active mask at all.
* If/when we extend the CPU target to support these operations (along with SIMD code generation, I assume), we will also need to synthesize an active mask on that platform, but the approach taken here (which pretty much relies on support for CUDA "cooperative groups") wouldn't seem to apply in the SIMD case.
* Similarly, the approach taken to computing the active mask here requires a new enough CUDA SM architecture version to support explicit cooperative groups. If we want to run on older CUDA-supporting architectures, we will need a new and potentially very different strategy.
* Because the new pass here changes the signature of functions that require the active mask (and not those that don't), it creates possible problems for generating code that uses dynamic dispatch (via function pointers). In principle, we need to know at a call site whether or not the callee uses the active mask. There are multiple possible solutions to this problem, and they'd need to be worked through before we can make the implicit active mask and dynamic dispatch be mutually compatible.
* Related to changing function signatures: no effort is made in this pass to clean up the IR type of the functions it modifies, so there could technically be mismatches between the IR type of a function and its actual signature. If/when this causes problems for downstream passes we probably need to do some cleanup.
* fixup: backslash-escaped lines
I did some "ASCII art" sorts of diagrams to explain cases in the CFG, and some of those diagrams used backslash (`\`) characters as the last character on the line, causing them to count as escaped newlines for C/C++.
The gcc compiler apparently balked at those lines, since they made some of the single-line comments into multi-line comments.
I solved the problem by adding a terminating column of `|` characters at the end of each line that was part of an ASCII art diagram.
* fixup: typos
Co-authored-by: jsmall-nvidia <jsmall@nvidia.com>
|
|
It was possible for a recursive call to `tryRemoveTriviailPhi` to remove/replace a phi node that was already in a list of removal candidates to be processed. When the recursive call returned and that candidate was again considered, its operands would have already been cleared, leading to an assertion failure.
This case is what was coming up in practice in a user shader, although I have not been able to reproduce the failure with a more minimal synthetic test yet.
This change also changes the SSA creation logic to avoid a runtime crash in the case of a trivial phi that only references itself (which was how the above bug surfaced to the user). The "fix" there is not ideal (it leaves a trivial phi behind), but should be enough to retain semantic correctness if user code ever causes that (corneer-case) code path to execute. It is also expected that such a trivial phi would be removed in later DCE passes anyway.
|
|
Before this change, global and function-scope `static const` declarations were represented as instructions of type `IRGlobalConstant`, which was represented similarly to an `IRGlobalVar`: with a "body" block of instructions that compute/return the initial value.
This representation inhibited optimizations (because a reference to a global constant would not in general be replaced with a reference to its value), and also caused problems for resource type legalization because the logic for type legalization did not (and still does not) handle initializers on globals (so global *variables* that contain resource types are still unsupported).
The change here is simple at the high level: we get rid of `IRGlobalConstant` and instead handle global-scope constants as "ordinary" instructions at the global scope. E.g., if we have a declaration like:
static const int a[] = { ... }
that will be represented in the IR as a `makeArray` instruction at the global scope, referencing other global-scope instructions that represent the values in the array.
This simple choice addresses both of the main limitations. A `static const` variable of integer/float/whatever type is now represented as just a reference to the given IR value and thus enables all the same optimizations. When a `static const` variable uses a type with resources, the existing legalization logic (which can handle most of the "ordinary" instructions already) applies.
Another secondary benefit of this approach is that the hacky `IREmitMode` enumeration is no longer needed to help us special-case source code emit for `static const` variables.
Beyond just removing `IRGlobalConstant`, and updating the lowering logic to use the initializer direclty, the main change here is to the emit logic to make it properly handle "ordinary" instructions that might appear at global scope.
One open issue with this change, that could be addressed in a follow-up change, is that "extern" global constants that need to be imported from another module (but which might not have a known value when the current module is compiled) aren't supported - we don't have a way to put a linkage decoration on them. A future change might re-introduce global constants as a distinct IR instruction type that just references the value as an operand (if it is available). We would then need to replace references to an IR constant with references to its value right after linking.
|
|
* Prefixing source files in source/slang with slang-
* Prefix source in source/slang with slang- prefix.
* Rename core source files with slang- prefix.
* Update project files.
* Fix problems from automatic merge.
|
